70- 13,980
BENTZ, William Frederick, 1940- THE MEASUREMENT OF COMMON STOCK PRICE RELATIVE f UNCERTAINTY: AN EMPIRICAL STUDY. f I I The Ohio State University, Ph.D., 1969 | Accounting £
University Microfilms, Inc., Ann Arbor, Michigan j t ...... ____ ...... _...... -i
<© Copyright by
William Frederick Bentz
I1970i I i
THIS DISSERTATION HAS BEEN MICROFILMED EXACTLY AS RECEIVED THE MEASUREMENT OF COMMON STOCK PRICE RELATIVE UNCERTAINTY:
AN EMPIRICAL STUDY
DISSERTATION
Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University
By
William Frederick Bentz, B.A., M.Acc.
******
The Ohio State University 1969
Approved by
Adviser Department of Accounting PLEASE NOTE:
Not original copy. Some -pages have very light type. Filmed as received.
University Microfilms ACKNOWLEDGMENTS
I gratefully acknowledge the constructive criticisms and suggestions provided by my dissertation reading committee, Professors Thomas J. Burns
(chairman), Diran Bodenhorn and Melvin Greenball. They have never failed to give prompt attention to this work in spite of the many demands on their time. Professor Burns deserves special thanks for creating a very favor able environment in which I could improve my research skills and pursue my research interests. As a dissertation advisor, he has been instrumental
in obtaining data, computer time, financial assistance, and all the other resources necessary to undertake an extensive empirical study. As an aca demician, Professor Burns has constantly demanded that I improve the rea soning behind each step in the dissertation, as well as suggesting ways in which the exposition of theory and results could be improved. Professor
Bodenhorn is responsible for much of the improvement in logic and internal
consistency which distinguish the dissertation from earlier drafts. Pro
fessor Greenball has made many substantive and methodological suggestions which have improved the dissertation, even though it was well in progress when he joined the committee.
Helpful suggestions have been received from Professors Cunnyngham, Cole
Nestel, and Lyle during various stages of the development of the disserta
tion.
Preliminary analyses of the data were run on the Ohio State University
7094 system, with the assistance of the College of Administrative Science
ii Data Center staff. Jim Boltz, Anita Gehr, and Marjorie Brundage provided
programming assistance as well as technical information.
Final computations and statistical analyses were run at the Kansas
University Computation Center. Jeff Bangert has provided assistance in the use of the program library, and in the operation of the GE 635 system.
Financial assistance has been provided by the Department of Accounting
at The Ohio State University in the form of teaching opportunities and re
search appointments. Support during much of my dissertation work was pro vided by The Haskins & Sells Foundation in the form of two Faculty Assis
tance Grants (1965-66 and 1966-67). Some financial support was also pro
vided by the Ohio Society of Certified Public Accountants in an award
called the Herman Miller Price (1966). The financial aid provided by these
sources is gratefully acknowledged since a doctoral degree would not have
been feasible without this support.
The secretarial assistance provided by the School of Business of The
University of Kansas is gratefully acknowledged. Mrs. Betty Bovee has
typed the dissertation with great care and cheerfulness. Mrs. Marcia Brown
typed many of the tables, and was responsible for duplication of the dis
sertation. Without the cooperation of Mrs. Bovee and Mrs. Brown, the dis
sertation would have been a much more difficult undertaking.
I am most thankful to my wife, Janet, and our children, Michael and
Jennifer, who provided encouragement and understanding throughout my grad
uate studies. In terms of unfilled needs, a family's investment in a doc
toral degree is never quite repaid.
iii VITA
June 14, 194C . . . Born - Dayton, Ohio
1962 ...... B.A., Economics, The University of Cincinnati, Cincinnati, Ohio
1964 ...... Public Accounting Internship, Ernst and Ernst, Columbus, Ohio
1964 ...... Internship, General Motors Corporation, Detroit, Michigan
1965 ...... M.Acc., The Ohio State University, Columbus, Ohio
1965-1967 ..... Teaching Associate, Department of Accounting, The Ohio State University, Columbus, Ohio
1967-1968 ...... Research Associate, Department of Accounting, The Ohio State University, Columbus, Ohio
1968-1969 ...... Assistant Professor of Business, School of Business, The University of Kansas, Lawrence, Kansas
PUBLICATIONS
Magazine Review: "The Mathematical Content of the Business School Curriculum" by David Novick, California Manage ment Review, Spring, 1966, which was reviewed in Management Services, July-August, 1966
Magazine Review: "Capital Budgeting of Interrelated Projects: Survey and Synthesis" by H. Martin Weingartner, Management Science, March, 1966, which was re viewed in Management- Services, September-October, 1966. TABLE OF CONTENTS
Page ACKNOWLEDGMENTS ...... ii
VITA ...... iv
LIST OF TABLES ...... vii
Chapter
I. INTRODUCTION AND SUMMARY ...... 1
A Statement of Objectives Outline of the Study Summary of the Results
II. THE INVESTMENT PROCESS ...... 28
Individual Investment: A Postulated Process
III. UNCERTAINTY AND PREDICTION MODEL ERROR ...... 45
Uncertainty: Some General Notions Uncertainty and Prediction Uncertainty Defined A Measure of Uncertainty Price Relative Variance as a Measure of Uncertainty Analytical Arguments for Attempting to Predict Price Relatives Summary
IV. THE PREDICTIVENESS OF COMMON STOCK PRICES AND PRICE RELATIVES 75
Introduction The Random Walk Hypothesis: A Review The Potentiality of Stock Price Prediction Naive Prediction Models An Economic Model of Stock Value
V. ANALYSIS AND INTERPRETATION OF THE PREDICTION MODEL’S PERFORMANCE ...... 122
Performance of the Exponential Smoothing Models Performance of the Growth Model
v Page
Implications for Measuring Uncertainty Potential Improvements in the Growth Model Predictions
APPENDIX
A ...... 156
B ...... 159
BIBLIOGRAPHY ...... 230
vi LIST OF TABLES
Table Page 1. Naive Prediction Model No. 1 - Prediction. Results for the Years 1958 Through 1967 Using the Best Smoothing Constant for the Ten Years Preceding the Year for which a Prediction is Being M a d e ...... 162
2. Naive Prediction Model No. 2 - Prediction Results for the Years 1958 Through 1967 Using the Best Smoothing Con stant for the Ten Years Preceding the Year for which a Prediction is Being Made ...... 164
3. Naive Prediction Model No. 3 - Prediction Results for the Years 1958 Through 1967 Using the Best Smoothing Con stant for the Ten Years Preceding the Year for which a Prediction is Being M a d e ...... 166
4. Naive Prediction Model No. 1 - Prediction Results for the Years 1958 Through 1967 Using the Smoothing Constants I n d i c a t e d ...... 168
5. Naive Prediction Model No. 2 - Prediction Results for the Years 1958 Through 1967 Using the Smoothing Constants I n d i c a t e d ...... 170
6 . Naive Prediction Model No. 3 - Prediction Results for the Years 1958 Through 1967 Using the Smoothing Constants Indicated ...... 172
7. Price Relative Prediction Performance with Respect to Mean-Squared Error for Sixty-one Companies over the Ten Year Period 1958 Through 1967 Using Three Exponen tial Smoothing Models with Both Shifting Smoothing Constants and Fixed Smoothing Constants ...... 174
8 . Price Relative Prediction Performance with Respect to Constant Bias for Sixty-one Companies over the Ten Year Period 1958 Through 1967 Using Three Exponen tial Smoothing Models with Both Shifting Smoothing Constants and Fixed Smoothing Constants ...... 176
vii Table page 9. Price Relative Prediction Performance with Respect to Proportional Bias for Sixty-one Companies over the Ten Year Period 1958 Through 1967 Using Three Expo nential Smoothing Models with Both Shifting Smooth ing Constants and Fixed Smoothing Constants ...... 178
10. Price Relative Prediction Performance with Respect to Correlation for Sixty-one Companies over the Ten Year Period 1958 Through 1967 Using Three Exponen tial Smoothing Models with Both Shifting Smoothing Constants and Fixed Smoothing Constants ...... 180
11. Price Relative Prediction Performance Ranking of Three Exponential Smoothing Models, Incorporating Shifting as Well as Fixed Smoothing Constants, by Industry for the Ten Year Period 1958 Through 1967 182
12. Comparative Price Relative Prediction Results Using Fixed Smoothing Constants versus Shifting Smoothing Constants for Sixty-one Companies from 1958 Through 1967 ...... 184
13. Comparative Absolute Prediction Errors Using Fixed Smoothing Constants versus Shifting Smoothing Con stants for Sixty-one Companies from 1958 Through 1967 ...... 186
14. Kolmogorov-Smirnov One Sample Test for the Normality of the Constant Exponent Smoothing Model's Pre dictions for Sixty-one Companies from 1958 Through 1967 where Shifting Smoothing Constants are Used ...... 188
15. Kolmogorov-Smirnov One Sample Test for the Normality of the Constant Exponent Smoothing Model's Pre diction Errors for Sixty-one Companies from 1958 Through 1967 where Shifting Smoothing Constants are Used ...... 190
16. Kolmogorov-SmirnovOne Sample Test for the Normality of the Constant Exponent Smoothing Model's Pre dictions for Sixty-one Companies from 1958 Through 1967 where Fixed Smoothing Constants are Used ...... 192
17. Kolmogorov-Smirnov One Sample Test for the Normality of the Constant Exponent Smoothing Model's Pre diction Errors for Sixty-one Companies from 1958 Through 1967 where Fixed Smoothing Constants are U s e d ...... 194
viii Table Page 18. Constant Exponent Smoothing Model's Mean-Squared Error versus Price Relative Variances as Measures of Uncertainty ...... 196
19. Analysis of Covariance Table for the Price Relative Variance and the Constant Exponent Smoothing Model's Mean-Squared Prediction Error for Sixty- one Companies Using Shifting Smoothing Constants ...... 198
20. Analysis of Covariance Table for the Price Relative Variance and the Constant Exponent Smoothing Model's Mean-Squared Prediction Error for Sixty- one Companies Using Fixed Smoothing Constants ...... 199
21. Analysis of Covariance Table for Price Relatives and the Constant Exponent Smoothing Model's Price Relative Predictions for Sixty-one Companies from 1958 Through 1967 Using Shift ing Smoothing Constants ...... 200
22. Analysis of Covariance Table for Price Relatives and the Constant Exponent Smoothing Model's Price Relative Predictions for Sixty-one Companies from 1958 Through 1967 Using Fixed Smoothing Constants ...... 201
23. Kolmogorov-Smirnov One Sample Test for the Nor mality of Growth Model Prediction Errors for Sixty-one Companies from 1958 Through 1967 ...... 202
24. Kolmogorov-Smirnov One Sample Test for the Nor mality of Growth Model Price Relative Pre dictions for Sixty-one Companies from 1958 Through 1967 ...... 204
25. Kolmogorov-Smirnov One Sample Test for the Nor mality of Price Relatives for Sixty-one Companies from 1958 Through 1967 ...... 206
26. Analysis of Covariance Table for Price Relative Variances and the Growth Model's Mean-Squared Errors for Sixty-one Companies ...... 208
27. Analysis of Covariance Table for Price Relatives and the Price Relative Predictions Produced by the Growth Model for Sixty-one Companies from 1958 Through 1967 ...... 209
ix Table Page 28. Analysis of Variance Table for Price Relatives Regressed on Price Relative Predictions from the Growth Model for Sixty-one Companies from 1958 Through 1967 ...... 211
29. The Optimal Linear Correction of the Price Rela tive Predictions Made by the Growth Model for Sixty-one Companies from 1958 Through 1967 ...... 212
30. Price Relative Prediction Error Decomposition for the Postulated Growth Model and for an Exponen tial Smoothing Model Tested on Sixty-one Com panies over the Ten Year Period 1958 Through 1967 ...... 213
31. Price Relative Prediction Performance of the Postulated Growth Model versus an Exponential Smoothing Model for Sixty-one Companies over the Ten Year Period 1958 Through 1967 ...... 215
32. Price Relative Prediction Error Decomposition for the Postulated Growth Model and for an Exponential Smoothing Model Using Shifting Smoothing Constants for Sixty-one Companies over the Ten Year Period 1958 Through 1967 ...... 217
33. Price Relative Prediction Performance of the Postulated Growth Model versus an Exponential Smoothing Model with Shifting Smoothing Con stants for Sixty-one Companies over the Ten Year Period 1958 Through 1967 219
34. Prediction Performance of Prediction Model No. 1 versus Prediction Model No. 3 Using Shifting Smoothing Constants ...... 221
35. Performance of Prediction Model No. 1 versus Prediction Model No. 3 Using Fixed Smooth ing C o n s t a n t s ...... 222
36. Price Relative Variance versus the Mean- Squared Prediction Error of Model No. 3 Using Shifting Smoothing Constants ...... 223
37. Mean-Squared Error versus Variance for Sixty- one Companies...... 224
38. Prediction Performance of Prediction Model No. 3 Using Shifting Smoothing Constants versus the Postulated Growth Model ...... 225
x Table Page 39. Mean-Squared Prediction Errors Using the Mean of Past Price Relatives versus the Price Relative Variance as a Measure of Uncertainty...... 227
40. Prediction Performance of the Arithmetic Mean of Past Price Relatives versus Model No. 3 Using Fixed Smoothing Constants ...... 228
41. Prediction Performance of the Arithmetic Mean of Past Price Relatives versus Model No. 3 Using Shifting Smoothing Constants ...... 229
xi CHAPTER I
INTRODUCTION AND SUMMARY
A Statement of Objectives
Rate of return on investment and the uncertainty associated with that rate of return are two major elements in capitai investment theory, mathe matical algorithms for selecting portfolios, financial utility theory, and the evaluation of prior investment performance. This study is concerned with the prediction of common stock price relatives and with the estimation of the uncertainty associated with those predictions.'*'
According to the neo-classical economic theory of the firm, a corpora tion should select investments so as to maximize stockholder wealth, which 2 is a function of return on investment and the uncertainty of that return.
In portfolio selection theory, the objective is to allocate an investor's 3 resources so as to maximize utility. In the economic and the finance literature, utility is a function of rate of return on investment and the .
*A common stock price relative is equal to the annual rate of return on investment plus 1.0. Annual price relatives are calculated by dividing the sum of cash dividends received during the year plus the cash value of shares held at year end, by the cash value of shares held at the beginning of the year, assuming no new investment. In this study, rate of return is used synonymously with price relative. 2 John Lintner, "Dividends, Earnings, Leverage, Stock Prices and the Supply of Capital to Corporations," The Review of Economics and Statistics, Vol. XLIV, No. 3 (August, 1961), pp. 243-269. 3 The classic work in this area is by Harry M. Markowitz, Portfolio Selection (New York: John Wiley and Sons, Inc., 1959), Part IV.
1 4 level of uncertainty associated with that rate of return. Therefore, when evaluating the economic success of a corporation, of an investor’s port folio, or of an investment fund for a large number of investors, the rate of return on investment and the uncertainty of that rate of return are two very important measures of performance.^ It seems clear that the measure ment and prediction of rate of return on investment, and the measurement and estimation of the uncertainty of the rate of return on investment, are significant elements of financial planning and financial performance mea surement .
The objective of this study.— Our purpose is the development and ap plication of the methodology necessary to measure the uncertainty of the common stock price relatives of individual corporations. The study is of practical importance because our measures of uncertainty can be used in other research studies, as inputs for the development of subjective prob ability distribution estimates made by financial analysts, or in evaluating the historical performance of common stocks. The uncertainty measurement methodology developed herein can be used to measure the past uncertainty of those common stocks for which we do not provide uncertainty estimates.
Our analysis begins with the development of a conceptual model of the investment process, because we must have a set of standard conditions for 6 which the analysis makes sense. A methodology for measuring the uncertain-
^See James Tobin, "Liquidity Preference as Behavior Toward Risk," Review of Economic Studies, Vol. XXVI, No. 67 (February, 1958), pp. 65-86.
^Peter 0. Dietz, "Measurement of Performance of Security Portfolios," Journal of Finance, Vol. XXIII, No. 2 (May, 1966), pp. 267-274.
^The concept of a set of standard conditions in measurement theory is provided by C. West Churchman, Prediction and Optimal Decision (Englewood Cliffs: Prentice-Hall, Inc., 1961). 3 ty of common stock price relatives, which is consistent with the conceptual model of the investment process, is then developed and applied to create measures of uncertainty for each of sixty-one major corporations.
Motivation for the study.— The two measures of uncertainty usually discussed in connection with common stock investment are (1) subjective estimates of the probability density function of expected returns on in vestment, and (2) the variance of observed rates of return about the mean rate of return for some block of investment periods.^ Subjective probabil ity estimates should be based predominately on past experience unless the investor has some reason to believe a shift has occurred in the level of uncertainty associated with a particular common stock, or set of common stocks. In any case, historical evidence of uncertainty is the usual starting point in formulating expectations about the future.
The variance of past price relatives about the arithmetic mean of a series of past price relatives is an appropriate measure of uncertainty whenever the price relatives can.be described as observations from one popu lation. If the price relatives are independent observations from a single population, then it is proper to use sample statistics to estimate the mo ments of the population's probability density function.
However, the variance of price relatives about the mean of a series of past price relatives is not believed to be a complete measure of uncer tainty. The variance is an incomplete measure because (1) the moments of the population's probability density function are not known, so there always
^Charles G. Ferreira, "Quantification and Measurement of Risk: An Empirical Study of Selected Common Stocks" (unpublished Ph.D. dissertation, School of Business, University of Washington, 1966). 4 exists a sampling error in the estimation of population moments, and be cause (2) the price relatives may not behave as if they are represented by a stable probability density function.
At this stage we might hypothesize either that the variance of past price relatives tends to overstate uncertainty, or that the variance of past price relatives tends to understate uncertainty. In Chapter III we argue that the mean-squared prediction error is the more meaningful measure of the uncertainty associated with a set of predictions. The mean-squared prediction error computed over our ten-year test period is equal to the sum of the squared differences between the ten actual annual price relatives of a company and the respective predicted price relatives, divided by the num ber of predictions, ten. Accordingly, the mean-squared price relative pre diction error is the standard by which we evaluate the variance of past price relatives as a measure of uncertainty.
On the one hand, we can argue that, on the average, squared errors of ex-ante price relative predictions may tend to exceed the ex-post dif ferences squared between the price relatives and the mean of these price relatives because (1) of sampling errors in the ex-ante estimates of the price relative mean, or because (2) the price relatives are not properly described by a stable probability density function. Both situations lead to a difference between the ex-post measure of the mean of a series of an nual price relatives and the ex-ante prediction of that mean. Therefore, prediction errors tend to be larger than one would expect based on the variance of annual price relatives about the mean price relative (ex-post).
On the other hand, we can argue that, on the average, squared errors associated with ex-ante price relative predictions may tend to be less than 5 the ex-post squared differences between the price relatives and their mean because (1) the mean may not contain all the information relevant to pre diction which is contained in the price relative series, or because (2) any correlation between the price relative series and a cue (prediction) variable is ignored in the calculation of differences between the price relatives and their mean value. If there exists a trend in a common stock's price relative series, then the price relatives are not independent obser vation* from a single population which is described by a stable probability density function, and the trend information should be used to predict the price relatives, not just the mean. Likewise, when another variable can be used to predict price relatives, then the average prediction error can be made smaller than the average difference between past price relatives 8 and their mean value.
• Prediction models.— As defined, the study of uncertainty is in terms of prediction errors, which necessitates the formulation of predictions.
We construct five prediction models which are used to predict annual price relatives for sixty-one common stocks over the ten-year period 1958 through
1967. Four of the five prediction models are based exclusively on the in formation contained in the time-series of past price relatives for that company. The fifth model incorporates economic information and corporate data in the formulation of price relative predictions.
The three linear prediction models are a constant model, a linear trend model, and a constant exponent model. If price relatives vary about a normal value, which is constant with respect to time, then that normal
O See section six of Chapter III. 6 value is the best prediction of future price relatives. Similarly, the linear trend model is based on the hypothesis that price relatives are equal to a constant amount plus a constant increment each year. The con stant exponent model is based on the hypothesis that the natural (Naperian) logarithm of a series of price relatives is constant with respect to time.
The three models can be formulated as follows:
(1.0) Constant model: R = a. it 1
(1.1) Linear trend model: = a. + b. • t It 1 1
a. (1.2) Constant exponent model: R = e ,
th where a^ and b^ are equation constants which are unique to the i securi ty, and R is the price relative of security i during period t.
The equation constants a^ and b^ are estimated for each company by means of exponential smoothing. The smoothing constant is allowed to vary from 0.0 through 1.0 and a set of prediction errors is generated for each smoothing constant tested. That smoothing constant which resulted in the smallest mean-squared prediction error for the prior ten years is used to estimate the equation constants, which are used to estimate next year's price relative. This smoothing constant selection procedure is repeated ten times in order to make ten annual predictions for each security. These models are described completely in equations 4.3 through 4.18, and in the text following those equations.
These three models are designed to incorporate information which is contained in the time-series of past price relatives.
The unweighted arithmetic mean of the prior ten price relatives is 7 also used to predict future price relatives. This method resembles the
constant model, but there are two important differences. The arithmetic mean calculation is not dependent on the sequence of the price relatives
in the sample period, whereas the constant model based on exponential
smoothing is influenced by the order in which the prior price relatives were observed. Also, each past price relative is weighted equally in the
fourth method, but not in the constant model^ where exponential smoothing
is used.
In addition to the use of naive prediction models, we develop an eco nomic model of common stock price relatives. The model is an economic model in the sense that measures of economic performance are used to pre
dict future price relatives, as opposed to using only, past price relatives
to predict future price relatives.
Economic growth, price stability, earnings stability, sales growth,
earnings growth, and past price relative performance are variables which
appear in many stock valuation models, and which are examined by profes
sional security analysts. Accordingly, the postulated economic model is
of the form:
where
Rt is the expected price relative for period t,
(t = 1948, 1949, .. 1967) for any security,
is a constant (i = 0 , 1 , 2 , • • • * 6)
Gt is the expected growth rate of gross national product
from the year t-1 to the year t, 8
0 ^ is one plus the average historical geometric growth
rate in operating income per share,
'S is one plus the average historical geometric growth
rate in sales per share,
Dt_^ is the difference between the price relative for
year t-1 and the average price relative for the five
one-year periods ending with year t-1 ,
Ut_^ is the variance of the changes in operating income
over the five one-year periods #nding with the year
t-1 ,
Vt_^ is the variance of price relatives over the five one-
year periods ending with year t-1 .
The prediction model summarized in equation 1.3 is the hypothesized
form of the model. Several variations of the six independent variables
included in 1.3 were tested, so that a total of fourteen independent vari
ables were actually subjected to empirical test. These fourteen variables
are listed on page 9.
A least-squares regression model, applied on a step-wise basis, pro
vides preliminary information about the ultimate form to be taken by the
prediction model.^ Because it is applied on a step-wise basis, the re
gression model tests indicate (1) how many variables are significantly cor
related with the past common stock price relatives of each company, (2)
9 This program is a completely rewritten, double precision, FORTRAN IV version of the original UCLA BIMED 34 Stepwise Regression Program. Hodson Thornber of the Center for Mathematical Studies in Business and Economics modified the original BIMED routine, Anita Gehr adapted the program to The Ohio State University 7094 System, and the author made some minor changes to adapt the program to the GE 635 System at The University of Kansas. ______Industry______Rubber Petroleum Beverage Goods Products Cement Textiles Last year's price relative less the average for the previous five years X X X X Last year's price relative , X The variance of the company's price relatives over the prior five years The variance of changes in operating income about the average change for the prior five years X X X X The average proportional price range over the prior five years X The average proportional price range over the prior two years The geometric average growth rate in operating income (before depreciation) per share over the prior six years XX The geometric average growth rate in operating income (before depreciation) per share over the prior two years X X The geometric average growth rate in net sales per share over the prior six years X The geometric average growth rate in net sales per share over the prior two years X X The annual growth rate in gross national product over the year for which a price relative is being predicted X The annualized growth rate a fourth-quarter gross national product prior to the prediction period X X The annual growth rate in gross national personal income over the year for which price relatives are being predicted X The annualized growth rate of fourth-quarter gross national personal income X 10 which set of the fourteen variables are most highly correlated with the past price relatives of each company, and (3) the order in which the inde pendent variables are admitted into the prediction equation, which is a * • measure of their relative importance.
At least six variables usually enter the step-wise regressions, so we have decided to use six variables in each prediction model. On an in dustry by industry basis we select those six measures, out of the fourteen possible, which are the most highly correlated with common stock price relatives over the period 1948 through 1957. Five different sets of six variables are used for the five different industries (see page 9) .
Preliminary tests of the data indicate that extreme predictions are sometimes generated by the model. Accordingly, predictions were bounded by the range of past price relatives, or by 0.8 and 1.50, whichever is the more restrictive set of bounds. This bounding process significantly af fects the predictions of 6 out of the 61 companies tested.
Once a variable set is selected for each industry, that same set is used for the entire test period, 1958 through 1967. For each company, ten prediction equations are constructed for the ten prediction periods. As illustrated in equation 1.3, the constants Bq , B^, ..., B^ are estimated by means of a least-squares regression model applied to the variable ob servations over the prior ten years. For example, the equation used to predict 1958's price relative for a particular company is constructed from the data available during the prior ten years, 1948 through 1957. Like wise, the data from 1949 through 1958 is used to construct the equation from which the 1959 price relative is predicted. This process is repeated- ten times for each company. 11
Nimrod and Bower have used this regression model approach to predic tion in the futures market for hog bellies.^ Their simulated investment program was profitable, so they achieved some predictive success. Also,
Whitbeck and Kisor constructed regression equations on a cross-sectional basis for a large number of firms, then they applied these regression equation coefficients to each company's data to estimate the intrinsic value of that company's stock.^ By purchasing these stocks for which the intrinsic value estimate greatly exceeded the current price, they were able to earn a favorable return. The results reported by Shelton also suggest that price predictions are possible, and that we are not restricted to earn- 12 ing an average rate of return on common stocks, or to lucky choices.
The Value Line Investment Survey represents an objective, but prag matic approach to common stock valuation. Their methodology has some em pirical credence since it is the methodology of an investment advisory group which is selling a service. Accordingly, our prediction model in corporates the methodology of Nimrod and Bower and an adaptation of the variables examined by Value Line as described by the founder, Arnold
Bernhard."^
^Vance L. Nimrod and Richard S. Bower, "Commodities and Computers," Journal of Financial and Quantitative Analysis," Vol. II, No. 1 (March* 1967), pp. 61-73.
^Volkert S. Whitbeck and Manown Kisor, Jr., "A New Tool in Investment Decision-Making," Financial Analysts Journal, Vol. 19, No. 3 (May-June, 1963), pp. 55-62. 12 John P. Shelton, "The Value Line Contest: A Test of the Predict ability of Stock-Price Changes," The Journal of Business, Vol. L, No. 3 (July, 1967), pp. 251-269. 13 Arnold Bernhard, The Evaluation of Common Stocks (New York: Simon and Schuster, 1959). 12
Because investors have different goals and objectives, Value Line evaluates and ranks companies by three different criteria: quality grade, appreciation potentiality in the next 3 to 5 years, and probable market performance in the next 12 months. Quality grade is dependent on the sta bility of earnings and dividends, as well as growth. A high quality com pany is a financially sound, low risk investment for which there is a low probability of failure, and a relatively low probability of missing divi dends. High growth rate is another attribute of quality because growth of earnings is an important prerequisite to dividend growth and price appre ciation. To summarize, a high quality stock is one which is growing in a stable, predictable manner.
Price appreciation potential over the next 3 to 5 years is another factor by which stocks are ranked. Long-run price appreciation potential depends on the expected performance of the economy in general, the expected performance of the industry, and the sales-earnings potential of each firm.
Because individual stocks are being ranked, economic conditions and indus try factors are only important in so far as they influence the price per formance of the firms being ranked. The average level of appreciation over a three to five year period is used in order to avoid the distortions that 14 might be caused by a cyclical variation in single years.
The other criteria by which stocks are ranked is probable market per formance over the next twelve months.
14 Bernhard, p. 119. 13
The third method gives us an indication as to whether a stock is cheap or dear now in relation to its own Intrinsic Value this year, ^ and whether it is cheap or dear in relation to all other stocks now.
Because only the price performance in the next twelve months is being con sidered, the third method is based on "an estimate of earnings and dividends in the next twelve months and the likelihood that the estimated earnings and dividends will command a price capitalization similar to the past."^
The theoretical analysis and the empirical evidence which provides the logical transition from the Value Line methodology to the model postulated in 1.3 are presented in Chapter IV. The methodology used by Value Line represents but one approach to common stock selection, but the continued success of The Value Line Investment Survey is evidence of the acceptance and interest in this relatively objective approach to investment selection.
Many authors believe that one cannot earn a rate of return which is above average unless one has inside information. By definition, all inves tors and financial analysts have access to generally available information, so that current market prices reflect the information currently available.
Some of the relevant studies are reviewed in Chapter IV.
The analysis of the five prediction models can be described in terms of the major steps involved. First, all five prediction models are used to make predictions over the ten year test period, then the mean-squared
4 prediction error is calculated for each prediction model, for each of the
^Bernhard, p. 110 .
^Bernhard, p. 110 . 14
61 companies. Next, the mean-squared prediction errors associated with the three naive time-series models are compared in order to determine which one is the best. We then compare the average mean-squared prediction error for the best naive time-series model with the average mean-squared pre diction error associated with the arithmetic mean model. The results of this comparison indicate which is the best naive model as well as indicating if there is any information in the time-series of past price relatives.
Next, we compare the average mean-squared prediction error associated with the economic growth model to the average mean-squared prediction error associated with the best naive model. The research hypothesis in the latter test could easily be stated negatively, but the economic model supposedly contains more variables and more information, so the research hypothesis is stated in a positive manner.
Hy thesis 1.— The proposed economic growth model produces better price relative predictions than the best of the naive prediction models. .
The final step is a comparison of the mean-squared prediction errors associated with the best prediction model of the five tested, with the variance of the past price relatives of each company about their own mean.
The variance is measured according to the equation 10 _ a . 4) vi = ^ (Rit" V 2/9’ t=l where i denotes the firm subscript
(i = 1 , 2 ..... 61)
is the price relative variance for company i 15
is the price relative for company i (i = 1 , 2 , 61)
during period t (t = 1958, 1967)
is the arithmetic mean of the price relatives for company i
(i = 1, 2, ..., 61) over the period 1958 through 1967.
As mentioned above, we believe that the variance is an incomplete
measure of uncertainty because the mean price relative for the period
relative for the period 1958-1967 is not known in advance, and because
we do not believe the price relative series can actually be described by
a distribution which is stable over time and under all types of conditions.
Hypothesis 2 is stated accordingly.
Hypothesis 2.— The variance of a series of observed price relatives
about the mean price relative for that series tends to understate the
uncertainty of those price relatives as measured by the mean-squared
error. As we have pointed out above, it would not be too surprising to
find that the price relative variance may overstate uncertainty. However,
the time-series of price relatives is not expected to contain sufficient
information to support the hypothesis that the price relative variance
overstates uncertainty.
In summary, we argue that the proper measure of the uncertainty
associated with a price relative over some prior set of investment periods
is the mean-squared prediction error, not the variance of a set of price
relatives about the mean price relative. It is hypothesized that the mean-
squared prediction errors associated with the best of the five prediction models tested are larger than the respective price relative variances. 16
Various other hypotheses are tested in conjunction with the selection of the best prediction model, the evaluation of the resulting prediction errors, the existence of risk classes, and the form of the relationship of the mean-squared prediction error to the price relative variance of each company.
Outline of the Study
The organization of this paper parallels the thought processes and the methodology used to attack the problem of measuring the past uncer tainty of common stock price relatives. In Chapter II we present a con ceptual model of the investment process which consists of six steps:
(1) the formation of investor goals, (2) the recognition of constraints on the investment process, (3) the creation of an algorithm for select ing a portfolio of securities, (4) the search for and description of acceptable securities, (5) the application of a predetermined algorithm to the set of potential security investments, and (6) the selection of an efficient set of securities. The description of steps one through three represents a theory of investor behavior which is meaningful under a specified set of conditions.
The theory of investor behavior is assumed to apply to a signifi cant number of investors, but not to all investors. With respect to goals, we assume that investors seek to maximize utility, where utility is a function of the relative increase in wealth. The relative increase in wealth over a year is a function of the relative increase in the price of each common stock owned. 17
Next, some constraints are placed on the investment process in order to make it more amenable to formal study. Consumption, once decided upon, is treated as a constraint because increased current consumption reduces the amount of current investment in wealth creating resources. Invest ment periods of one calendar year, and the holding of cash dividends until year end are two other constraints placed on the investment process.
The mean, variance model is the most widely accepted and well-defined approach to the selection of efficient portfolios of common stocks. The method is operational because algorithms have been developed which per form the computations necessary to select the most efficient portfolio from among a set of alternative investments.
The mean, variance model is based on the assumption that investors are risk averse, and that risk can be adequately described by the variance or standard deviation of rates of return. Accordingly, expectations about common stock price relatives and the uncertainty of those expectations are the most important factors in the investment process, and thus are the most deserving of isolation for further study.
The concept of the uncertainty of common stock price relatives is discussed in Chapter III. Some general comments about traditional dis tinctions between risk and uncertainty are presented in section one.
Next, we define a common stock price relative and indicate how it is measured under various circumstances. The price relative is discussed in
Chapter III because it is the subject of our predictions— that variable which 18
is uncertain. In section three, uncertainty is defined as the extent to which price relatives can be predicted one year in advance. The prob
ability density function of the prediction errors is the basis for measur ing uncertainty. The mean-squared prediction error is the second moment about zero of the prediction errors, which is our measure of uncertainty
(predictability) associated with past price relative predictions.
The variance of past price relatives about the mean of those price relatives is frequently used as a measure of uncertainty. We argue that
the variance is a poor measure of uncertainty because (1) the mean is not known at the time predictions are being made, (2) price relatives may not properly be represented by a stable probability density function, (3) the price relative series may contain some information, such as trend, which the variance calculation ignores, and (4) another predictor vari able may be highly correlated with the price relative series.
A mathematical argument for attempting to predict price relatives is presented in section six of Chapter III. The variance of price relatives about the expected price relative tends to overstate uncertainty whenever there exists a predictor variable which is correlated with the price rela tive series because the conditional variance, given the predictor variable, is less than the unconditional variance of the price relatives.
The improvement of price relative expectations or predictions is im portant from a market point of view, as well as from the point of view of individual investors. More intelligent analyses of corporate performance should produce better predictions of long-run profitability, and thus im prove the allocation of capital funds by the market process. Moreover, 19
individual investors can earn above average rates of return only by being
better predictors, because the probability of earning high rates of re
turn on randomly selected portfolios is small.
A part of the literature concerning the time-series behavior of com
mon stock price changes bears on the subject of price relative prediction.
In regard to short-term price changes in common stocks and other specula
tive price series, the general concensus is that average price changes
tend to be zero and the variance of price changes tends to be slightly
greater than is expected for a normally distributed random variable. This
viewpoint, the random-walk hypothesis of stock price behavior, is reviewed
in Chapter IV. The time series of stock price changes is important be
cause a common stock price relative is the sum of the dividend yield, and
the price change divided by the price at the start of the period, plus 1.0.
The naive prediction models, designed to predict future price relatives based on the behavior of past price relatives, are influenced by the na
ture of stock price changes over time. The random-walk studies concen
trate on short term price changes, so they do not preclude the possibility
of an upward drift in stock prices over annual periods.
The random walk literature is also reviewed in order to determine
the implications of this hypothesis of stock price behavior with respect
to the predictability of common stock price relatives using econometric models. Advocates of the random walk hypothesis not only accept that prediction may be possible, but they also argue that perfect markets, or nearly perfect markets, are necessary in order to explain random price movements. Adjustments to information about economic forces in the economy 20
are rapid and complete in highly efficient markets. Prediction of those
economic factors that influence stock price changes should lead to pre
diction of stock price changes, and thus prediction of price relatives.
Four naive prediction models are developed to predict common stock
price relatives based on past price relatives. A constant model, a linear
trend model, and a constant exponent model are. tested. The constant terms
in these models are estimated by means of exponential smoothing. In ad
dition to these three models, we use a simple arithmetic mean of the prior
ten price relatives in order to predict next year's price relative.
The final section of Chapter IV contains the development of the eco
nomic prediction model, also referred to as the growth model which is
described on page 7. The value Line Investment Survey Method and several
empirical research studies are summarized, and their implications for
our model explained. The analysis explaining the inclusion of each
variable in the prediction model, and a statement of the model, conclude
Chapter IV.
The relative predictive performance of the naive prediction methods
and the proposed growth model is evaluated in Chapter V. The five in
dustries selected for study are the beverage producers, cement producers,
integrated domestic oil producers, textile makers and producers of tire
and rubber goods. The financial data for the firms was provided by the
Binary-Type Annual Industrial Compustat magnetic tape library distributed by Standard s Poor's, Inc. After eliminating those companies for which
twenty years of financial data is not yet available, there are 61 companies
in the five industries to be included in the sample. Predictions are 21 made for the years 1958 through 1967, inclusive.
Summary of the Results
A capsule summary of the overall performance of the five prediction models is presented below.
Average Mean- Prediction Method Squared Error
Naive Models:
Constant model 0.142
Linear model 0.148
Constant exponent model 0.140
Mean of the prior ten price relatives 0.133
Economic model 0.175
The three naive time-series models performed well and the results are interesting in several respects. The constant exponent model is generally superior to the other two models in minimizing the mean-squared prediction error over the test period 1958 through 1967, as we had expected. The ranking of the naive models is significant at the 99.9% confidence level, and the Kendall Coefficient of Concordance was 0.64, which indicates a generally high degree of agreement among the rankings (see Table 7).
However, the difference between the constant model and the constant exponent model is minor, and it is not significant at the 95% confidence level when the two models are compared directly.
The problem of selecting an optimal exponential smoothing constant for each of the three naive prediction models, for each firm, proved to be an important issue. Each of the three naive prediction models was implemented in two ways. One method involves finding that smoothing con
stant value which results in the smallest mean-squared prediction error
for the test period. The other method involves using a set of smoothing
constants, one for each year. For example, equation constants needed to predict 1959's price relative, for a particular company, are estimated by means of the exponential smoothing constant which results in the smallest mean-squared prediction error over the previous ten years, 1949 -
through 1958. For 1960's prediction, the smoothing constant which results in the smallest mean-squared prediction error for 1950 through 1959 is used to estimate the equation constants. For this method, the smoothing constant may change from year to year to year, so that there is a set of ten optimal smoothing constants associated with each price relative series.
The same smoothing constant tends to be used over time for each company, but there are many cases for which several different smoothing constants are used over the ten year test period.
The use of shifting smoothing constants is our simulation of the investment prediction process. The determination of an optimal smoothing constant, ex-post, was done in order to determine the maximum possible benefit of improved smoothing constant selection procedures. For the constant exponent model, the potential average reduction in the mean- squared prediction errors is 14%. The details of this analysis are contained in tables 12 and 13.
The Kolmogorov-Smimov test indicates that the prediction errors resulting from the constant exponent model are not normally distributed, although a visual inspection of a histogram of the prediction errors does 23 indicate significant non-normality. A comparison of the observed interval frequencies with those of a standardized, normally distributed random vari able indicates that there is a greater concentration of prediction errors about the mean error, and less concentration in the tails of the error distribution than is expected. As a result, there is no evidence that the variance of prediction errors is undefined. Tables 15 and 17 provide the details of the tests of normality.
The fourth naive prediction model, the use of' the mean of the prior price relatives to predict next year's price relative, performed better than the constant exponent model (see Table 41). As a result, it is the best of the four naive prediction methods.
The performance of the postulated growth model was interesting, but disappointing. The average mean-squared error from the growth model was about 20% greater than that of the constant exponent model using shifting smoothing constants. This difference is significant at the 99.9% confidence level for the Wilcoxen matched-pairs, signed-ranks test, and is significant an the 99.9% level for the t-test. We must reject the hypothesis that the growth model is better than the naive models.
In formulating the economic model, it is not possible to specify in advance how each variable should best be measured. By using a stepwise regression process, those variables which have the highest historical explanatory power, with respect to price relatives, are selected for use in the final prediction model. The growth rate of sales and the growth rate of operating earnings before depreciation are highly correlated (par tial correlation) with common stock price relatives, as is the uncertainty of operating income before depreciation. For some industries, two year 24 growth rates are more highly correlated with price relatives than six- year average growth rates. In other industries, the longer term average growth rates dominate. In all cases, one of the six measures of general economic activity enters the regression equations at about the fifth step. The general economic measures include annual and fourth quarter annualized growth rates for gross national product, gross disposable per sonal income, and personal expenditures on durable goods. A more complete description of the variables used in the prediction equations for each industry is presented in Chapter V.
The prediction errors resulting from use of the postulated growth model appeared to be approximately normal in distribution when reviewed in histogram form. However, the Kolmogorov-Smimov test indicates (Table
24) that these price relative prediction errors are not normally distri buted. A comparison of the observed interval frequencies with those of a standardized, normally distributed random variable indicates that there is a greater concentration of prediction errors about the mean error, and less concentration of large errors in the tails of the error distribution than is expected. These results are consistent with the analysis of the prediction errors from the constant exponent prediction model.
The major purpose of this research is the measurement of uncertainty.
The mean-squared prediction error is the criterion by which we evaluate uncertainty. We argue that predictions should be based on data available to investors at the time predictions are actually being made. Because uncertainty is measured in terms of prediction errors, that model which minimizes the mean-squared prediction error is the best prediction model. 25
Consistently, the best available measure of uncertainty is the mean-squared prediction error provided by the best prediction model.
If a new prediction model can be develpped which incorporates informa tion available to investors at the time predictions are made, and which results in smaller prediction errors, then that model should be used to measure uncertainty. Until such an improved model is developed, the mean-squared error of predictions based on the mean of prior price rel atives best measures uncertainty.
Hypothesis 2 is accepted^ The variance of a series of price relatives about the mean price relative for that series does tend to understate uncertainty as measured by the mean-squared prediction error, because of the difficulty of predicting the mean. All of our prediction models re sulted in an average mean-squared prediction error which was significantly larger than the average price relative variance, for the companies tested
(see Table 37).
The failure of the three naive prediction models to provide better predictions than the use of a mean of the prior ten price relatives leads us to two possible conclusions. One conclusion is that there is little or no information in the series of past price relatives which is not cap tured by the simple arithmetic mean. The other conclusion is that price relatives can be described by a mean-reverting process for the majority of companies. In our tests of the naive prediction models, the smooth ing constants tended to approach 1.0 for a few companies, which indicates that the process associated with those price relatives may not be mean- reverting. However, our over-all results indicate that price relatives tend to be mean reverting. We have further evidence that the variance may be a misleading measure of uncertainty- We found no linear first degree o< second degree
equation which can be used to relate the mean-squared prediction error
from a price relative to the variance estimate of that price relative.
In other words, the mean-squared error was not simply larger than the variance estimates by some constant amount or by some proportionality
factor. The importance of this finding is that the relative uncertainty
of two price relatives will vary, depending on which measure of un
certainty is used. Ordinal rankings, as well as interval scale com parisons, may be influenced by the measure selected. For example, let
the variance of the price relatives of company B equal the variance of
the price relatives of company A. Their respective mean-squared prediction
errors may be equal, or the mean-squared error for B may be greater than,
or less than, that of A. Moreover, let the variance of B's price relatives
equal two times the variance of A's price relatives. Our findings indicate
that the mean-squared error of company B's price relatives may be three
times as large as the mean-squared error of A's price relatives.
Our inability to find a statistical relationship between the mean-
squared prediction errors, and the price relative variances for the
sixty-one companies in the sample has implications with respect to one's
ability to separate firms into risk classes. As summarized in Table 20 we found significant mean effects in the industry classifications of mean-squared prediction errors, even after adjustment by linear regression
for the price relative variances. The within-industry regression co
efficient is clearly significant, but industry differences in the mean- 27 squared errors remain after adjustment for the regression. These results support the belief that using mean-squared prediction errors to define risk classes will result in a different classification than results from used price relative variance as the measure of uncertainty.
To summarize, the mean-squared error of predictions based on the arithmetic mean of the prior ten price relatives is believed to be a better measure of uncertainty than the price relative variance about the mean price relative for the prediction period. CHAPTER II
THE INVESTMENT PROCESS
Much has been written about the operation of economic systems and the
problem of investment alternative selection within the context of a compet
itive market economy. In spite of the complexity involved, many writers
have been attempting to capture the essential elements of the processes by which people make investment decisions. Indeed, one cannot study the in vestment performance of individual common stocks, mutual funds, or other
forms of investment, if the goals of investors have not been determined be
forehand.
It seems safe to say that the needs of investors vary greatly, and
that the ways in which one might model the investment process are numerous.
Our approach to this problem is to begin with a conceptual model of the
investment process which has its origins in accepted theories of rational
economic behavior. These theories are generally accepted by most re searchers in finance and economics, so they represent the current state of development in economic theory.
Individual Investment: A Postulated Process
The three classes of elements entailed in any purposive act "are (1)
the decision-maker, (2) a set of alternative actions, and (3) a set of g o a l s . "17 With respect to the investment process, the decision-maker is
17C. West Churchman, Prediction and Optimal Decision (Englewood Cliffs, New Jersey: Prentice-Hall, Inc., 1961), p. 137.
28 the investor, who may be representing a large group of other people, or who may be investing his personal resources. The set of alternative courses of action consists of the set of all common stock portfolios that each investor might select. Portfolio selection involves decisions as to whether or not certain securities should be included in the portfolio, as well as decisions regarding the proportional dollar amount to be invested in each security to be included in the portfolio. The set of goals con sists of an operational specification of the investment results each inves tor wants to achieve.
The three elements of the purposive act of investment are made opera tional by describing the investment process, as we perceive it, in terms of six major elements: the formation of investor goals, the recognition of constraints on the investment process, the creation of an algorithm for selecting a portfolio of securities, the search for and description of ac ceptable securities, the application of the algorithm to the set of poten tial security investments, and the selection of an efficient set of securi ties. The fourth step, a search and description of securities, is the focus here, but that step cannot be discussed independently of tha three preceding ones.
Investor objective: step 1 .— Imputing a specific goal to an individ ual's behavior, while maintaining some ..relevance to the behavior of actual investors, is a difficult task. At the present time, we know very little about the collective values of man., and very little about the collective 18 values of the subset of man which we shall label common stock investors.
1 fi See Churchman, p. 3. A major portion of Churchman's book is de voted to an analysis of the methodology needed to ascertain personal values. 30 *
However, in order to measure performance, which is goal achievement, we must know more about the objectives of investors with respect to their in vestment activities.
We assume a goal for investors, which is based on accepted economic theories about man's behavior. In particular, w e ‘assume that investors maximize expected utility, where expected utility is a function of the ex pected rate of return on common stock investments, and the uncertainty of that expected return as measured by the variance of its probability density function. As is argued below, these assumptions made about investor goals are customary in studies of investment performance. Although widely used, these basic assumptions do limit the study and preclude generalization of the results obtained here to all investment activities, or to all common stock investors.
We begin by accepting the expected utility maxim as a proper descrip tion of rational behavior. In this context, the rational investor "acts as if he (1) attaches numbers (utilities) to each possible outcome and (2) 19 chooses that option (or strategy) with the largest expected value." Axiom- 20 atic presentations of the expected utility maxim are presented by Markowitz, and von Neuman, 21 while a more philosophical discussion is presented by
- 22 Churchman.
19 Michael C. Jensen, "Risk, The Pricing of Capital Assets, and the Evaluation of Investment Portfolios," The Journal of Business, Vol. XLII, No. 2 (April, 1969), pp. 167-247.
^ H a r r y M. Markowitz, Portfolio Selection; Efficient Diversification of Investments (New York: John Wiley and Sons, 1959).
21John von Neuman and Oscar Morgenstern. Theory of Games and Economic Behavior (Princeton, New Jersey: Princeton University Press, 1953). 31
Next, we assume that theinvestor maximizes a single objective func tion which is of the form
(2.0) U = f (C^, Wjj C2, W2; ...; Cn , Wn , E), where
U is utility,
C_£ is the dollar amount of consumption during year i
(i = 1> 2, •«•, n),
W^ is the unconsumed wealth remaining at the end of period i
(i = 1, 2, ..., n),
E is the net cash equivalent value of the investor's estate
at the time of death, and
n is the expected or probable maximum life of the individual
investor.
We ignore institutional investors to the extent that they do not have ob jective functions of the form of 2.0 or are investing for a set of indi vidual investors. Individual investors can choose mutual funds or trust managers based on their beliefs about the ability of these agents to maxi mize their (the investors') personal utility functions.
Equation 2.0 is formulated to emphasize the interdependence of con sumption and wealth. Increased current consumption must decrease wealth, which will usually decrease future increments in wealth, and thus decreases potential total lifetime consumption. Regardless of how the investor makes the choice between consumption and investment, he must maximize annual in vestment return in order to maximize total life-time consumption plus estate value, whether or not they are discounted values. In other words,
22 Churchman, Chapter 8. 32 to maximize equation 2.0, the investor must select the most profitable set of investments available, assuming that utility is proportional to the present value of future consumption expenditures and estate value. The ap parently divergent consumption and investment objectives of investors can be recognized in at least two ways. A multiple goal .structure can be for mulated, or the investor's noninvestment objectives can be treated as con straints on the investment process. The latter approach is adopted here because the more complex goal structure is beyond the scope of this paper.
Consumption, once decided upon, becomes a constraint because it limits the amount of money available for investment.
At this point some comment about the maximization assumption is use ful. It has been stated that "Most human decision-making, whether indi vidual or organizational, is concerned with the discovery and selection of satisfactory alternatives; only in exceptional cases is it concerned with 23 the discovery and selection of optimal alternatives." The assumed ob jective of an investor can be tentatively summarized as follows: given some dollar amount of cash to be invested, and a set of potential security investments, the investor will select a portfolio of securities which will maximize his utility, where his utility is a function of the expected re turn from the portfolio and the uncertainty associated with that return.
This assumed objective is not necessarily inconsistent with the observation that man seeks satisfactory alternatives, not optimal alternatives. For non-professional investors, the cost of searching for good investments may be very high, which would limit his search to those popular securities for
9 ^ James G. March and Herbert A. Simon, Organizations (New York: John Wiley and Sons, Inc., 1958), pp. 140-141. 33
which information is readily available. The same is true for the profes
sional administrator of the investment funds of an organization, except
that the list of securities that are evaluated is much larger.
As stated, the wealth maximization assumption is an accepted represen
tation of actual investor behavior. The assumption includes no normative
statements about how much money an investor should commit to investment in
securities in order to achieve some lifetime objective. We need only to
observe that people and organizations do invest in corporate securities.
More importantly, the knowledge that investors only consider some small
subset of all investment alternatives does not affect the assumption as
stated. We only assume that once some set of securities is included in the
investment process, the investor will prefer greater utility to less util
ity. In other words, man may be a satisficer with respect to his search
for alternatives, because of perceived cost and time constraints, but given
a set of alternatives, there is no reason to suspect that he will not max- 24 imize utility.
The assumption that investors attempt to maximize expected utility
must be followed by a specification of the general form of a utility func
tion. Since we have regarded current consumption as a constraint on in
vestment, we can reduce equation 2.0 to the form
24According to Jensen, p. 172, Fama has shown that an investor will appear to behave as though he is maximizing E |. U (Ct» w t+j)j even though he is faced with the multi-period consumption-investment decisions indi cated in 2.0. Therefore, the assumption of period by period utility maxi mization seems justified. Eugene Fama, "Multi-Period Consumption - Invest ment Decisions," Report No. 6830 (Chicago: University of Chicago, Center for Mathematical Studies in Business and Economics, June, 1968). 34
(2.1) U = f (Wj.)j where W^ is defined as before. To avoid scale problems, 2.1 is restated in terms of the rate of return on beginning wealth, plus 1.0, which is equal to and is denoted R^. The reformulation is:
(2.2) U = f (Rt), where
R t ’
Restating these formulations in terms of expected values, we have:
(2.3) E [u (Rfc)] = the expected utility of the price relative, Rfc.
The investor is assumed to select investments so as to maximize E ^ U (Rj.)]
As before, we are faced with the need for information about investors that is not directly available. We need to know how investors react to uncertainty and how they interpret the meaning of uncertainty.
The usual method of incorporating uncertainty into an investment de cision is to assume that the rate of return or the price relative is a random variable for which possible values can be described in terms of a probability density function. If the probability density function is
Gaussian, then the first two moments of the density function can be used to specify all of its higher, non-zero moments. In such cases, it is ap propriate to assume that the utility function of the investor is a function of the mean and the standard deviation of the probability density function of the return from a particular security, or from any linear combination of 25 security returns (i.e., a portfolio).
25 Eugene Fama, "Risk, Return, and Equilibrium: Some Clarifying Comments," Journal of Finance (March, 1968), pp. 29-40; William Breen, "Homogeneous Risk Measures and the Construction of Composite Assets," Journal of Financial and Quantitative Analysis, Vol. Ill, No. 4 (December, 1968), pp. 405-413; and J. Tobin, "Liquidity Preference as Behavior Towards Risk," Review of Economic Studies (February, 1958), pp. 65-86. 35
The problem of non-gaussian distributions is discussed in Chapter III. At this point we will only note that extensive use has been made of- the mean, variance model, and that at least some empirical evidence supports its use 26 as a description of risk adverse investors. Many investors may not be risk adverse, or their desire for low risk may change with changes in 27 wealth levels. As a result, we are only concerned with that subset of all risk adverse investors for which a quadratic utility function is ade quate. The general form of such utility functions is
(2.4) U = R - ba, where R is the mean (expected) rate of return, b is a constant which reflects a particular investor's aversion to risk, and a is the standard deviation of R.
The measures of uncertainty discussed in this study are incomplete whenever the deviations of actual return from expected return are not rep resented by a Gaussian probability density function, and when the inves tor's utility function is not quadratic. In theory both conditions must hold for the under-specification to exist. As a practical matter, inves tors may not be influenced by small deviations from either condition.
Constraints on investment: step' 2 .— For purposes of this study, two constraints are imposed on the investment process. One constraint is the
See Markowitz; Fred D. Arditti, "Risk and the Required Return on Equity," Journal of Finance, Vol. XXII (March, 1967), pp. 19-36; and Kalman J. Cohen and Jerry A. Pogue, "An Empirical Evaluation of Alternative Portfolio Selection Models," Journal of Business, Vol. XXXX, No. 2 (April, 1967), pp. 166-193. Many of the sources referenced in the bibliography in corporate the variance or standard deviation as the measure of uncertainty. 27 Milton Friedman and Leonard J. Savage, "The Utility Analysis of Choices Involving Risk," The Journal of Political Economy, Vol. LVI, No. 4 (August, 1948). 36
limitation of investment periods to one year. This fixed investment period
constraint was made necessary by the decision to consider the problem of
prediction over time. Each prediction period within a series of prediction periods must include the same interval of time in order to be comparable.
The effect of increasing the investment time period is a separate issue, which should be tested separately.
By using finite investment periods, we recognize that investors do not
accumulate wealth over some very long period, but- invest for some purposes within their own expected life spans. At the present time, very little is known about the typical investment period duration for any investor type,
so the extent to which this assumption deviates from actual practice is un known.
The other constraint is that dividends are held until the end of the investment period or are consumed immediately. In either case, the annual return consists of the cash dividends received, plus the price change that took place during the year.
Evaluation models; step 3 .— Given a set of alternative investment opportunities, and given a well defined utility function, a selection pro cess can be defined so that utility can be maximized. Inputs to the selec tion process include information about the common stock, as well as infor mation about the utility function. If common shares can be described in ratio scale measures, and the utility function can be formalized, then a formal algorithm can be developed to perform the process of selecting an optimal portfolio from a set of investment opportunities. Four of the more refined methods of selecting portfolios are the state-preference approach, the relative strength method, the mean, variability approach, and the 37
dominance theorm.
In the state-preference approach, securities are resolved into distri
butions of dated contingent claims to income, defined over the set of all 28 possible states of the world. Each state of the world is specified and a
probability assigned to that state of nature. In addition, each possible
amount of income from each investment must be specified for each state of
nature that has been specified. Hirschleifer argues that this approach in
volves a more precise statement of uncertainty in that each event is as
signed a probability.
The state-preference approach is not viewed as an empirically relevant
method for portfolio selection for various reasons. First, an algorithm
has not been developed for the multiple investment opportunity case, even
though Hirschleifer has provided an example involving a single commodity
and two mutually exclusive states of nature. Lacking a precise algorithm
for selecting a portfolio, it is difficult to determine exactly what the
information requirements of such an algorithm will be for the multiple in
vestment case. Second, it is almost impossible to conceptualize the com
plete specification of all possible states of nature, and all possible con
tingent claims for each state of nature. With respect to security invest ments, the essence of uncertainty is not our inability to specify outcomes
in advance, but our inability to assign a meaningful probability to each
outcome. Hirschleifer has assumed away this essential element of uncer
tainty in order to formulate his theory. As a result, his theory is of
28 J. Hirschleifer, "Investment Decision Under Uncertainty: Choice- Theoretic Approaches," The Quarterly Journal of Economics, LXXXIX, No. 4 (November, 1965), pp. 509-536. 38
little assistance in the actual portfolio selection process in an uncertain
world.
A second portfolio selection method is based on the relative strength 29 criterion of investment performance as described by Levy. In summary, it
is postulated that stocks performing well in recent market movements will
continue to perform well in the future. Past performance is compared by
ranking the percentage price changes of a set of stocks. The highest posi
tive percentage increase in a period is given a rank of 1, while the
largest percentage price decrease is given a rank of N, where N stocks are
being ranked. This ranking process can be performed over any number of in
vestment periods and over any investment period duration.
Levy did find that high performance stocks do tend to continue to per
form well, and that low performance stocks do tend to continue a record of
poor performance. As is true in any such study, it is possible that his
results are related to the time periods over which the tests were made.
The most important limitation of the study was explained by Levy him
self. No measure of uncertainty was used to determine if greater rates of
return are related to greater levels of uncertainty. Usual hypotheses
about the theoretical relationship of higher rates of return to uncertainty would suggest that high rates of return persist due to greater uncertainty,
and will continue so long as uncertainty is high. Because Sharpe has found
that mutual funds with higher rates of return do tend to have higher stan- 30 , dard deviations of annual return, Levy s results may be entirely
29 Robert A. Levy, "Relative Strength as a Criterion for Investment Selection," The Journal of Finance, Vol. XXII, No. 4 (December, 1967), pp. 595-610. 30 William F. Sharpe, "Risk-Aversion in the Stock Market: Some explained by the relative uncertainty of the stocks selected. High rate of return stocks will continue to command a high rate of return so long as the associated uncertainty of return is high. Low rates of return usually are associated with stability and low uncertainty. In summary, the relative strength criterion of investment performance has not been sufficiently validated, because Levy's findings can be explained in terms of traditional utility theory. Also, the relative strength criterion is counter to the random-walk theory of price behavior which is discussed more fully below.
The relative strength criterion of common stock selection is thus not a theoretical base upon which.to build a selection algorithm, from which we can impute the stock description input requirements.
Although the relative-strength method may not be an adequate criterion for investment selection, Levy's findings do have some significance for the process of price relative prediction. Of the three ratio-type measures of price performance which Levy tested, that price ratio which yields the most consistent intertemporal ranking of companies is the current week's ending stock price divided by the average ending price for the 27 weeks prior to and including the current week. Two hundred companies are ranked weekly for the 260 week period from October 24, 1960, through October i5, 1965.
Levy found that companies tend to maintain a relatively high or low price ratio over the entire test period. The implication of these results is that past price ratios are helpful in predicting future price ratios.
Levy's results have implications for price relative prediction because both price relatives, as defined above, and his price ratios are measures
Empirical Evidence," The Journal of Finance, Vol. XX, No. 3 (September, 1965), pp. 416-422. 40 of relative price increase. Upward price movements result in high price relatives and high price ratios. Similarly, downward price movements re sult in low price relatives and low price ratios. Therefore, Levy's find ings are expected to hold for price relatives, as well as for price ratios.
Although we do not test Levy's proposed method of stock selection, the naive prediction models which are evaluated in this study are based on the hypothesis that past price relatives are useful in predicting future price relatives.
The mean, variability approach to portfolio selection is the more 31 fully developed method currently being proposed. The expected utility of a common stock's return is assumed to be a function of the expected (mean) rate of return and the standard deviation of the rate of return. The exact trade-off between return and uncertainty is a function of each investor's utility function.
Operationally, the mean, variability approach requires information about the expected rate of return of each stock, the standard deviation of that rate of return, and the covariance between each stock's"rate of return and every other stock's rate of return. Theories about equilibrium condi tions in capital asset markets, as well as empirical studies of stock prices and rates of return, indicate that investors must accept higher
31 Some of the major works in this approach include: . H. Makower and J. Marschak, "Assets, Prices, and Monetary Theory," Economics N. S., Vol. V (1936), pp. 261-288; Harry M. Markowitz, Portfolio Selection: Efficient Diversification of Investments (New York: John Wiley and Sons, Inc., 1959); James Tobin, "Liquidity Preference as Behavior Toward Risk," Review of Eco nomic Studies, Vol. XXV (1958), pp. 65-86; and more recently, D. E. Farrar, The Investment Decision Under Uncertainty (Englewood Cliffs,. N.J.: Prentice- Hall, 1962). A recent contribution of ideas is provided by William F. Sharpe, "Capital Asset Prices: A Theory of Market Equilibrium under Condi tions of Risk," Journal of Finance, XIX (September, 1964), pp. 425-442. 32 levels of uncertainty in order to earn higher rates of return. Markets may temporarily deviate from equilibrium conditions, but the tendency is to move toward equilibrium. Any one investor may be able to find an invest ment which is expected to yield a higher rate of return, with no increase in uncertainty. However, large numbers of investors are not able to do so.
Using the mean, variability approach to portfolio selection insures that the level of uncertainty (as measured by the standard deviation) is mini mized for whatever rate of return is desired. A basic assumption of this method involves the use of the standard deviation of rate of return as a measure of uncertainty. For a Gaussian probability density function, the third moment about the mean and all higher moments are either equal to zero, or can be expressed in terms of the mean and variance. As a result, the mean and variance (standard deviation squared) completely describe the den sity function, and no other information is needed.
If the probability density function of the rate of return is non-
Gaussian, then one must know all of the moments, or at least all of the non zero moments of the probability density function in order to completely spe cify the probability density function of the rate of return. Samuelson has quite properly argued that one ought to consider the first four moments of the probability density function of each security's rate of return in any 33 portfolio selection model. He concentrates on the first four moments
■^ama, "Risk, Return ..., pp. 29-40; Sharpe, "Risk Aversion ..., pp. 418-419; and Sharpe, "Capital Asset ..., pp. 425-442.
^3?aul A. Samuelson, "General Proof that Diversification Pays," Journal of Financial and Quantitative Analysis, Vol. II, No. 1 (March, 1966), pp. 1-13. 42 *
because it is difficult to explain in non-mathematical terms why an inves
tor should be interested in the tenth moment about the mean, whereas the
second, third and fourth moments — which represent variability, skewness
and kurtosis — can be explained in a layman’s terminology. Another major
factor for considering only four moments of a stock's rate of return proba
bility density function is the limitation of data. Higher moments tend to
be unstable in repeated sampling experiments, which implies that they are 34 of little value in the portfolio selection problem.
Because we know very little about the form of investor's utility func
tions, it is discouraging to argue that we need to know all n-moments of a
probability density function in order to maximize any investor's n-degree
utility function. The dominance theorm is presented as an alternative
method of ordering uncertain prospects, which requires relatively little 35 information about an investor s utility function for uncertainty. To
summarize, the probability density function g is said to dominate a second
probability density function f, if and only if, the cumulative density
function of g (denoted G(x)) is less than or equal to the cumulative den
sity function of f (denoted F(x)), that is: G(x) < F(x) for all xel, the
strict inequality holding for at least one xel, where x is the value of a
random variable X, and I denotes the range of values taken on by X. In
34 Paul G. Hoel, Introduction to Mathematical Statistics (Second edition, New York: John Wiley and Sons, Inc., 1954), p. 48. 35 Josef Hadar and William R. Russell, "Preference Ordering and Stock- astic Dominance" (unpublished Working Paper No. 17, Case Western Reserve University, November, 1968); M. K. Richter, "Cardinal Utility, Portfolio Selection and Taxation," Review of Economic Studies, Vol. XXVII (1960), pp. 152-166; and J. P. Quirk and R. Saposnik, "Admissibility and Measur able Utility Functions," Review of Economic Studies, Vol. XXIX (1962), pp. 140-146. 43
the investment case, we are saying that one security is generally preferred
over another, if higher rates of return are more probable than lower rates
of return, relative to the second investment security.
In theory, the dominance theorm is more general and thus less restric
tive than the mean, variance or moment approach because it holds for all
concave (monotonic) utility functions. Theories of risk aversion imply
that utility functions of investors are usually concave, and this issue can
be tested in the laboratory. There is little corresponding evidence to
suggest that we will ever be able to specify an investor's utility function
to the degree required by a strict interpretation of the mean, variance
theory of portfolio selection when extended to include the first four mo ments of the probability density function.
Nevertheless, the dominance theorm of stock selection is the more
troublesome approach in an empirical sense, because one must know what the probability density function of one security is in order to determine if it dominates another density function. At the present time, the solution to this problem is not clear.
Both the mean, variance approach and the dominance theorm require knowledge about the probability density function of each security's rate of return. Empirical methods of determining probability density functions in volve goodness-of-fit tests or use of sample estimates of the probability density function moments. In this study, sample data is examined and pre sented as historical evidence upon which an investor could form subjective estimates of probability density function moments. Due to the limited amount of data for each firm, only expected return and the variability of return are evaluated. To measure skewness and kurtosis with small sample 44 evidence would be a waste of time and effort because the results would be highly suspect.
Search for and description of acceptable securities: step 4 .— The process by which investors become aware of potential investment opportuni ties is called the search process. At some point in time, an investor be comes aware of a particular corporation's stock through a broker's tip, some financial news, or various other types of communications. Once aware of a particular stock, many courses of action are open to the investor.
The initial set of information elements transmitted to the investor may be sufficient for him to decide to buy the stock, or to reject it. If the initial set of information is not sufficient for the investor to make a decision, then more information must be collected.
The search process is a continuous one in the sense that some initial awareness of a security will lead to more and more information, until a de cision can be made. If the decision is to purchase the stock, then the stock's performance must be reviewed periodically. Current rejection of a security may lead to either permanent rejection or to another evaluation at a later date.
Because we have accepted the mean, variability method of selecting portfolios as a starting point, the primary items of information needed for stock selection are the expected price relative and the associated standard deviation of the price relative. We shall now turn our attention to the proper specification and measurement of these two items of information. CHAPTER III
UNCERTAINTY AND PREDICTION MODEL ERROR
\
Uncertainty is defined as the extent to which the future can be speci fied in the present. In this case, uncertainty is the extent to which a common stock's price relative can be predicted or specified one year in ad vance. We are concerned with an objective dimension of uncertainty, which is the historical predictability of a common stock's price relative over a series of one year periods. The statistic used to measure historical pre dictability is the mean-squared difference between the predicted annual price relatives and the actual price relatives subsequently observed over a period of ten years.
A brief review of some traditional notions of uncertainty is presented as a prelude to our own analysis of the concept of uncertainty and the ap propriate measure of uncertainty.
Uncertainty: Some General Notions
Investment is the process of exchanging a known amount of cash for the right to receive a greater amount of cash in the future. When the amount of cash to be received during a future period is known at the time of the exchange, the investment is said to take place under conditions of certain ty. When many different amounts of cash return are possible, investment is said to take place under conditions of uncertainty, because the actual rate of return that will be earned over some predefined period of time frequently
45 46 will not be equal to the rate of return that was expected at the start of
that investment period.
A dictionary definition of uncertainty is the "quality or state of be- 36 ing uncertain; lack of certainty; doubt..." Conversely, certainty is de
fined as "that which is certain or sure; the truth; the fact; also a cer- 37 tain account. A certain or definite number or quantity." Such comments and definitions illustrate the vagueness of the concept of uncertainty which leads to a variety of possible measures of uncertainty.
Weston has said that "reality is a continuum of cases running from cer- 38 tainty to increasing degrees of uncertainty." For discussion purposes, authors have frequently divided this continuum into two or more categories of cases according to the quality of information we have about each case.
In this context, information quality is synonomous with the degree to which probabilities can be objectively specified. The objectivity of a discrete probability estimate is defined as the variance of the probability estimates about the true probability. When the true probability is unknown, the ob jectivity of a discrete probability estimate is defined as the variance of a set of estimates about the arithmetic mean of all such estimates. For continuous probability density functions, the objectivity of the probabili ty estimates is defined as the variance of the probability density function parameter estimates about the true parameters, or about the arithmetic mean
•^Webster's New International Dictionary of the English Language (2d ed.; Springfield, Massachusetts: G. and C. Merriman Company).
Fred Weston, "The Profit Concept and Theory: A Restatement," The Journal of Political Economy (April, 1934), pp. 154-155. 47
of all estimates of each parameter where the true parameters are unknown.
Knight for example, is frequently quoted for distinguishing between 39 risk and uncertainty. He uses the term risk to refer to those situations
for which the probabilities of alternative outcomes are known. In general, probabilities may be known because of some physical characteristics of the experiment, or by empirical calculation of the relative frequencies of al
ternative outcomes for a large number of experiment replications. The physical characteristics of. a die imply that the probability of any number appearing on the top surface is 1/6 when the die is thrown, because each die has six faces and each face should have an equal chance of appearing on top if the die is square and its center of gravity is at its geometric cen ter. For those outcomes which do not have their probabilities implied by some physical characteristics of the experiment, as does the die, the rela- 40 tive frequency approach can be used. The relative frequency of an out come is the proportion of all trials on which each particular outcome oc curs. The important point is that the same experiment must be repeated a large number of times in order to determine the relative frequencies of the various possible outcomes. If each experiment cannot be repeated, then the relative frequency approach has less validity because the resulting rela tive frequencies relate at best to a set of similar, but not equivalent, experiments.
•^Frank H. Knight, Risk, Uncertainty, and Profit (Boston: Houghton Mifflin Company, 1921), pp. 19-21.
^®The experiment is the throwing of the die, and the event or outcome is the integer value of the top surface of the die. In Knight's terms, the outcomes of an experiment are said to be uncer
tain if the probability of each of those outcomes is not known The un
certainty case is usually associated with those events and activities for
which we have little experience. Without such experience, one may have
little basis for evaluating the probabilities of various possible outcomes,
and so the outcomes are regarded as uncertain. All of the possible outcomes
may not even be known, which creates another form of uncertainty.
Marshak divided the continuum of the cases running from certainty to 42 increasing degrees of uncertainty into four classes. These classes are:
1. Incomplete information, nonstochastic case,
2. Incomplete information, stochastic case,
3. Complete information, stochastic case, and
4. Complete information, nonstochastic case.
As used here, information is the ability to specify the alternative out
comes. All possible outcomes are assumed known in the complete information
cases, while all of the outcomes cannot be specified in the incomplete in
formation cases. The term stochastic implies that the probabilities of the defined outcomes are specified, or are somehow known. The probabilities of alternative outcomes are not known in the nonstochastic cases.
The distinction between risk and uncertainty made by Knight is impor
tant to us because it is frequently quoted in economic studies of invest ment under conditions of uncertainty. Likewise, the more modern terminolo gy incorporated by Marshak is frequently used in conjunction with discus-
42 Jacob Marshak, "Role of Liquidity under Complete and Incomplete In formation," American Economic Review, Suppl. XXXIX (May, 1949), pp. 183-184. 49
sions of the investment decision under uncertainty. These frequently used
terms have special meaning to many people, so our use of the terms must be
made clear, and must be contrasted with traditional definitions.
Using Marshak's terminology, we have complete information with respect
to the possible rates of return that can be earned on investment in a com
mon stock. All the possible rates of return can be specified in advance if
we are willing to aqcept the existence of an upper bound on market price
per share. Given an absolute upper bound on price, a rate of return can be
calculated for each possible combination of cash dividends and ending mar
ket price because the number of possible combinations is finite. If no up
per bound on ending market price existed, then the number of possible rates
of return would not be finite, and one could never specify all possible
rates of return. Even with an upper bound on price, the number of possible
rates of return is very large. Although it is very likely that investors will disagree about the probable maximum price of any common stock share,
agreement on a very high upper bound is plausible. As a practical matter,
then, we are able to specify the possible rates of return on investment in
any common stock, with few errors.
Specification of the probability of each rate of return outcome is
another matter. We argue that it is very difficult, if not impossible, to
specify in an objective manner the probability of each rate of return out
come. The observation that the investment services, accountants, research
ers in finance, and investors do not assign a probability measurement to
each possible rate of return outcome for each common stock serves as prima
facie evidence for this point of view. 50
If we assume that the probability of each possible rate of return can not be objectively specified, then what approach does one take? We shall suggest a specific approach to the uncertainty measurement problem after first discussing a more philosophic methodology question.
At least two approaches to the problem of uncertainty specification and measurement can be visualized. One approach is to describe and define a concept of uncertainty, and then construct a measure of uncertainty which is consistent with the concept. A second approach is to select a measure of uncertainty, and then define uncertainty in terms of that measure. The approaches are similar in that the end result is a concept of uncertainty and a measure of uncertainty which relates to the concept.
Neither approach can be judged correct or incorrect by logic alone.
The theoretical correctness of a measure of uncertainty is a matter of the / *5 collective judgement of the relevant scientific community. This situ ation prevails regardless of the method used to define and measure uncer tainty. Usefulness is another criterion by which we might evaluate the relative merits of different concepts and measures of uncertainty.
Neither the usefulness criterion nor the criterion of general accep tance by a scientific community are helpful in the initial evaluation of a proposed change in current practice or theory, which is our purpose. Our approach consists of three stages: (1) the definition of uncertainty in terms of predictions, (2) arguments in support of the particular definition adopted, and (3) the specification of a measure of uncertainty which is consistent with the definition proposed. Because our definition of uncer-
^ S e e Churchman, Chapters 4 and 5. tainty is developed within the context of a particular conceptualization of
the investment process, it is not being presented as the only concept of uncertainty which is relevant to the general process of investment.
Uncertainty and Prediction
The first step in any analysis of uncertainty is a specification or description of the subject of a prediction. The subject of a prediction is 44 the function which determines the description of an event. Consistently, an event is being predicted because it is uncertain, and that event must be
described in some manner so that we can verify its occurrence. In this
context the subject of prediction is the common stock price relative, which is defined below.
The price relative as a measure of return.— Earnings on investment can be measured in a variety of ways. Some of the more frequently used measures are the rate of return on investment and the price relative, which is a transformation of the rate of return on investment. The rate of re turn is the rate of increase in wealth over some period of time. The an nual rate of return represents the relationship of wealth at year end to wealth at the start of the year. For example, let
W± be the wealth that exists at the start of year t, for invest ment i, and let
W. be the wealth that exists at the end of year t, for investment i. i > t The rate of return on investment i over period t is denoted R OR^, and is defined such that:
44S. A. Ozga, Expectations in Economic Theory (Chicago: Aldine Pub lishing Company, 1965), Chapter 1. 52
(3.1) (1 + EORlt) - W1>t.
In the case of common stock investments, the wealth associated with invest ment i at the start of period t is N. „ , • P. t where N. , is the x,t-l i,t-l’ i,t-l number of shares of security i owned at the start of period t, and P i»t—1 is the price per share of security i at the start of period t, which is to say:
(3.2) W. . = N. . -P. , . i,t-l i,t-l i,t-l
Most common stocks pay cash dividends during each year and intermittently change the number of shares outstanding through stock splits and stock div idends. Cash dividends become a part of stockholder wealth and may be in corporated into the wealth calculation in a variety of ways.
One approach is to assume that dividends will be reinvested immediately upon receipt in fractional shares of the same security. The number of shares of security i owned at the end of period t is thus N. + F. , x , t x , t where F. is the decimal number of shares of security i that the investor is able to purchase during period t with his period t cash dividend pay ments. Another approach is to assume that cash dividends are held until year end and thus are reinvested in the same security at the start of the next investment period. A third approach is to assume that cash dividends are reinvested in whatever securities seem to be the best buy at the time the cash dividend is received.
The first approach of assuming immediate reinvestment upon receipt of cash is not used in this study because the price and quarterly dividend data necessary to do so are not readily available. The difficulty and cost of collecting this data does not seem to be justified in terms of the ex- pected affect on the results of this study, and the nature of the alterna
tive assumption that is made. The third approach of assuming that cash
dividends are invested in the most profitable alternatives available at the
time of cash dividend receipt is probably the best representation of the manner in which portfolios are actually managed. However, the performance
of individual investments cannot be evaluated when cash flows are reinves
ted in other securities. Because we are concerned with predicting the price relatives of individual securities, the approach of assuming rein vestment in other securities cannot be used in this study.
It is assumed that cash dividends are held until year end, or are used for consumption purposes. As a result of this assumption, end of period wealth, denoted W. , is defined as i,t’
(3.3) W. = N. . • (D. . + P. .), where D. . i s the dollar amount i,t i,t i,t i,t i,t of cash dividends per share paid during year t. As before, N. denotes the number of shares owned at the end of year t.
The term N. . in equation (3.2) is equal to the value of N. in 1 y C 1 1 j t equation (3.3) whenever there are no changes in the shares owned by the investor. The two events leading to a restatement of the number of shares owned, are stock dividends and stock splits. Simple increases in the num ber of shares held by any particular investor due to purchases do not af fect our rate of return calculations, but stock dividends and splits do af fect the relationship of the number of shares held at period start to the number of shares held at period end.
First, consider the case of stock dividends. If a 2% stock dividend is distributed during period t, then 54
(3.4) N. = 1..02 . N. j, assuming no other changes for security i X , t X ,t—I have taken place. Because this study is a historical analysis of stock prices, our adjustment must be in terms of current shares outstanding. So
for each share outstanding at the end of period t, there are 1/1.02 or
.9804 shares outstanding at the end of year t-1 on an equivalent investment basis. The year used as the base period is the most recent year used in
the study, or 1967 in this case. Each previous N. is adjusted so that l , t per share statistics represent a given percentage ownership in the company,
or a common investment base.
Stock splits were the second complicating factor in computing wealth
changes over time. Suppose there is an X for Y stock split, which means
each owner receives X shares of security i for each Y shares held before
the split. The relationships of shares owned at the beginning of the year
to shares owned at the ending of the period is expressed as
(3.5) N. = N. , (X/Y). Thus for each share of security i owned X y C X 5 t — I
at the end of period t, there were Y/X shares owned at the start of the period t, assuming the split took place during period t and there were no
other changes.
The terms N. . and N. are adjusted whenever there is a stock split X)t*l 1)t or a stock dividend. Hereafter, no explicit mention will be made of this I adjustment because it will always be made, and the adjustment accounts for
the fact that N. may differ from N. ,, for any i and t. X , t X)t"X Restating the rate of return formula, we have
W. - W. . (3.6) ROR. = ^ --- , which also can be restated as ’ Wi,t-1 55
w t (3.7) ROR. = 77-*------1, where (3.7) is the computational form. X,t i,t-l
The price relative for a security i over some period t is defined as
W i t (3.8) R. = t —, W. . > 0. Notice that R. = 1 + ROR. . i,t w_^ ^ >t i ,t
Due to the limited liability feature of corporations, stock market prices are non-negative, even in the case of liquidation. Because W. 1»t is almost always greater than zero, the price relative is almost always greater than zero. In contrast, the rate of return is negative when the price decline exceeds the value of dividends, which frequently happens when investment periods of one year are used. Price relatives are more conve nient to process because of this non-negativity feature. The basic measure of return on investment used in this study is the price relative as des cribed in (3.8).
The prediction aspect of uncertainty.— An investor can estimate the current dollar value of his common stock investments by referring to stock market prices or other evidence of the net cash price of each investment.
For each common stock, the valuation process usually involves multiplying the ending market price per share by the number of shares currently owned.
Total wealth is then determined by adding the dollar value of all the indi vidual common stock holdings.
Suppose the aforementioned investor would like to know something about the dollar value of his wealth at the end of the current calendar year.
Because stock prices and cash dividends change over time, the investor's wealth may assume many different dollar values by year end. Since many 56
values of his wealth are possible, the investor either must predict the
most likely ending value of his wealth, or he must describe the possible
wealth values in terms of a probability density function. He may predict
total wealth directly, or he may predict the ending dollar value of each
common stock and then combine the results to construct a prediction of the
dollar value of total wealth.
; In Chapter II we established that investors need to predict the future
value of wealth in order to decide upon current consumption levels. The
need to predict the dollar value of one's wealth at the end of one year is
thus built into our model of investor behavior.
The need to predict future stock prices and cash dividends is actually
greater than the above comments imply. If an investor wants to maximize
the wealth-consumption function described in 2.0, then he must invest in a
combination of the most profitable investments available. The most profit
able investments are those for which next year's price relative will be
greatest (we have assumed a one-year investment horizon). Obviously, one
must formulate some predictions about each common stock's price relative
for next year in order to make a selection of the most profitable ones.
To summarize, an investor must predict future price relatives in order
to evaluate investment alternatives — especially if an attempt is being made to maximize or increase rate of return on investment.
To note that an investor must predict price relatives over some in vestment horizon is only a start. Next, we must consider the nature of
these predictions and the resulting prediction errors.
When only two outcomes are possible, predictions are either correct or
incorrect. The football predictor either picks the eventual winner of the 57
Super Bowl, or he does not. Likewise, a baseball fan either correctly
picks the winner of the World Series, or he does not.
For common stock investors the notion of correct versus incorrect pre
dictions is of little relevance because price relatives are seldom predic
ted in advance. Since each price relative can take.on many different val ues, it is unlikely that any investor will be able to predict which value will actually be earned over the next year. However, a skilled predictor may make predictions that are nearly correct in the sense that the predic
ted price relatives are close to the price relatives subsequently earned.
It is this closeness between predicted price relatives and observed price relatives which we analyze in more detail.
Uncertainty Defined
In the investment context, uncertainty is defined as the extent to which price relatives cahftbe predicted one year in advance. Historical evidence and expectations about the probability density function of the price relative prediction errors provide the basis for a measure of the extent to which price relatives can be predicted.
The evaluation of predictions through an evaluation of the prediction 45 error probability density function is widely used and accepted.
This definition of uncertainty provides for the estimation of a prob ability density function from historical evidence, or from any other source.
Markowitz has described how a security analyst might crudely structure his beliefs about the probable rate of return that each common stock can be
45 James C. Naylor, "Some Comments on the Accuracy and the Validity of a Cue Variable," Journal of Mathematical Psychology, Vol. 4, No. 1 (February, 1962), pp. 154-161, and Henri Theil, Applied Economic Forecast ing (Chicago: Rand McNally and Company, 1966). 58 46 expected to earn. The beliefs of the analyst reflect the totality of his
experience, which includes any data about the firm which he has absorbed.
The bases of the analyst's expectations are varied and are not well defined.
We are interested in the historical evidence of uncertainty which is
provided by past prediction performance, where the predictions are made by
means of well defined models. Consistently, the remainder of this study is
concerned with the development and application of a measure of historical
uncertainty. The measure will be historical in the sense that it is de
rived from historical data and relationships. Such a measure is considered
important because most estimates about the future start with interpreta
tions of past performance and behavior. Our ability to draw any conclusions
about the confidence we have in our parameter estimates is unknown at pres
ent.
A Measure of Uncertainty
Uncertainty is defined in terms of the probability density function of
prediction errors. Ideally, there would be no difference between the ex-
ante prediction of a particular price relative and the ex-post measure of
that price relative. As a practical matter, however, the predicted price
relative of any frequently traded common stock is apt to vary significantly
from the actual price relative that is earned.
The mean of a series of prediction errors should tend to zero if we
are to minimize prediction errors through constant improvement of the pre
diction model. Given this objective to reduce prediction errors, the best measure of uncertainty is the second moment about zero of the prediction
lif\ Markowitz, Chapter 2. 59
47 errors, which is the mean-squared prediction error.
The mean-squared error is the arithmetic average of the differences
between the numerical value of an observed price relative and the numerical
value that was predicted for that same price relative. The formula is
(3.9) MSE = ln (R - R .. )2 /n, where J1 t—1 JJ-1'
it is the actual, observed price relative for common stock i
over the period t, 'V, R ^ t is the predicted numerical value of as made by pre
diction model j as of the end of period t-1, and
n is the number of periods involved. In this case, n equals
10 (years).
In addition to being useful as a measure of total uncertainty, the MSE
statistic is amenable to division into two component parts. This division
provides a useful basis for analyzing the nature and source of prediction
errors in general. Because this separation of the MSE into component parts
can be expressed in terms of traditional statistical measures, we start with a discussion of prediction errors and the relevant statistical mea
sures of these errors,
First, observed price relatives are compared to predicted price rela
tives in order to obtain a series of differences, D.. . as indicated by Jit’
^ T h e mean-squared error is a measure commonly used in statistics. For example, see W. Allen Wallis and Harry V. Roberts, Statistics: A New Approach (New York: The Free Press, 1956), p. 448. The formulation of mean-squared error is also widely used in various contexts, such as in Yuri Ijiri, The Foundations of Accounting Measurement (Englewood Cliffs: Pren- tice-Hall, Inc., 1967), p. 142; and Miltin Friedman, The Interpolation of Time Series by Related Series (Technical Paper No. 16, New York: National Bureau of Economic Research, 1962), p. 14. 60
'b (3.10) D.. = R.^»- R..^, where j denotes the prediction model used, jit it jxt
i denotes the common stock for which price relatives are being predicted
and t denotes the year. Two rather common sample statistics, the mean and
the variance of the prediction errors, are defined as
(3.11) MN.. = In D . . /n = DD. .., and Ji Lt=1 Jlt: J’i
(3.12) VAR.. = l n (D - 5 )2 /n. JA t=l J
Using the relationship defined in 3.10, formula 3.9 can be put in the form
2 (3 ,13) MSE.. = yn D .. /n. Ji L t=l Jit
These two component parts of the MSE are the mean prediction error squared and the variance of prediction errors about their mean error. For- 48 mula 3.12 can be restated as
(3.i« vae^. - v 2 /- - [rM »jlt /»]2. which means equation 3.13 can be restated as,
(3.15) MSE.. = VAR.. + (MN..)2 . Ji Ji Ji
The mean prediction error, the variance of prediction errors, and the mean-squared error are operational measures of bias, precision and accura cy, respectively. Bias is the extent to which a prediction model or pro- 49 cess persistently over or under predicts. Excessive overprediction im- 'V plies that Rjit will exceed R^t more frequently than not, or that large
48rhe reason for using the biased estimate of o^2 as shown in 3.14 is explained in Appendix A^ Actually the right side of equation 3.15 is an unbiased estimate of o , or MSE... e ji ^Naylor, pp. 155-156. 61
prediction errors are more often associated with overprediction than with
underpredictions. Conversely, excessive underprediction implies that IL 'b will exceed more often than not, or that the larger prediction errors
made by the model are associated with underpredictions more often than with
overpredictions. Due to the formulation of the difference term (3.10) ex
cessive overprediction results in a negative average error, and underpre
diction results in a positive average error.
The term precision has been used by Cochran and Cox to denote the
closeness with which a measurement approaches the average of a long series
of measurements, made under similar conditions.^ In this study, the term
precision shall be used to denote the closeness with which a prediction
error approaches the average prediction error over the: time period during which the model was operated. The difference between the two uses of the
term precision is in the set of conditions under which the measurements or
predictions take place. In the context of measurement described by Cochran
and Cox, repeated measures are being taken of a particular characteristic
of a person or an object, ceteris paribus. In our prediction context, con
ditions are changing and these changing conditions are only partially rec
ognized in the prediction model. Therefore, one has a series of'single ob
servations in the prediction case, whereas one has repeated observations within a single time period in the measurement case. The basic idea that
the measurement (or prediction) differs from the true value by some error
is common to both cases, however, so the term precision is adopted for use in our context of prediction.
G. Cochran and G. M. Cox, Experimental Designs (2d ed.; New York: John Wiley & Sons, Inc., 1957), p. 16. 62
As mentioned before, the operational measure of imprecision is the variance of prediction errors about the mean prediction error. Precision is thus unaffected by any bias that exists in the series of prediction errors. This implies that a prediction model can result in precise pre dictions which may or may not be biased predictions. Likewise, a predic tion model can result in relatively imprecise predictions which may or may not be biased predictions.
Accuracy is the term used to signify the closeness with which the pre dicted value of a price relative approaches the numerical value actually obtainedAccuracy is thus a measure of the extent to which prediction is achieved, since the ultimate goal of prediction is the minimization of differences between predicted price relatives and the actual price rela tives subsequently obtained. Note that accuracy can be determined without regard to the source of the predictions.
The accuracy of a set of predictions is reflected by the mean-squared error of the resulting prediction errors. More precisely, formula 3.13 is 52 a measure of the inaccuracy of a set of predictions. The term inaccuracy is used because the value of M S E ^ increases as inaccuracy increases, and this relationship is easier to comprehend than the inverse relationship that exists between the numerical value of and accuracy as defined.
As indicated in formula 3.15, bias and imprecision both contribute to
51 This definition follows a parallel definition of accuracy in mea surement by Cochran and Cox, op. cit., p. 16, and is the same definition used by Naylor, p. 156, in the context of the predictive performance of a cue variable.
■^Naylor is-responsible for this use of the term inaccuracy. Ijiri, p. 143, uses essentially the same formulation to define reliability in accounting measurements. 63 prediction errors. Because the mean-squared error is increased by the mean error, squared, it becomes quite apparent that finding unbiased predictions
is very important. Moreover, formula 3.15 provides an exact indication of
the trade-off that is involved between biased predictions and imprecise predictions. Bias should be reduced if and only if the can be re duced. Likewise, imprecision should be reduced if and only if the can be reduced. Whenever bias and imprecision can be reduced simultaneous ly, there is no trade-off and MSE.. is reduced, of course. . Ji The utility function.— A slightly different form of the usual quadra tic utility function is implied by the definition and measurement of uncer tainty adopted here. The utility function stated in 2.4 is restated as
A A (3.16) U = R - b oe, where R is the predicted price relative, b is a constant which reflects the particular investor's aversion to uncertainty, and is the variance of prediction errors about zero. MSE is the statis tic by which we estimate a ■ e
Price Relative Variance as a Measure of Uncertainty
The range and the standard deviation are two popular measures of un certainty. As a starting point, consider the range of the possible values of a price relative as a measure of uncertainty. Lange has argued that people "need not, and usually do not visualize an exact probability distri- 53 bution of possible prices." Further, people only consider a practical range, which excludes extreme values, he argues. Lange does not provide any theoretical or empirical evidence to support this hypothesis, but his
53 Oscar Lange, Price Flexibility and Employment (Bloomington, Indiana: The Principia Press, Inc., 1952), p. 29. 64
comments serve to indicate at least one notable proponent of the use of
range as a measure of uncertainty.
Suppose that a price relative A has ranged rather uniformly between
.90 and 1.50 over the last twenty years. Suppose that a second price rela
tive, price relative B, has also ranged between .90 and 1.50, but that all
but two price observations are between .95 and 1.35. Clearly, the range of
price relative A equals the range of price relative B. Investment B is
preferred, however, because there is less variability in investment B, than
there is in investment A. The range is not sensitive to any differences
within the range, and thus is a less complete measure of uncertainty than
the standard deviation.
The standard deviation of return per dollar of investment is another 54 popular measure of uncertainty. It is the positive square root of the variance of a variable about its mean value, and thus is a measure of dis
persion.
Under certain conditions, the standard deviation is a relatively com
plete measure of uncertainty. If a price relative were properly described
as a random variable such that each price relative is independent of every other price relative, and the probability density function of the random
See Jacob Marshak, "Money and the Theory of Assets," Econometrica, VI (October, 1928), p. 320; H. Makower and Jacob Marshak, "Assets, Prices and the Monetary Theory," Economica, N. S., V (August, 1938), p. 272; Albert Gailord Hart, "Risk, Uncertainty, and the Unprofitability of Com pounding Probabilities," Studies in Mathematical Economics and Econometrics, 0. Lange, R. McIntyre, and T. Yntema, editors (Chicago: The University of Chicago Press, 1942), pp. 110-118; Donald E. Farrar, The Investment Deci sion Under Uncertainty (Englewood Cliffs: Prentice-Hall, Inc., 1962); H. Markowitz, Portfolio Selection (New York: John Wiley & Sons, Inc., 1959); P. A. Samuelson, "General Proof that Diversification Pays," Journal of Financial and Quantitative Analysis, II, No. 1 (March, 1966), pp. 1-13; and many others. 65
variable is symmetric, then the standard deviation is a relatively complete
measure o£ uncertainty.
Even under these conditions the variance understates the difficulty of
predicting price relatives because the true mean of the price relatives is
unknown. If price relatives are properly described by a random variable
with a stable probability density function, then the best prediction of
each year's price relative is the mean price relative, which is estimated
by means of sample data. Prediction errors will be caused by the variabil
ity of the price relatives as well as by sampling errors in estimating the
mean value of the price relatives.
The problem created by use of sample estimates of the true mean may
not be too great. The standard deviation of the sample means is a function
of the number of observations and the variance of the price relatives
(o = a/ n ) . This implies that the variability of the price relatives, R and the sampling errors associated with estimating their mean, are both
proportional to the standard deviation of the price relative population.
As a result, the standard deviation of past price relatives may be an ade
quate measure of uncertainty under the conditions stated.
There are criticisms of the standard deviation which are much more
basic than the simple issue of sampling error. Suppose price relatives are
not described by a stable probability density function. If a much more
complex stochastic process better describes the price relative data, then
the standard deviation of price relatives over some past period does not
represent uncertainty. It represents variability about a mean, but that
mean may have little relationship to data from other time periods.
I 66
The standard deviation of price relatives should understate uncertain
ty under these circumstances because the time model of the stock price re
latives is unknown, and because the mean price relative for the test period may not have been predictable. It is difficult to speculate as to the exact effect of using the standard deviation to represent the uncertainty of price relatives that are generated by an unknown stochastic model.
In contrast it is conceivable that the price relative variance could overstate uncertainty. In time series data, some trend frequently exists,
and that trend could be used to predict future price relatives. As a mea sure of variability about a mean, the standard deviation calculation ig nores any trend information which is contained in the price relative time series. The variability about the predictable trend may be less than the variability about the mean, which would imply that the standard deviation overstates uncertainty.
In addition to trend, the value of a price relative may be related to other variables. Using these relationships, one might be able to make pre dictions which are highly correlated with the price relative series. The differences between the price relatives and their mean may be greater than
the differences between the price relatives and the price relative predic tions based on other variables, so that the standard deviation would over state uncertainty.
In the next section we present more formal arguments for attempting to predict common stock price relatives in order to reduce uncertainty. 67
Analytical Arguments for Attempting To Predict Price Relatives
In general, prediction is useful if it serves to reduce uncertainty.
The manner in which prediction can lead to reduced uncertainty is dis cussed in this section.
A summary of the argument can be stated as follows: Let R be a ran- 2 dom variable, and let p be the mathematical expectation of R and a be the variance of R. It is assumed that the variance of a random variable is an adequate measure of its uncertainty, so prediction of actual obser vations of R is desirable if the conditional variance of R, given the prediction, is less than the variance of R.
The validity of this statement can be demonstrated by considering the relevant conditional probability arguments. Accordingly, let
Rit a rand°m variable which represents the ratio of the market
price of security i at the end of period t to the dollar amount
invested at the start of period t,
P2£t be the mathematical expectation of and
2 a „ . be the variance of R._. 2it it
Let X. be a random variable for which and R.„_ have the joint prob- lt xt it ability density function • Dropping the subscripts
for convenience, the conditional probability density function
of R, given that X=x, is
(3.17) f(r|x) = f (r,x)/f^(x), f^x) > 0, where f^x) is the marginal probability density function of X.'*’’
-^The following discussion can be found in many standard mathematical 68
By definition.
(3.18) E(r|x) = /^r-f(r|x) dr, and from 3.17
(3.19) E(r |x ) - £ 4r
The conditional expectation of R given X=x, is a function of x alone.
If it is a linear function of x, denoted by g(x), then it can be written
in the form g(x) = a+bx. Expression 3.18 can thus be restated as:
(3.20) E(R|x) = f_l dr = a+bx, or
(3.21) /“r-f(r,x)-dr = (a+bx)*f1(x)
Integrate both members of 3.21 on x to obtain
00 00 00 / rf„(r) dr = af fn(x) dx + b f x f-(x) dx, — 00 2 — 00 — CO 1 which by substitution of symbols, can be restated as
E(R) = a + b E(X), or as
(3.22) ^2 = a + by^.
Multiply the members of 3.21 by x and then integrate on x to obtain
f f x x f(r,x) drdx = af xf-(x) dx + bf x2 f1(x) dx, — 00— 00 ' 7 ' — 00 — 00 X which is by definition,
(3.23) E(XR) = a •E(X) + b-E(X2).
From the facts that E(X2) - [E(X)]2 = a^2 or E(X2) = a^2 + y^2, and
statistics texts. This particular development is from Robert V. Hogg and Allen T. Craig, Introduction to Mathematical Statistics (2d ed.; New York: The MacMillan Company, 1965), Section 2.2, pp. 54-65.
56Hogg, op.cit., p. 63. 69
E(XR) - y . ^ = E[ (X-y^) • (R-y2) ] - Pi2°la2’ exPression 3*23 can be restated as: 2 2 (3.24) P12CTlCT2 + m1P2 = aPl + b ^al + ui
The simultaneous solution of equations 3.22 and 3.24 yields
a 2 °2 b = an^ a = y2 - ~ so ^*20 can be restated as
°2 (3.25) E(R|x) = (y2 - P12 ~ *1^) + Pl2‘cr2/,C^1*X, which reduces to
°2 (3.26) E(r|x) = y2 + ~ * (x-y-^) > t*ie conditi°aal expectation of
R given X=x.
We shall now consider the conditional variance of R given that X=x, for the case where the conditional variance of R is given by
(3.27) E {[R - E (r|x)]2 | x} =
£ > [r - y2 - P12 (x-P^ ] 2 • f (r|x) *dr, or
f 2 t(r-y2) “ P12 (x-y^ ] 2 • f(r,x)*dr
f-L(x)
The variance is non-negative and at most a function of x alone, so f multiplication by f^(x) and integration of 3.27 on x will be non-negative.'
Recall that E[h(x)] = / h(x) f.(x)»dx by definition, so letting h(x) equal ™°° _L the right hand member of 3.27, we have
57Ibid., p. 64. 70
^(r-u2)-p12 (x-)J1)2 • f(x,r)dr j f^x) dx (3.28) E(h (x)) = CC
0 0 0 0 = f j a (a) J „ „ -Joh OOm OO
2
E [«-u2)2j -2p12^ E [tt-UjMR-Pj)] + Pi22 ^ 2 E [«-“l>2]
„ 2 2 0 2 2 2 2 = °2 2p12 o7 p12°la2 + P12 2 °1 1 al
2 „ 2 2 2 2
" °2 12 a2 p12 2
= c22 d - » 122) ; o
The expected value of the conditional variance of R, given X=x, is 2 2 equal to a2 (1-p ). Notice that the higher the correlation coefficient of
X and R, the smaller is the expected variance. Perfect correlation im
plies perfect prediction and a correlation coefficient of 1. If p = 1, 2 p = 1 and the conditional variance of R is zero.
The above formulation is based on the assumption that the costs of
prediction are zero and that uncertainty should be reduced, ceteris pari
bus. The postulation that some investors seek to maximize wealth implies
that if expected return is unchanged, then less uncertainty is preferred
to more uncertainty. Less uncertainty is preferred becasue the probability
that actual wealth will be within ± X% of expected wealth is increased as
the uncertainty is decreased.
* 71
When prediction is not cost free, the extent to which the investor will undertake prediction, or pay for predictions, depends on the utility he has for less uncertainty. The utility he has for less uncertainty is reflected in the marginal rate of substitution of more predictability
(less uncertainty) for expected return.
The utility functions of some investors can theoretically be ex pressed as"^ 2 (3.29) U(R) = y^ - Acr^ , where R is the financial rate of return,
U(R) is the utility of satisfaction derived from earning R,
y^ is the mathematical expectation of R,
2 is the variance of R, and
A reflects the curvature of the investor's utility of money
function.
From this illustrative utility function, we can demonstrate the effect of prediction costs, in a somewhate elementary manner. Expression
3.29 becomes
(3.30) U2 (R) = (y^C) - A(o12 - Aa^2) , or
U2 (R) = (y^-C) - A a ^ - d - A),
where C is a prediction cost per dollar of planned invest
ment and 2 A is the proportional reduction in resulting from pre
diction.
5%onald Eugene Farrar, The Investment Decision Under Uncertainty (Englewood Cliffs, N. J.: Prentice-Hall, Inc., 1962), p. 15; and Milton Friedman and Leonard J. Savage, "The Utility Analysis of Choices Involving Risk," The Journal of Political Economy, Vol. LVI, No. 4 (August, 1948), pp. 279-304. 72
Prediction will take place if and only if ^ ( R ) > U(R). The condi- 2 2 tion ^ ( R ) > U(R) implies ^( R ) - U(R) = AXo^ - C. Because AXo^ - C 2 must be greater than zero, AXa^ > C must be satisfied for prediction to take place.
The discovery and specification of utility curves is beyond the scope of this project, so the determination of prediction costs is of little value. We can point out that as the dollar volume of investment increases, the prediction cost per dollar of investment decreases. This relationship helps to explain the reason for the extensive analysis staffs of large investors as opposed to the rather naive methods of individuals.
We have established that if price relatives are a linear function of a cue variable, then the conditional variance of the price relatives may be reduced. The analysis shall now be extended to show the relationship of predictions to the mean-squared prediction error, which is the measure of uncertainty adopted here. The mean-squared prediction error may be de composed as follows:
(3.31) I (R - R)2/n = (R - R)2 + (S - rS )2 + (1 - r2) S 2 where p A A
R is the mean price relative,
R is the mean predicted price relative,
Sp is the standard deviation of the price relative predictions,
is the standard deviation of the actual price relatives, and
r is the product-moment correlation coefficient of predicted
and actual price relatives.
Substitute the symbol R for a + bx, where x is the predictor or cue vari- 73 able discussed in equation 3.20. Under ideal conditions, predicted price relatives will be equal to actual price relatives, so that the only source of prediction error will be the lack of perfect correlation between x and
A R, (p-^2 < 1) • This statement can be verified by noting that if R = R, then
X A A R = R, and R - R = 0. The statement R = R is a special case of the state- SA SA ment R = b*R, for b* = 1. Because b* = r — , and b* “ 1, we have 1 = r — b b P P and S = rS., or S - rS. = 0. p A ’ p A The first two terms of 3.31 are thus zero, which leaves the term 2 2 S. (1 - r ), which is equivalent to equation 3.28. A We have now shown how the use of a prediction variable will lead to a reduction in the mean-squared prediction error where the linear relation ship between the predictor variable and the variable being predicted is known. The effect of biased predictions (R = a + bR) on total uncertainty is discussed earlier in this chapter and in Chapter V.
Summary
Uncertainty is defined as the extent to which price relatives can be predicted one year in advance. The extent to which price relatives can be predicted is reflected in the probability density function of the predic tion errors. The mean-squared prediction error is used to measure uncer tainty since it is the second moment about zero of the prediction errors.
The variance of past price relatives about a mean is not believed to be a good measure of uncertainty because the variance calculation ignores information contained in the price relative series, as well as ignoring the fact that a cue variable may exist for which the conditional variance of the price relatives may be less than the unconditional variance. More importantly, the use of a mean of past price relatives to predict future price relatives assumes a specific model of stock price relatives which may not fit the data.
The relationship of predictions to the price relative variance and to the mean-squared prediction error are discussed in detail. Having estab lished predictability as the aspect of uncertainty which we wish to study, we now turn to the problem of formulating predictions. CHAPTER IV
THE PREDICTIVENESS OF COMMON STOCK PRICES
AND PRICE RELATIVES
In spite of the large number of common stock selection methods, prac ticing investment analysts do seem to fall into two clusters*^, €ne cluster being the technical analysts who chart stock price movements in order to determine the proper time to buy a particular common stock, and the other cluster being those analysts who examine growth, earnings stability, price/ earnings ratios, and other measures of economic performance in order to evaluate intrinsic value. Intrinsic value is a concept of long-run value to which market price should converge.
The nature of technical analysis is discussed in the introduction to this chapter. In the second section, we review the random walk hypothesis, and then we analyze its implications for price relative prediction in sec tion three. The remainder of the chapter is concerned with the development of prediction models. Three naive models are presented as one approach to the formulation of price relative predictions. They are linear, first de gree equations for which the constants are estimated using exponential smoothing of past price relatives. A fourth naive method is the use of the mean of the prior ten price relatives as the prediction of the next price relative. An economic model which simulates The Value Line Investment Sur vey methodology is developed in the final section of this chapter.
75 76
Introduction
For many years attempts have been made to predict the future prices of individual common stocks, as well as future movements in the various common stock price averages which reflect general market trends. A superficial examination of the prediction problem might lead one to conclude that past prices or price changes should be of some value in predicting future prices and future rates of return. However, two conflicting theories about stock price movements have emerged over the years, and they are of fundamental importance whenever stock price predictions are being considered. Cootner went so far as to say "... it is hard to find a practitioner, no matter how sophisticated, who does not belie^Ie that by looking at the past history of prices one can learn something about their progressive behavior, while it is almost as difficult to find an academician who believes that such a 59 backward look is of any substantial value."
"Technical analysis" is the name given to the process of investigating past stock price movements in order to predict future price movements.
Through a careful examination of past common stock prices, a good analyst can presumably determine what future market movements will be. Because the analyst must be correct in his predictions more often than not, in order to attract a clientele, it is safe to say that a proponent of technical analy sis would argue that price increases or decreases can be predicted more often
59 Paul H. Cootner (editor), "Introduction," The Random Character of Stock Market Prices (revised edition; Cambridge: The M.I.T. Press, 1964), p. 2. 77
than is possible by using a chance model, or by making naive predictions.
Chartists and technical analysts do not argue that they can predict the mag
nitudes of the future changes in common stock prices.
In order to state a theory of technical analysis, we must observe the
practice and note the writings of analysis practitioners because technical
analysis is an art which is practiced by many different types of people,
and not a field of scientific investigation. Because of the current state
of this art, however, the theoretical construct of technical analysis is
not in a generally accepted, testable form. The theory of technical analy
sis that appears below is thus an imputed theory.
Underlying the use of technical analysis is the theory that common
stock prices tend to move in identifiable patterns, and that these patterns
will be repeated in future years. The various forces that affect stock
prices must therefore come into play in specific sequences and in magnitudes
comparable to previous periods. The significant idea is that there are cer
tain stock price patterns which tend to be repeated.
One implication of the technical analysis approach is that meticulous
study of the historical series of a stock's prices and knowledge of the
current price movements of that stock can lead to above average trading pro
fits. Another implication is that prediction models of future common stock
prices should incorporate information about the recent price movements of
the stock, or should include the predictions of a skilled analyst in order
to be complete models in an informational sense.
Due to the lack of objective, well defined estimates made by technical analysts, it is not possible to include such estimates in a formal prediction model. Moreover, some characteristics- of stock price behavior have been 78 tested and the conclusion drawn is that formal prediction models based on past prices are of little use.
The second, and conflicting theory of stock price behavior is that a common stock price series resembles a random walk. The level of a stock's price is not believed to be random, but the price changes from period to period are random and unpredictable. Any trends or complex patterns that appear in the series of past prices are due to a stochastic process, and are not apt to be repeated in the future. If the process generating com mon stock price changes is a random walk, then technical analysis is use less, and there is no need to search for the various patterns that techni cal analysts supposedly look for.
In the context of this project, the problem is to determine if know ledge of past prices is of any substantial benefit in predicting future price relatives (rate of return). If past prices or price changes are of little benefit in predicting future price relatives, then other means must be used to predict future price relatives. Conversely, if past prices or price changes are useful in the prediction of future price relatives, then any model used to predict future price relatives should make use of such knowledge. The method used to evaluate the potential prediction value of past stock prices or price changes is a review of the formal, empirical studies of stock price behavior. The results of this review are presented in the next section. 79
The Random Walk Hypothesis: A Review
Technical analysis persists in stock market circles even though some empirical evidence supports the hypothesis that changes in the price of a common stock behave as if they were generated by a simple chance model.
Roberts gives two reasons for the lack of a "widespread recognition among financial analysts that the patterns of technical analysis may be little, if any thing, more than a statistical artifact.One possible explana tion is that the usual methods of graphing stock prices gives a picture of successive 'levels' rather than of 'changes', and levels can give an arti ficial appearance of 'pattern' or 'trend'. A second is that chance behav ior itself produces 'patterns' that invite spurious interpretations.^ The patterns and trends which exist in historical market price data cannot be used to predict future market prices, unless these same patterns continue into the future. The evidence suggests they do not.
Two heuristic arguments in support of the random-walk hypothesis pro vide an intuitive introduction to the theory. "If the stock market behaved like an imperfect roulette wheel, people would notice the imperfections and 62 by acting on them, remove them." Information about market prices is gen erally available, so every investor has the opportunity to discover consis tent price behavior. General knowledge of the behavior would probably re sult in market activity which would change the pattern, and thus eliminate it.
^Harry V. Roberts, "Stock-Market 'Patterns' and Financial Analysis: Methodological Suggestions," Journal of Finance, Vol. XIV, No. 1 (March, 1959), p. 1.
^Roberts, p. 2. 62 Roberts, p. 7. 80
A second argument which is implicit in Roberts' comments is the more
fundamental one. If a simple mechanical device (chance model) can be cre
ated which duplicates many of the characteristic features of stock-price
movements, then stock-price movements can be described in terms of that
chance model. The best model is that model which duplicates more charac
teristic features of stock-price movements, or does a better job of dupli
cating certain important characteristics. At the present time, no method
or set of criteria has been proposed which would provide a convincing basis
for selecting a particular model as being the best model of stock prices.
The .models of stock price behavior developed to date are not completely
satisfactory because different models seem to fit the empirical data, even
though they conflict in a rather basic way. The stock-price literature is
extensive, so only the better known of these models are discussed and con
trasted.
The random-wallc hypothesis of stock price behavior was first proposed 63 by Louis Bachelier. His model was of the form
(4.0) p(t) - p(t-l) = e(t), t=l,2,..., "where p(t) denotes the price of a
security at time t and e(t) represents a Gaussian process of independent variables, that is, E j e(t)*e (t+T) != 0, T ^ 0, T integer, and for every
64- time tg, e(tg) has a normal distribution." Two basic ideas are involved
in the Gaussian behavior hypothesis of Bachelier. One is that bits of in-
63 L. J. B. A. Bachalier, "Theorie de la Speculation, Gauthier- Villars, Paris, 1900; translated in Paul H. Cootner (editor), The Random Character of Stock Market Prices (revised edition; Cambridge: The M.I.T. Press, 1964), Chapter 2.
^ S. James Press, "A Compound Events Model for Security Prices," The Journal of Business, XL, No. 3 (July, 1967), p. 318. formation are essentially independent over time. Good news does not tend to follow good news or bad news, with any pattern in other words. If the news reaching investors is serially independent, then price changes should be independent also. Secondly, if the news of information is received in a large number of small bits, spread uniformly over time, then by the central limit theorm, the distribution of stock price changes tends to its limiting distribution, the Gaussian or normal distribution. Because the work of
Bachelier was so advanced in 1900, very little testing of his theory ap peared until the 1920's and 1930's.
Working and Slutsky later found that a series generated by summing ran- 65 dom numbers is similar in appearance to an economic time series. * 66 These two works thus provided support for the earlier work of Bachelier, however, empirical work by Cowles and Jones cast some doubt on the independence of price changes. They found a first order serial correlation in the first dif- 67 ferences of some price 4-ndex series. In a presentation and subsequent pa per dealing with the analysis of economic time-series, Kendall also found some evidence of non-random behavior in cotton prices, but the rest of his work supported the random-walk hypothesis of Bachelier, even though that was •68 not the original purpose of this study. Kendall studied twenty-two time
65 H. Working, "A Random Difference Series for Use in the Analysis of Time Series," Journal of American Statistical Association, XXIX, (1934), pp. 11-24. 66 E. E.« Slutsky, "The Summation of Random Causes as the Source of Cyclic Processes," Econometrica, V, pp. 105-146 (1937), translation.
^ k . Cowles and H. Jones, "Some A Posteriori Probabilities in Stock Market Action," Econometrica, V (1937), pp. 180-294.
68M. G. Kendall, "The Analysis of Economic Time Series — Part I: Prices," Journal of the Royal Statistical Society, XCVI, Part I (1953), pp. 11-25. 82 series, and the series of cotton prices was the one series that appeared to be a notable exception to the pattern of random price changes.
Working independently of each other, Work i n g ^ and Alexandei?® discov ered that the use of an average of weekly price observations as a monthly price observation introduces first-order serial correlation in the first differences of a price series. This serial correlation is created even though the original data is a random walk. Cowles noted the validity of this argument and revised some of his earlier conclusions, because some of his indexes involved averaging of observations within his basic period. 71
This new information also explained Kendall's findings with respect to cotton prices because part of that series involved the use of averaged prices whereas the other indexes he studied did not.
An improvement in the basic model was proposed by various writers at about the same time. That improvement is the use of a logarithmic trans formation. The model is thus restated as
(4.1) logg p(t) - loge p (t-1) = e(t), where e(t) is distributed as in 4.0.
The logarithmic transformation is based on the observation of past stock price performance. The natural logarithms of price changes fit a normal distribution better than the price changes themselves do. M. F. M.
69 H. Working, "Note on the Correlation of First Differences of Aver ages in a Random Chain," Econometrica, XXVIII, No. 4 (October, 1960), pp. 916-918.
^Sidney S. Alexander, "Price Movements in Speculative Markets: Trends or Random Walks," Industrial Management Review, II, No. 2 (May, 1961), pp. 7-26.
^Alfred Cowles, "A Revision of Previous Conclusions Regarding Stock Price Behavior," Econometrica, XXXVIII, No. 4 (October, 1960), pp. 909-915. 83
Osborne attributed the empirical behavior of price logarithms to the Weber-
Fechner law. "The Weber-Fechner law states that equal ratios of physical
stimulus, for example, of sound frequency in vibrations per second, or of
light or sound intensity in watts per unit area correspond to equal inter
vals of subjective sensation, such as pitch, brightness, or noise. The value
of a subjective sensation', like absolute position in a physical space, is
not measurable, but changes or differences in sensation are, since by experi
ment they can be equated, and reproduced, thus fulfilling the criteria of 72 measurability." By hypothesizing the Weber-Fechner law, Osborne is saying
the absolute level of stock prices is of no significance because only changes
in prices can be measured by traders or investors. The loge of the ratio
p(t)/ p(t-l) is equal to loge p(t) - loge p(t-l), as in 4 .1.
The limited liability feature of investment in corporations and the
capital budgeting model provide some basis for the argument that investors
should a priori expect something like a log-normal distribution. Fisher has
argued that "traders know the distribution of price changes must be skewed.
The size of negative changes is restricted by the limited liability feature of the corporation. Price of a share of stock cannot fall below zero. In
the upward direction there is no such limit. At any momemt each firm, cur rently controlling only a small portion of the world's wealth, has some small chance of great (unbounded?) growth with a consequent rise in the price of its stock. Thus expectations must be skewed, having a lower, but no upper bound. Further, the most probable changes are small rather than ex treme .
72 M. F. M. Osborne, "Brownian Motion in the Stock Market," Operations Research, VII (March-April, 1959), p. 146. 84
Professor Fisher further suggests that such an expected distribution may not be strictly log-normal, but may simply be similar to a log-normal 73 distribution, i.e, skewed, and with its lower bound at zero."
The present value approach to the valuation of future income provides another reason why one might expect the changes in the logs of prices to be more appropriate for analysis than arithmetic first differences. A log normal distribution arises when some variant of a law of proportional ef fect is at work. The combination of the present value formula and a posi tive discount rate suggests that the discount rate itself is a source, perhaps the source, of the proportionality of price changes to prices, and the attendant normality of changes in the logs of prices.
Investors seek yields which are acceptable, and prices change in order to bring expected income and the desired yield into equilibrium. "Thus prices are subjected to numerous small multiplicative adjustments as traders estimate and reestimate yields. Even in the absence of uncertainty, price changes are proportional to price since the present value formula implies that prices change at r percent per year."^
In summary, the logarithmic transformation provides a model that fits the empirical data better.^
73 Arnold B. Moore, "Some Characteristics of Changes in Common Stock Prices," Cootner, The Random..., pp. 142-43.
^Moore, pp. 144-145.
^ F o r this quote and a more extensive discussion of the logarithmic transformation of prices, see Moore, pp. 142-145. 85
The independence assumption of the revised model is still subject to 76 question. Taussig has argued much earlier that market prices fluctuate within a range or penumbra. General consensus about underlying supply and demand factors prevents the movement of prices outside the penumbra, but lack of knowledge about current, specific market conditions results in a great deal of fluctuation within the penumbra.
Working^ provided a slightly different view of the same issue. He postulated that traders received information in a substantially independent series. The gradual interpretation and dissemination of information results in a series of price changes, however. If each trader cannot significantly influence the market, then price movements will consist of a series of very small movements.
Neither Taussig's analysis nor Working's analysis provide a very useful basis for predicting the behavior of stock prices over time. Price changes may be correlated within Taussig's penumbra, or they could tend to be inde pendent, depending on the time period selected. The correlation of price changes in Working's theory should depend on the time period studied and the relationship of the time period to the frequency of major bits of information.
A rapid flow of information might tend to reduce the probability of trends due to the existence of conflicting new items in various stages of dissemina tion to all traders.
76 F. W. Taussig, "Is. Market Price Determinate?", Quarterly Journal of Economics, XXXV (May, 1921), pp. 394-All.
^H. Working, "A Theory of Anticipatory Prices," American Economic Re view, XLVIII (May, 1958), pp. 189-199. 86
78 Moore found that weekly differences in logs are negatively correla ted, even though a price index he constructed was not. He attributed his findings to two factors; the standard deviation of the price relative of the security and the amount of dividend being paid. His findings involved a small sample of stocks (30) and he only suggested that his study provided an indication of possible sources of negative correlation. He did acknow ledge that Taussig and Working may have provided the reasons for a corre lation of price changes.
Cootner has argued that price changes are not independent, and that 79 they do not form a Gaussian or normal distribution. He postulates that there are two distinct types of investors. One group consists of people with occupations unrelated to the stock market. Costs of search are high for people in unrelated occupations, so this group of investors tends to accept current market prices and price movements as good measures of the value of a stock. A second group might be called the professionals. The professional group includes all of those investors who have access to in formation about the basic profitability factors of a corporation. The stock market is expected to behave as a random walk except in those cases in which prices tend to move away from some equilibrium level. If a stock becomes underpriced relative to its basic earning capacity, then the pro fessionals will begin to buy it and thus push the price up. If a stock becomes overpriced, the professionals will sell shares they acquired pre viously or will sell short, thus forcing the price downward. The cumula-
78 Moore, pp. 146-147.
^ P a u l H. Cootner, "Stock Prices: Random vs. Systematic Changes," Industrial Management Review, Vol. Ill, No. 2 (Spring, 1962), pp. 24-45. 87 tive action of professional investors thus creates a set of barriers beyond which prices will not tend to move. Even though Cootner admits a great deal of the movement of stock price change can be described by a random walk, there are some significant implications of his formulation. Most price changes will tend to be independent, but those price changes in the direction of a nearby barrier will frequently be followed by price rever sals. These price reversals at the barriers are used by Cootner to ex- 80 plain the slight negative autocorrelation found independently by Moore 81 82 and by himself. Larson and Alexander also provided some evidence that changes in futures prices are not independent.
A second significant element of Cootner's analysis is the statement that the distribution of price changes over short periods of time would be more leptokurtic than the normal distribution. The existence of barriers on prices implies that there should be more short or small price changes, and more large price changes than one expects from a normal distribution of price changes. Price movements toward a barrier are cut short, whereas they would not be in Bachelier's model. Likewise, large price changes be- . tween the barriers would exist more often than is consistent with a normal 83 distribution of price changes; Sample evidence reported by Fama supports
80 Moore, pp. 146-147. 81 Arnold B. Larson, "Measurement of a Random Process in Futures Prices," Food Research Institute Studies, I, No. 3 (November, 1960), pp. 313-324. 82 Sidney S. Alexander, "Price Movements in Speculative Markets: Trends or Random Walks," Industrial Management Review, II, No. 2 (May, 1961), pp. 7-26. 83 E. F. Fama, "The Behavior of Stock Market Prices," Journal of Business, XXXVIII, No. 1 (1965), pp. 34-105. 88 the hypothesis of a leptokurtic, long tailed distribution with non-zero mean for logged price changes.
Mandelbrot has proposed a model which differs in important ways from 84 that of Bachelier and that of Cootner. He argues that past research has over emphasized the approximate normal distribution of price changes, and ignored the obvious deviations from normality. 85 His suggested model is the difference equation
(4.2) logeP(t) - logep(t-l) = e*(t), where e*(t) is a memeber of the family of distributions called stable Pare tian.
Mandelbrot's model is a generalization of Bachelier's model because the Gaussian distribution is a special case of the general family of stable
Paretian distributions. This model has some appeal because it is more gener al than the Gaussian hypothesis, and because it explains some of the erratic behavior of successive price changes observed empirically. A major disadvan tage is that the variance of a stable Paretian distribution is not finite, except in the special case of a Gaussian distribution. Much of the portfolio literature is based on the existence of a finite variance, and all of the existing portfolio literature would need to be reformulated in terms of stable Paretian distribution parameters. Furthermore, many statistical tools used in econometric studies are based on the existence of finite variances,
®^Benoit Mandelbrot, "The Variation of Certain Speculative Prices," Journal of Business, XXXVI, No. 4 (October, 1963), pp. 394-419.
®^Kendall (op. cit.) found that weekly price changes of Chicago wheat and British common stocks were almost normally distributed. Moore (op. cit.) found that his sample of 30 common stocks from the New York Stock Exchange were also approximately normally distributed. 89 and may be of much less value than originally thought, if the variables be ing studied do not have finite variances.
Fama supports Mandelbrot's hypothesis in that he believes that
"... the stable Paretian hypothesis is more consistent with the data than 86 the Gaussian hypothesis." He examined the distributions of daily first differences of log price of each of thirty stocks included in the Dow-Jones
Industrial Average.
Mandelbrot's hypothesis has been attacked on the basis that it is only 87 an alternative model. Cootner criticizes Mandelbrot's casual methodology and his use of spot prices, as well as arguing that the evidence is not very compelling. Godfrey, Granger and Morgenstern examined some securities using spectral analysis and concluded that "no evidence was found in any of these series that the process by which they were generated behaved as if it pos sessed an infinite variance."^®
Press recently postulated a statistical model of stock price which has a finite variance and is consistent with the empirical data. "The model is similar to previous analyses of stock market price behavior in that logarith mic price changes are assumed to be independent (random walk models). How ever, this model differs from earlier work in that the logarithmic price changes are not assumed to follow some stable distribution (which might
86 Eugene F. Fama, "Mandelbrot and the Stable Paretian Hypothesis," Journal of Business, XXXVI, No. 4 (October, 1963), p. 428. 87 Paul H. Cootner, "Comments on the Variation of Certain Speculative Prices," in Paul H. Cootner (editor) The Random Character of Stock Market Prices (Cambridge: The M.I.T. Press, 1964), pp. 333-337. 88 M. D. Godfrey, C. W. J. Granger, and 0. Morgenstern, "The Random Walk Hypothesis of Stock Market Behavior," Kylos, XVII (1964), p. 13. possibly be normal). Instead, the logarithmic price changes are assumed to follow a distribution that is a Poisson mixture of normal distributions. It is shown that the analytical characteristics of such a distribution agree with what has been found empirically. That is, this distribution is in gen eral skewed, leptokurtic, more peaked at its mean than the distribution of a comparable normal variate, and has greater probability mass in its tails 89 than the distribution of a comparable normal variate." Press' model has been applied to ten common stocks, but no comprehensive testing of the rela tive merits of this model and those previously hypothesized has been made.
At this stage one can only argue that the Press model is consistent with the empirical evidence, and that the Mandelbrot model is consistent with some em pirical evidence. Both models represent the empirical data better than the
Gaussian model of equation 4.1.
In summary, no single model of the changes in stock market prices has been thoroughly tested and generally accepted. Existing models do imply that changes in price appear to be nearly independent. Slight negative autocorrelation may be induced by the effect of prices going ex-dividend, and by the limiting barriers on price created by the cumulative activities of professional traders. Distributions of logarithmic price changes tend to be leptokurtic and more long-tailed than the Gaussian or normal distri bution. Knowledge of past prices is of little expected benefit in the pre diction of price relatives.
®^Press, p. 317. 91
The Potentiality of Stock Price Prediction
Some writers have assumed that the random-walk hypothesis of stock price behavior implies that stock price changes are not predictable. This assumption is false. "When statisticians hypothesize that the course of a
stock's prices describes a random walk or Brownian motion, they do not imply
that a skilled student of the subject cannot forecast price changes. They merely imply that one cannot forecast the future based on past history
Moreover, if one examines the random-walk literature carefully, then it becomes apparent that predictions of intrinsic value are a basic part of the random-walk theory. Cootner states some heuristic arguments behind the 91 random walk theory as follows:
If any one group of investors was consistently better than average in forecasting stock prices, they would accumulate wealth and give their forecasts greater and greater weight. In the process they would bring the present price closer to the true value. Conversely, investors who were worse than average in forecasting ability would carry less and less weight. If this process worked well enough, the present price would reflect the best information about the future in the sense that the present price, plus normal profits, would be the best estimate of the future price.
Presumably, this process of rewarding good forecasters of intrinsic value, and eliminating poor forecasters, is a continuous process which gradually improves forecasts of intrinsic value.
Further evidence that predictions of intrinsic value are a fundamental part of the random-walk hypothesis of stock price behavior can be found in
90 Paul H. Cootner, "Introduction - Part II," in Paul H. Cootner (editor), The Random Character of Stock Market Prices (Cambridge: The M.I.T. Press, 1964), p. 80.
^^Cootner, "Introduction ...," p. 80.
9 92 the models of Cootner, Working, and Larson. The flexible barriers to stock price movements that Cootner envisions are based on the ideas professional 92 investors have about the intrinsic value of each security. Professional investors have the time and the resources to investigate the economic power of each company, and are able to formulate some expectations about the future dividends or earning capacity of the firm. These expectations are translated into a market price range beyond which the stock price will not be allowed to move without corrective action. Professionals would have no basis for formulating these ideas about the extent to which a stock is underpriced or overpriced if the stock's price were not closely related to some economic forces.
Working argues that current prices reflect the best projections made 93 by those involved in the market. Price changes arise due to revised ex pectations and not due to the haphazard movement of prices in the short run as suggested by Taussig. 94 The expectations discussed by Working are expec tations about basic supply and demand forces. Because truly new information emerges randomly, or in predictable trends, price movements tend to be ran dom in his model. "Working's model of price movement in an ideal market is 95 thus a random walk; that is, price is the cumulation of random movements."
92 Paul H. Cootner, "Stock Prices: Random vs. Systematic Changes," op. cit. 93 H. Working, "A Theory of Anticipatory Prices," American Economic Review, XLVIII (May, 1958), pp. 189-199. 94 F. W. Taussig, "Is Market Price Determinate?", Quarterly Journal of Economics, XXXV (May, 1961), pp. 394-411.
^ArnoId B. Larson, "Measurement of a Random Process in Future's Prices," Food Research Institute Studies, I, No. 3 (November, 1960), p. 316. 93
Working's model is tested by Larson and is found to be more consistent with the behavior of changes in corn futures prices than the model of
Taussig, which stressed the haphazard movement of prices within a penumbra.
The penumbra in Taussig's model is determined by supply and demand forces, but most of the short-run movement of prices is based on predicted price trends in the market, not fundamental economic forces.
Although there may be some information in a historical stock price series, such as slight autocorrelation, the information is generally insuf ficient to be of economic value in predicting future price changes.^ As a result, estimates of intrinsic value must be based on more fundamental eco nomic forces. The valuation models developed below are an attempt to incor porate economic forces in the prediction of common stock price relatives.
One should not conclude from the above analysis that past price rela tives are not useful in the prediction of future price relatives. If the relatives are mean-reverting, then past data provides an estimate of that mean. On the other hand, if the price relatives follow a random walk, then the most recent price relative is the best prediction available for next year's price relative. In both cases, some information about prior price relatives is absolutely necessary.
^ F o r a more complete discussion of this point see Robert A. Levy, "The Theory of Random Walks, A Survey of Findings," American Economist, Vol. II, No. 2 (Fall, 1967), pp. 34-48. 94
Naive Prediction Models
A price relative prediction model must reduce price relative uncer tainty if it is to be of any predictive value. Price relative uncertain ty is reduced whenever the conditional variance of a set of observed price relatives, given the respective predicted price relatives, is less than the variance of the observed price relatives without regard to the predictions.
This concept of uncertainty reduction is discussed in detail in section six of Chapter III. However, for any small sample of predictions, the rela tionship of actual and predicted price relatives will deviate from that equality relationship which is expected at the time the predictions are being made. For such samples, total uncertainty consists of two elements.
One is the conditional variance of the price relatives, given the set of price relative predictions, while the second is the portion of the predic tion errors which is created by the lack of knowledge about the true linear relationship between the actual and the predicted price relatives. When combined, these two components of uncertainty are equal to the mean-squared prediction error, which is the sum of the squared differences between pre dicted price relatives and the corresponding observed price relatives, di vided by the number of differences.
The information contained in a past series of common stock price relatives may be used in\arious ways to make predictions about future price relatives. At one extreme, one might use the most recent price relative as the prediction of next year's price relative. At the other extreme, one might use an average of the price relatives observed for many years in the past as a prediction of next year's price relative. The 95 proposed naive prediction models involve various weightings of past price relatives that include these two extremes, as well as thirty-nine grada tions in between. Prediction models using only past data in the same series are called naive models because they incorporate no other economic variables in the formulation of predictions. In this case, specific eco nomic forces and the relationships of these forces to common stock price relatives are ignored by the naive prediction models. However, it should be noted that the general class of naive prediction models includes those involving sophisticated mathematical techniques which cannot be regarded as simple, or trivial. , n 97
The naive models are comparable with the growth model method in terms of cost to operate, assuming that the basic data is available in useable form. They are also comparable in the sense that they are well-defined, objective methods which can be duplicated by others. Such would not be the case if the predictions were made by individuals using more intuitive methods
Three of the four models of stock price tested for price relative pre diction performance are a constant model, a linear trend model and a con stant exponent model. The constant terms in each equation are estimated by means of exponential smoothing. The optimal smoothing constant, or set of constants, is allowed to change from company to company. More details of the smoothing constant selection method are presented in Chapter V.
The three models of stock price tested, along with their respective
^Charles F. Roos, "Survey of Economic Forecasting Techniques," Econometrica, Vol. XXIII, No. 4 (October, 1955), pp. 363-395. 96 smoothing statistics, coefficient estimates (equation constants) and price relative prediction models are summarized in equations 4.3 through 4.18:
Constant model;
(4.3) Model of the process: R$t = A^gt
(4.4) Process observed with noise: R^fc = + e^
(4.5) Smoothing statistic: + ^i i t -1
(4.6) Estimate of coefficient:
A (4.7) Prediction: Rit+1 = Ai0t
Linear Model;
(4.8) Model of the process: R$ = + A^^fcT
(4.9) Process observed with noise: R ^ = A^Qfc + A^-^T + e^t
(4.10) Smoothing statistics: = aRit + ^ H t - 1
(4.11) S2 .t - aSU t + B S ^
(4.12) Estimate of coefficients: A ^ ^ = 2 S ^ t -
Ailt - “/S(SU t - s2it>
(4.13) Prediction: Rit+1 = AiQt + Allt for T = 1
Constant exponent model; Ai0t (4.14) Model of the process: R* = e
Afot + (4.15) Process observed with noise: R_^t = e
(4.16) Smoothing statistic: = aloge R^t +
(4.17) Estimate of coefficient: A^q = S ^ t 97
A -n (4.18) Prediction: = e 1
R*t is the true price relative of security i (i = 1, 2, ..., 61) for
period t (t = 0, 1, 20)
R^ is the price relative observed for security i (i = 1, 2, ..., 61)
for period t (t = 0, 1, 20)
A R^ is a prediction of R^t
Sfit* are smoothing statistics as defined
are equation coefficients (constants) for security i
(i = 1, 2, 61) for period t (t = 1, 2, ..., 20)
e. is an error or disturbance term which is normally distributed it with zero mean, constant variance and is independent of past
disturbance terms and past price relatives.
The initial conditions for the constant model and the linear trend model are determined in the same manner. The initial value for S., _ (t = 1, lit 2, ..., 5) is equal to the arithmetic mean of the first five price rela tives observed during the years 1948 through 1952 (t= 1, 2, ..., 5). For 5 any t (t= 1, 2, ..., 5), = j£l where j is the time subscript in the summation index. The initial value of is assumed to be zero 2it for 1948 (t = 1) only. The smoothing statistic for the constant exponent model is determined in a similar manner, except that the data is in loga- 5 rithm form, or S.._ = log R. ,/5 for t = 1, 2, ..., 5. lit j=l e ij The annual price relative is decomposed into the dividend yield ele ment and the price change element in order to formulate some expectations about the value of the smoothing constant, a. First, consider a price 98
relative such that:
(4.19) Rlt - (D.t + Plt)/Ple.x where
is the price relative for security i (i = 1, 2, 61)
and t is period t (t = 1, 2, 20)
D is the cash dividend paid during period t for security i
is the equivalent price per share for security i for
period t .
The expression in 4.19 can be separated into two parts such that
D P (4.20) R = ^ + ^ — , where t-1 t-1
is the annual price relative for year t
is the amount of cash dividends paid during year t
P is the common stock price at the end of year t assuming
adjustment for all stock dividends or stock splits.
The company subscript, i, has been omitted for convenience and clarity.
If the expected price change is zero, we can write the equation
(4.21) E (P - Pfc i) = 0, and since P^_ ^ is known at the start of year t,
(4.22) E (Pfc) - P = 0 or
(4.23) E (Pfc) = Pfc in which case
P. (4.24) E 1 'E (PJ = 1 ?t-l Pt-1 1
If the expected price change is some constant percentage, representing the upward drift in common stock prices, then 99
(4.25) E (Pt - Pt_1) = wPt_1
(4.26) E (Pfc) = Pt_1 + wPfc_1
(1 + w) P t-1 (4.27) E = 1 + w, t-1 t-1 when w is the constant percentage of upward drift.
Let Y be a random variable with mean My which is the dividend yield on common stocks. We are assuming that dividend yield is mean reverting in the short run even though shifts in mean yield may occur on occasion.
Since E (pt;/I>t j_) = My, we can saY that
(4.28) E (D /Pfc p j_) = My + 1 + w, the mean dividend yield plus a constant, 1 + w.
If 4.28 represents the behavior of price relatives, then one would expect a low smoothing constant because the price relative series will fluctuate about an average equal to My + 1 + w.
In contrast, suppose that the expected change in dividend yields is zero. The price relative can be decomposed as follows
D D , P P' t t-1 t t-1 (4.29) R.-V- l P , P „ t-1 . t-2P t-1 , P t-2 ^ M. D ; Dt t-1 + E It t-1 (4.30) E (Rt - Rt_1) = E ; y P P t- 1 t-2t-1 t-2 t-1 t-2t-1 but the first term on the right side of 4.30 is zero by assumption, so
P t-1 (4.31) E (Rt - Rt_1) = 1 + w - p « 0 . t-:2
The nature of the price relative depends on the nature of the dividend yield. The assumption that annual price changes are zero has not been validated, but the assumption has been validated with respect to short term price movements (see review of random walk). If we are free to as sume that annual price changes tend to zero, or some constant percentage of drift, then the nature of dividend yield becomes the important issue.
If the expected dividend yield is zero, then the expected price relative is If the expected dividend yield is equal to the mean of a stable distribution (My), then the expected price relative is equal to My + 1 + w.
Model No. 1 or Model No. 3 should outperform Model No. 2 whenever dividend yield is mean-reverting and there is some constant proportional drift in the changes in stock prices. The expected smoothing constants for such a process should be less than 0.3 for a and greater than 0.7
(1 - a) for 3. Because the dividend yield is non-negative, and the expec ted price change is zero or slightly positive for most firms, the expected price relative for most firms will be greater than or equal to the pre vious year's dividend yield plus 1.0. Individual price relative observa tions frequently drop below 1 .0 , but the expected price relative is some what above 1 .0 .
Smoothing constants are selected for Model No. 1, for each firm, in two ways. One method is to use Values between zero and one in increments of .025 for a, on all twenty price relative observations; then calculate the mean-squared prediction error for each value of a; and then adopt the smoothing constant which resulted in the lowest mean-squared prediction error for the test period from 1948 through 1967. For each company there exists an optimal smoothing constant, ten predictions resulting from use 101 of the optimal a, and ten prediction errors. The second approach is to use forty-one smoothing constants from 0.0 through 1.0 in increments of
0.025 for years one through ten (1948 through 1957). That smoothing con stant which resulted in the smallest mean-squared prediction error over the base period 1948 through 1957 is used to estimate Awhich is the prediction of the price relative for 1958. That smoothing constant which results in the smallest mean-squared prediction error for the period 1949 through 1958 is used to estimate A^q , which is now used to predict the price relative for 1959. This process must be repeated for the remaining predictions on a year to year basis so that each annual prediction is based on the statistic provided by the best smoothing constant for the prior ten years. When the process is completed, there are ten price relative pre dictions and ten price relative prediction errors as was the case for the first method. However, there may be as many as ten different smoothing constants used to make those predictions. The usual case is to have one or two different smoothing constants for the entire ten years of predic tions .
Model No. 3, which is the constant exponent model, is presented in equations 4.14 and 4.15. This model is based on the hypothesis that the annual rate of return varies about the company's average rate of return, which is determined by the growth rate of earnings and dividends, and by the level of uncertainty associated with that company. The expression for the geometric average rate of return is:
(4.32) (1 + r)n = (1 + r ^ U + r2)...(l + rn) where r is the average rate of return and r^, r^j are the rates 102 of return for the respective periods 1, 2, n. If some rate of re turn is associated with a particular corporation's uncertainty level and growth rate, then for any t (t < n), E(1 + r ) = 1 + r, because E(rfc) = r so long as the growth rate and uncertainty level do not shift. The rate of return for any year, plus 1 .0 , is equal to the price relative of that year. Let the logarithm of the annual price relative be symbolized by
Rt> and the average price relative be symbolized by R. The averaging method used to approximate R is:
(4.33) R = aR^ + a(l - + 01 (i " a^ ^ t-2 * ’ ’ ’ ” 01
In this way the more recent price relatives are given more weight than they would receive from a simple, unweighted average of the price relatives dur ing the period from 1948 up to the year for which a prediction is' being made. The estimate of R is denoted R. Model No. 3 averages the logarithms of past price relatives in order to approximate the weighted geometric average relative, whereas Model No. 1 is a weighted arithmetic average of past price relatives.
As in the case of Model No. 1, two methods are used to select the op timum smoothing constant, or set of constants, for each firm. One method involves finding the best single smoothing constant for the years 1948 through 1967. The other method involves the use of that smoothing constant which minimizes prediction errors for the prior ten years to estimate the statistics used to predict the next year's price relative. These are the same two approaches described more fully with respect to Model No. 1.
Model No. 2, the linear trend model, is based on empirical evidence about stock price changes and rates of return earned on common stocks in 103
prior periods. Farrar was forced to simulate investor expectations in
his study of uncertainty, and in doing so he tested some naive prediction 98 models. In predicting growth stocks' annual price changes, he found that
the continuation of the previous year's arithmetic change in price is more
likely than the prediction of no price change, or the prediction of a con stant rate of change. In other words, predicting a continuation of last year's arithmetic price change results in smaller prediction errors than
results from the prediction of a zero price change or from the prediction
that last year's rate of price change will continue for another year. Most of the companies in the sample being used in this study are not growth stocks, but Farrar's findings are potentially relevant for some of the
firms that are growing at above average rates.
Translated into price relative terms, the predicted price change is positive for growth stocks. Although Farrar did not test for an average amount of annual price change, it is reasonable to think in terms of an average price change because the expected price changes are positive for growth stocks. Suppose that a growth stock does experience a constant annual increase in stock price over the years. The price change element of the price relative will be declining because the relatively constant amount of price increase is being divided by larger and larger share prices. If dividend yield is increasing, then it is possible that in creases in the dividend yield offset the declining ratio of price increase, so that the price relative is constant. If dividend yield is constant, then the total price relative will decline over time because the price
98 Donald E. Farrar, The Investment Decision Under Uncertainty (Engle wood Cliffs: Prentice-Hall Inc., 1962), p. 57. 104 change ratio declines. Under these conditions, one would expect to find a trend in the price relatives.
An examination of general movements in the rates of return earned on a broad sample of common stocks also suggests that a linear trend model 99 may be relevant for limited periods of time. For example, there was an increasing trend in the cumulative rate of return from 1945 to 1956. The fact that the cumulative rate of return was sharply increasing implies that the annual rates of return were also increasing. There is little evidence of trend in subsequent years, but the earlier trends suggest that linear trend models may be appropriate in predicting price relative trends. The evidence is not strong.
The naive prediction models described in equations 4.3 through 4.18 are appropriate models of common stock price relatives under varying cir cumstances. First, if the price relative series of a common stock does contain information, such as trend, then this information will be incorpo rated into the price relative predictions of the linear trend model. Under different circumstances, the price relatives may follow a random walk. The most recent price relative should.be the best prediction of next year's price relative if the relatives do follow a random walk. Either the con stant model or the constant exponent model can readily adapt to the random walk case because the smoothing constant (a) can vary from 0.0 through 1 .0 .
However, if the price relatives are best described by a stable prob ability density function, then the mean of the past price relatives should
^Lawrence Fisher and James H. Lorie, "Rates of Return on Investments in Common Stock: The Year-by-Year Record, 1926-65," The Journal of Business, Vol. XXXX, No. 3 (July, 1968), pp. 1-26. 105
provide the best prediction of next year's price relative.
Accordingly, a fourth method of predicting future price relatives is
the use of the mean of the prior ten price relatives to predict the next
price relative. The prediction model may be described as:
(4.34) R.- = £n R. . /n, where j=l * L J
R^t is the common stock price relative for company i for
period t,
R. is the prediction of R. , xt Xt *
n is the number of observations (10 in this case).
Superior price relative prediction performance of this model will serve as
evidence that the price relatives can be represented by a stable probabili
ty density function.
In summary, four naive methods of predicting price relatives are for mulated. Three models incorporate equation constants estimated by means
of exponential smoothing, while the fourth method incorporates a simple
average of the prior ten price relatives. The resulting prediction errors
are evaluated in order to determine the best overall naive prediction model.
The results of these prediction performance tests are discussed in Chapter
V in conjunction with the evaluation of the economic growth model's perfor mance .
An Economic Model of Stock Value
The thrust of the random-walk hypothesis is that technical analysis
is not useful in buying common stocks. Even though brokers and analysts
still claim that technical analysis can lead to more timely purchases, 106
academic research seems to reflect the general agreement that technical
analysis is not useful. The empirical evidence in support of the random walk hypothesis is extensive and technically sophisticated, but it is
somewhat naive in that the statistical tests performed do not test for
the complex patterns envisioned by the technical analyst. Nevertheless, academic researchers seem content to dismiss the notion of technical ana
lysts.
A somewhat different situation exists with respect to the determina
tion of the intrinsic value of a common stock. Intrinsic value is an es
timate of the true value of a common stock, which depends on the long-run profitability of the corporation. The determination of intrinsic value not only is important in the selection of the best common stock invest ments for long investment periods, but it also helps indicate which stocks seem to be underpriced or overpriced at the current time.
Both academic researchers and financial analysts have devoted a great deal of time and effort to the determination of common stock values. Among practitioners, the Value Line Investment Survey is one of the most widely known approaches to the determination of intrinsic value. The Value Line method is objective enough to be of interest to academic researchers.
On the academic side, extensive research has been devoted to the de velopment of normative, as well as positive, common stock valuation models.
In addition, various studies have attempted to determine the predictive power of accounting measures of performance. The normative models are de signed to indicate the relationship of common stock values to other econo mic variables, for a given set of assumptions. Empirical models involve 107
testing hypothetical models of stock price in order to determine which
theory fits the data best. Both the empirical (positive) and the normative
valuation models are designed to isolate and define those variables which
influence stock prices.
The VLIS approach and the models constructed and tested by academi
cians share a common assumption — that stock prices and stock price
changes are predictable. Unless this assumption is made, then their ex
tensive activities are not purposeful.
The steps included in our development of an economic model for pre
dicting common stock price relatives are (1) a statement of the inherent
limitations of stock valuation models, (2) a brief review of some findings
from other empirical studies, (3) an explanation of the Value Line method,
(4) arguments in support of including certain variables in the model, and
(5) specification of the model.
Limitations of stock valuation models.— All common stock valuation models suffer from the same problem — uncertainty. Our inability to pre
dict aspects of the future makes price prediction very difficult, regard less of the nature of the resource being priced.Predicting the price of the commodity steel is as difficult as predicting the net earnings to be reported by United States Steel, Inc., in other words. Nevertheless, improved predictions should theoretically lead to better decisions in an economic sense, and so the prediction process is worthy of study.
Prediction models based on theories of economic behavior are generally
^Slany of these general comments on stock valuation model construc tion are from William J. Baumol, "Problems in the Construction of Stock Valuation Models," paper read at the Spring Seminar of the Institute for Quantitative Research in Finance, April 25, 1968. 108
considered more fruitful because there is a greater probability that theo
retically defined relationships will continue in the future, whereas spuri
ous ones will not. Baumol has clearly stated the issue, which is:
The ad hoc models may do better as forecasters. If one is lucky and happens to hit on a mathematical relationship that cap tures for the moment the psychology of the investing public, good predictions may emerge. But there is no reason to believe that the ad hoc model that would have done well in 1960 will perform satisfactorily in 1967. We have no grounds on which to expect that the psychological climate will remain stationary over time, and perhaps only by the wildest coincidence will a mathematical relationship derived by ad hoc processes long follow successfully the vagaries of investor models. This means that even the rela tively successful among ad hoc models are likely to fail us when we need them most. It is the turning points in the behavior of the market that we want most help in predicting, and it is pre- cisely there that the ad hoc models will most probably go wrong.
With these limitations and prediction problems in mind, we turn to some of the findings of others, and a review of the Value Line Invest ment Survey Method.
Empirical evidence.— There is an extensive literature of finance and accounting in which the effect of financial variables on stock prices, or stock price changes, is tested. Many articles have been concerned with the usefulness of financial ratios in predicting the failure of corpora tions. Horrigan suggested that the usefulness of financial ratios may 102 have been overlooked in recent years. Horrigan had some success in predicting the long-term credit standing of corporations with respect to 103 bond ratings. Beaver has had some success in predicting failure using
^^Baumol, "Problems ... ." 102 James 0. Horrigan, "Some Empirical Bases of Financial Ratio Anal ysis," Accounting Review, Vol. XL, No. 3 (July, 1965), pp. 558-568. 103 James 0. Horrigan, "The Determination of Long-Term Credit Standing with Financial Ratios," Empirical Research in Accounting: Selected Studies, 109
104 ratios of accounting variables, but he also found that market price de clines may serve as more timely predictions of failure.These findings may be the result of investor predictions of failure due to early trends in the ratios as well as other information. The evidence presented in
Beaver's study suggests that accounting variables may be anticipated by the market, and thus are of little value in formulating price relative predictions.
There is evidence, however, that stock prices do move jointly with accounting numbers, and that different accounting measures do influence in- vestment«. decisions.j • • t Roper,, 106 „ Horngren, 107 Cerf, . 108 and . „Burns 109 . have used . the questionnaire or interview approach to determine the extent to which accounting data is useful in various types of decisions. Bruns,
1966, pp. 44-62. A supplement to Journal of Accounting Research, Vol. 4 (1966).
^^William Beaver, "Financial Ratios as Predictors of Failure," Empirical Research in Accounting, 1966, pp. 71-111. A supplement to Journal of Accounting Research, Vol. 4 (1966).
^^William Beaver, "Market Prices, Financial Ratios, and the Predic tion of Failure," Journal of Accounting Research, Vol. VI, No. 2 (Autumn, 1968), pp. 179-192. 106 Elmo Roper, A Report on What Information People Want about Policies and Financial Conditions of Corporations, Vols. I and II (New York: Con trollers Institute Foundation, Inc., 1948.
^^Charles T. Horngren, "Disclosure: 1957," Accounting Review, Vol. XXXII, No. 4 (October, 1957), pp. 598-604. 108 Alan R. Cerf, Corporate Reporting and Investment Decisions (Berkeley, California: Institute of Business and Economic Research, 1961).
Research in progress. 110 William J. Bruns, Jr., "Inventory Valuation and Management Deci sions, " The_^£counting_Revoew, Vol. XL, No. 2 (April, 1965), pp. 345-357. 110
111 112 Dyckman, and Jensen have used simulated data to determine if students or financial analysts make different operating decisions or different in vestment choices given accounting data in controlled experiments. Their results were mixed, and at most one can say there is sometimes a difference 113 in decisions under certain experimental conditions. Greenball used a simulation to determine the relative performance of different measures of income with respect to criteria such as bias and predictability.
Many empirical tests of the association of stock prices and account- 114 ing variables have been performed. Gordon tests various models of stock price on a cross-sectional basis in order to find that model which best represents the stock valuation process. Dividends, growth, earnings in stability and even leverage were found to be significant factors in the valuation of common shares. Gordon's results were very good in terms of the relationship between the linear regression coefficients and their stan- 115 dard.errors. Staubus has tested various measures of income by using
111 Thomas R. Dyckman, "The Effects of Alternative Accounting Tech niques on Certain Management Decisions," Journal of Accounting Research, Vol. II, No. 1 (Spring, 1964), pp. 91-107; and "On the Investment Deci sion," The Accounting Review, Vol. XXXIX, No. 2 (April, 1964), pp. 285-295. 112 Robert E. Jensen, "An Experimental Design for Study of Effects of Accounting Variations in Decision Making," Journal of Accounting Research, Vol. 4, No. 2 (Autumn, 1966), pp. 224-238. 113 Melvin N. Greenball, "The Accuracy of Different Methods of Account ing for Earnings — A Simulation Approach," Journal of Accounting Research, Vol. XI, No. 1 (Spring, 1968), pp. 114-129.
^\lyron J. Gordon, The Investment, Financing and Valuation of the Corporation (Homewood, Illinois: Richard D. Irwin, Inc., 1962). 115 George J. Staubus, "Testing Inventory Accounting," Accounting Review, Vol. XLIII, No. 3 (July, 1968), pp. 413-424; "Statistical Evidence of the Value of Depreciation Accounting," Abacus (August, 1967), pp. 3-22; and "Alternative Asset Flow Concepts," Accounting Review, Vol. XLI, No. 3 (July, 1966), pp. 397-412. Ill
alternative inventory and depreciation methods in order to find that in
come measure which is more highly correlated with discounted stock prices.
The particular accounting method chosen does seem to influence the correla
tion between the accounting income measure used and common stock prices.
Benston uses various averaging techniques to determine if one method of averaging past annual rates of increase in earnings results in a sig- H 6 nificantly increased partial correlation with stock prices. His re sults were generally poor, which is some evidence that predictions may be difficult. Benston does not make predictions, and his is a cross-sectional study. Nevertheless, he does use price relatives instead of prices, and he is concerned with annual rates of change instead of the raw variables.
Benston eliminates general market effects by means of a simple linear regression equation of the price relatives on Fisher's link relative, which is an average of the price relatives for securities listed on the N.Y.S.E.
He is analyzing the residuals from this linear regression. Industry mean- effects are isolated by means of dummy variables. Part of the reason for
Benston's poor results may be the fact he has eliminated some of the inter action effects by using simple linear regression to eliminate market ef fects, of which each company is a part.
The implications of these studies with respect to intertemporal pre diction formulation are not clear. In many cases, it is not clear to what extent the results obtained are dependent on the manner in which the study is formulated, or on the variables selected. As a result, many conflicting
11/' George J. Benston, "Published Corporate Accounting Data and Stock Prices," Empirical Research in Accounting: Selected Studies, 1967, pp. 1-54. A supplement to the Journal of Accounting Research, Vol. V. 112
conclusions may be drawn from similar studies. The conclusions drawn here are based on intuitive information instead of hard facts. One conclusion is that cross-sectional studies frequently ignore trend information, and therefore are somewhat limited in explaining the current price or price relative for a company because price changes are a function of changed ex- 117 pectations. The level of earnings is not so important as its deviation from expectations. A second conclusion is that variables that are impor tant in distinguishing between firms may not be important in explaining intertemporal price changes for any one firm. Dividend yield may help ex plain interfirm differences, but any particular firm's dividend may be relatively constant over time, and thus is of little predictive value. A third conclusion is that one cannot separate out industry and economy-wide forces in the prediction process because these factors must be predicted, as well as the interaction effect of general and specific economic forces.
Based on these conclusions, we cannot predict very accurately the possi bility of making annual price relative predictions on a firm by firm basis.
There are technical problems created by small sample sizes, but here again it is easy to be pessimistic or optimistic without empirical foundation for that attitude.
The Value Line Investment Survey.— Early in his description of the
VLIS approach, Bernhard notes that the present value of a common stock is equal to the sum of all future dividends of the stock, discounted by the 118 current interest rate. Although he notes that this statement is true,
117 See William H. Beaver, "The Information Content of Annual Earnings Announcements," Empirical Research in Accounting: Selected Studies, 1968, a supplement to the Journal of Accounting Research, Vol. VI.
■^Bernhard, p. 39. 113 he also considers it unfathomable because nobody knows what future divi dends are!
Our preliminary attempts to find a time series of the projections made by Value Line were not successful, so we adopted their general ap proach to valuation. We are not reconstructing Value Line predictions because we do not measure variables in the same manner, and because we can not duplicate the professional judgment factors which are embodied in some of the VLIS projections. However, similar variables and similar methods are used in our prediction model.
The Value Line approach involves ranking stocks with respect to (1) quality grade, (2) price appreciation potential over the next three to five years, and (3) probable market performance over the next twelve months.
These three factors are ranked separately, so each investor must decide which factors are relevant to him, and how they should be weighted.
Quality grade depends on growth and stability. Growth is measured in terms of percentage changes in earnings per share and dividends per share over one-year investment periods, three-year periods, and five-year periods. All possible percentage changes are divided into nine intervals and a code number from -4 to +4 is assigned to each interval. The growth of earnings or dividends falls into an interval for each time period con sidered and a code is assigned accordingly. The total of the code numbers for dividends is added to the total for earnings, multiplied by two, to create a composite growth index for earnings and dividends.
Stability is measured in terms of price stability. The price stabil ity ratio is an average annual price range over the previous eleven years, 114
119 adjusted for a secular trend factor.
Based on correlation studies, the VLIS analysts have concluded that stability should be weighted three times as much as growth in creating a quality index. After calculating the quality index, they rank more than eight hundred companies, and then place each company into one of nine quality grades. Judgment may be exercised to the extent of moving a stock up or down one grade, or even two.
Price appreciation potential estimates are based on a hypothesis about the general economic environment three to five years in the future. Each corporation's sales are then predicted based on the relationship of that company's sales to national income statistics in the past, and based on the probable performance of the industry in the hypothesized economic en vironment. Estimates of earnings and dividends are based on projections of gross margin rates and the past relationship between working capital and sales. Price estimates are made by using the average dividend yield rate and average price/earnings ratio associated with this firm in past years.
By comparing current price to the projected price, the Value Line analysts have a measure of price appreciation potential which can be ranked across companies.
Estimates of probable market price performance are based on the rela tionship of the moving-average of each stock's price over the prior 52 weeks, to the Value Line rating of intrinsic value. The intrinsic value rating is based on estimates of earnings and dividends in the next twelve months. The relationship among each corporation's intrinsic value, earn-
119 Bernhard, p. 46.
/ 115 ings per share, and dividends per share is based on the past relationship among average price per share, average earnings per share, and average di vidends per share, for a group of companies.
An empirically testable model of stock price.— The Value Line In vestment Survey method and the results of empirical research studies combine to provide insights and information about those factors which influence stock price relatives. An analysis of these insights and findings forms the theoretical basis for the formulation of an economic prediction model.
Economic theorists have argued that highly profitable investments
(r "s' k) represent a form of compensation for risk bearing, a return cre ated by monopoly position, or a reward for entrepreneurship and innova- 120 tion. According to the risk compensation theory of profit, higher re turns must be associated with higher risks so that expected returns, after deducting possible losses on some investments, will be comparable to the returns available on low risk investments. The well-accepted theory that above average rates of return are required by investors whenever risk is 121 above average, is supported by empirical studies of common stock returns. To conclude, both economic theory and empirical evidence support the hy pothesis that high price relatives should be associated with high uncer tainty levels, and low price relatives should be associated with low levels of uncertainty.
120 J. Fred Weston, "The Profit Concept and Theory: A Restatement," Journal of Political Economy (April, 1954), pp. 152-170.
■^^Sharpe, "Risk Aversion...," pp. 152-170; and Fama, "Risk Return ...," pp. 29-40. 116
Although the theoretical relationship of uncertainty to common stock price relative levels is clear, the proper method of measuring uncertainty is not obvious. Value Line has concentrated on price stability as a mea sure of risk, where their price stability indes is an average price range over the past eleven years, after adjustment for secular trend.
Another candidate for the measure of uncertainty is a measure of 122 earnings stability, such as that used by Gordon. In his cross sectional study, Gordon found that the stability of net earnings, with interest ex pense added back, is significantly correlated with common stock price.
The stability of earnings seems to be the more important measure of risk with respect to long-term investments, because investors can with stand temporary declines in price — so long as they are temporary. How ever, we are assuming fixed investment periods for which temporary price fluctuations will be important if they occur at year end. Investors prob ably are more sensitive to price stability than to operating income sta bility whenever their investment horizons are less than a year or two.
Rather than choose one measure of uncertainty over another, on an a-priori basis, the step-wise multiple regression algorithm will provide evidence as to which measure is the more highly correlated with past common stock relatives, on a company by company basis.
High rates of growth also create an element of uncertainty. Recall that growth takes place when investments are available for which the cor poration can earn a rate of return in excess of that rate required by in-
122Gordon, The Investment..., Chapter 12. 117 vestors. Very high rates of return are associated with risky investments, and with monopolistic returns. We have already noted that high risk pro jects are associated with high rates of return. But consider the high rates of return made available through marketing innovations, lower costs, improved products, and other circumstances that put one's product in a preferred or monopoly position. These circumstances tend to be short-lived, or at least they are not expected to be perpetual in nature. Uncertainty arises because one cannot always predict the duration of these'high return investments.
The prediction of growth rates is very important because past growth rates tend to be extrapolated into the future. When projected prices and dividends of high growth companies are discounted back to the present at market discount rates, the resulting stock price is a reflection of inves tor expectations about growth, and not about current dividend yields or normal price/earnings ratios. Therefore, any unexpected indications that a company's high growth rate is beginning to slow down should result in a significant price decrease. Even though a corporation's growth has been nearly constant over past years, an unexpected, permanent decrease in its growth rate will result in a substantial price drop. 123 Bernhard, Gordon, and W ippem have used growth variables to measure quality grade, to explain common stock price levels, and to explain earn ings/price ratios, respectively. Consistently, we expect that high growth rates will be associated with high price relatives, and that low growth
• 123 Bernhard, p. 41; Gordon, The Investment..., Chapter 12; and Ronald F. Wippern, "Financial Structure and the Value of the Firm," The Journal of Finance, Vol. XXI, No. 4 (December, 1966), pp. 615-633. 118
rates will be associated with low price relatives, ceteris paribus. One
would expect some interaction effects between growth rates and price sta
bility, but high growth rates, which are also stable, may have an impor
tant influence on common stock price relatives even in those cases where
stock prices have been stable too. Conversely, common stock prices and
price relatives may not be stable in those cases where growth rates are
low and stable. In summary, the inclusion of growth rates in the predic
tion model is based on the VLIS method, and is supported by other empiri
cal studies.
The rate of sales growth and the rate of growth of operating earnings
are the two measures of growth used in our model. Sales growth is used
because shifts in demand for a corporation's product frequently cause an
increase in sales without an immediate increase in profits. As capacity
is increased, and the corporation is better suited to handle the increased
sales volume, profits can be expected to increase. In other words, in
creases in the rate of growth of sales are expected to precede increases
in the rate of growth of earnings,
Growth rates in operating income are used in the model because earn
ings growth may not respond to sales growth in some companies. For these
cases, the better measure of long-run growth prospects is the rate of growth
in operating income. Operating income is used instead of net income be
cause fewer transitory factors affect operating income. The less stable net income series is not expected to produce measures of growth that are
stable enough to be useful in the prediction of common stock price rela
tives . 119
The availability of investment opoortunities on which a corporation can earn an above average rate of return is dependent on factors exogenous to the firm as well as factors endogenous to the firm. One very important exogenous factor may be described simply as the general condition of the economy as is reflected by employment levels, corporate profits, produc tion levels, and other similar measures. When business is good through out the economy, more people are employed, which increases total expendi tures on consumer goods and services. As the production of these goods and services approaches capacity, managers look to increased investment to cut costs and to increase capacity. As a result, groups of corpora tions, such as those producing machine tools and those constructing other productive facilities, undergo a rapid increase in activity and profit ability when the economy is in a rapid expansion phase. Likewise, reve nues of the transportation industry are higher during periods of high economic activity, and are lower during economic slow downs. Some indus tries, such as food processing, are less dependent on general economic conditions for growth and profits.
Value Line ranks corporations with respect to price appreciation potential over the next three to five years. Their starting point is a hypothesis about the economic environment in three to five years. In this case we only duplicate the Value Line method to the extent of pre dicting general economic activity one year in advance, since we have assumed a one-year investment horizon. The Value Line prediction of price appreciation could not be exactly duplicated here because so many 120 professional judgments and estimates are involved in predictions of sales and earnings potential.
We use the annual business forecasts of the Prudential Insurance Com pany of America in the prediction model. Many factors influence general business conditions in a rather complex manner. Fiscal policy, monetary policy, war labor disruptions, technological change and many more intan gible factors are among the determinants of economic expansion. All rel evant factors are supposedly included in the forecasts made for Prudential.
In addition to quality grade and price appreciation potential, the
Value Line Investment Survey ranks common stocks with respect to expected price performance over the next year. Our problem differs somewhat since we are attempting to predict future price relatives, not future prices.
One measure of possible market performance over the next year is the difference between the most recent price relative and the average of the price relatives for the five years prior to the year for which predic tions are being made. If last year's price relatives were abnormally low, then one might expect the stock's price to rebound this year, resulting in a high price relative. Conversely, an abnormally high price relative in one year may be followed by a low price relative the next year, due to an adjustment in the stock's price. The measure of expected price relative performance is analogous to the Value Line measure of the extent to which a stock is overpriced or underpriced.
The economic model of stock price is of the form:
B1 B B B, B B, (4.35) Rt = BQGt 0 ^ 1St^ D ^ 1Ut_1Vt.1, where 121
Rt is the annual price relative for period t (t =■ 1948,
1949, ...» 1967) for any security,
is a constant (i = 0 , 1 , ..., 6),
Gt is the growth rate of gross national product from the
year t-1 to the year t,
0^ ^ is one plus the average historical geometric growth
rate in operating income per share,
^ is one plus the average historical geometric growth
rate in sales per share,
D , is the difference between the price relative for year t-1 t-1 and the average price relative for the five one-
year periods ending with year t-1 ,
Ut_^ is the variance of the changes in operating income
over the five one-year periods ending with t-1 ,
V ^ is the variance of price relatives over the five one-
year periods ending with year t-1 .
A detailed description of the manner in which equation 4.35 is applied to companies in the five test industries is presented in Chapter I, while the prediction results are presented in Chapter V and summarized in Chapter
I.
Preliminary tests indicate that the multiplicative form of the model presented in 4.35 fits the data better than an additive model. This finding is consistent with many of the cross-sectional studies relating common stock prices to various independent variables. CHAPTER V
ANALYSIS AND INTERPRETATION OF THE
PREDICTION MODELS’ PERFORMANCE
The predictive performance of the four naive prediction models and the
economic growth model is reported in this chapter. Based on the mean-
squared prediction error criterion, the constant exponent model was the best of the three naive time-series models. The constant exponent model performed significantly better than the linear trend model, but only slight ly better than the constant model.
Prediction based on the mean of the prior ten price relatives, that is naive model number four, were better than those of the constant exponent model. These results lead us to the conclusion that price relatives tend to be mean-reverting, and that the time series of past price relatives con tains little or no information not contained in the arithmetic mean of that series.
The economic growth model failed to provide better predictions than the best naive method, thus leading to the rejection of hypothesis 1. The weaknesses of the economic growth model include insufficient data to achieve stable prediction equation coefficients, and an insufficient emphasis on anticipatory data among the independent variables.
In general, the variance of the price relatives for each security for the test period 1958 through 1967 is significantly smaller than the mean-
122 123
squared prediction errors associated with the best prediction model. Hy pothesis 2 is accepted on the basis of these findings.
Performance of the Exponential Smoothing Models
The three naive time-series models described in equations 4.7, 4.13, and 4.18, are
A (5.0) Model Is R.. = A. it 10
(5.1) Model 2: R.. = A. + A.- xt io xl
A. *LO (5.2) Model 3: R >4. = e where xt
A R^ is the predicted price relative of security i for
period t,
A^ , A ^ are prediction equation constants that are
estimated by means of exponential smoothing, as
described in Chapter IV.
Because A^q and A ^ are estimated by means of exponential smoothing,
the performance of each prediction model is influenced by the selection of
the smoothing constant, a. The a priori expectation is that the smoothing
constant will tend to be slightly more that 0 .0 , thus, weighting the past
observations heavily. In contrast, if the smoothing constant is equal to
1 .0 , then only the most recent price relative is used to predict the next
price relative.
Two methods of selecting optimal smoothing constants for each company are used in this study. One method is to predict price relatives for 1958
through 1967 by means of the prediction models 4.7, 4.13 and 4.18. Each 124 model Is tested forty-one times per year because the equation constants
A, and A,, are estimated with the forty-one smoothing constant values io il a = 0.0, 0.025, 0.05, 0.075, ...» 1.0. The best smoothing constant is that constant which results in the lowest mean-squared prediction error for 1958 through 1967, inclusive, in this method, the smoothing constant is said to be fixed because it is applied to the entire period over which predictions are being made.
The second method involves the selection of a smoothing constant on a year-by-year basis. As before, values of the smoothing constant are al lowed to range from 0.0 through 1.0 in increments of 0.025 while making predictions for the ten year period prior to the period for which a predic tion is being made. That smoothing constant which results in the smallest mean-squared prediction error over the prior ten years is used to estimate
A^q and A ^ , and thus the price relative of the next period. The values of
A^q and A ^ used to make predictions in 1958 are estimated by means of the smoothing constant resulting in the smallest mean-squared prediction error for the year 1948 through 1957, inclusive. In 1959, the value of A^q and
Aji used to make the price relative prediction will be those estimates which are made by the smoothing constant which resulted in the smallest mean-squared prediction error over the previous ten years, or 1949 through
1958 in this case. This method may result in ten different smoothing con stants being used for the ten predictions made during the test period, 1958 through 1967. The smoothing constants are said to be shifting in this method, because the smoothing constant used for any one company may change from period to period depending on past results. 125
Ranking with respect to mean-squared error.— Two expectations are
being tested In the evaluation of the three models of stock price rela
tives. One expected result is that the constant models (1 and 3) will out
perform the linear trend model. The basis of this expectation is the fact
that no significant trend is expected in price relatives, even though a
trend may exist in dividends and the stock price. Price relatives are com
posed of the dividend element and the price appreciation element. Based on
the assumption of a constant proportional increase in stock price, the ex
pected price relative is a constant, and does not increase over time (see
equation 4.28). If there is no upward drift in prices, then the expected
price relative is still a constant, and does not increase over time (see
equation 4.31).
Irving Fisher and John Burr Williams have written what are now con
sidered classic works in the theory of interest and the theory of invest- 124 ment value. Their methods of representing the profitability of an in vestment by the annual rate of return which equates a stream of cash pay ments to a current price is still well-accepted today as evidenced by the
fact that many current authors of finance textbooks find it useful to asso
ciate one average discount rate (geometric rate) with each given level of 125 uncertainty. As discussed with respect to the economic model, the dis count rate would remain the same until uncertainty changes, or until there
^■^Irving Fisher, Theory of Interest (New York: MacMillan, 1907); and John Burr Williams, The Theory of Investment Value (Amsterdam: North- Holland Publishing Company, 1964). 125 See Ezra Solomon, The Theory of Financial Management (New York: Columbia University Press, 1964), and Eugene M. Lerner and Willard T. Carleton, A Theory of Financial Analysis (New York: Harcourt, Brace and World, Inc., 1966). 126 is a shift in profitability of investments throughout the economic area.
An average rate of return is determined according to the following expression:
(5 .3) (1 + k)n - (1 + k-jHl + k2)(l + k3)...(l + kn), where k± is the rate of return for period i (i = 1 , 2, ..., n),
n is the number of periods,
k is the geometric average of the k^'s, and
(1 + k^) is the price relative for period i.
An average price relative is determined in the following manner.
(5.4) Rn = R2, ..., Rn » where
Rn = (1 + k)n , and
R.^ = (1 + k^), for i = 1 , 2 n.
(5.5) n log R = log R^ + log R 2 + ... + log Rr
log R1 + log R? + ... + log R (5.6) R = Antilogarithm of ------—
Naive model number three uses the calculation method described in 5.6, with alternative weighting schemes tested. Equal weights are implied by expres sion 5.6.
Based on traditional theory, and the arguments provided in equations
5.3 through 5.6, we expect that one geometric average rate of return is associated with each price relative, which implies naive nodel number three will provide the better predictions.
Past performance of the three prediction models using shifting smooth ing constants is summarized in Tables One, Two and Three. Past prediction 127 performance for the same three models using the optimal fixed smoothing constant for each model, and for each company, is summarized in Tables Four,
Five and Six. Using two methods of selecting smoothing constants for each of three prediction models results in six sets of predictions, and thus six measures of the mean-squared prediction error for each company.
The comparative performance of the three naive prediction models and the two methods of selecting smoothing constants is based on the ranking of the six measures of the mean-squared prediction error that result from the six sets of predictions made for each company. The model that performs best has the smallest total ranking over the sixty-one companies, because the ranking is from smallest to largest.
Model No. 3 performed best for the sixty-one sample companies over the ten year period 1958 through 1967. This model outperformed the other two models when an optimal fixed smoothing constant was used for each company and it also outperformed the other two models when shifting smoothing con stants were used.
The overall ranking was significant at the 99.9% level of confidence as indicated by the chi-square statistic of 123.76 reported in Table 7.
The Kendall Coefficient of Concordance is 0.43 out of a possible 1.0, which is also significant at a level of confidence well above 99.9%. We can con clude that the sample rankings are not due to chance, but are due to a real difference in the performance of the three prediction models. One could expect to observe similar findings in other test samples.
Additional evidence of the general preference for model three is gained from a review of the model rankings on an industry by industry basis. As indicated in Table 11, Model No. 3 was the best model in all but 128
one industry, textile producers. The industry rankings were all signifi
cant at the 99.9% level of confidence except the Tire and Rubber Goods In
dustry which is significant at the 99% level. An examination of the in
dividual textile producers does not reveal any conclusive explanations for
the unusual results found in this industry. Dividend patterns seem to be
more irregular among the textile producers than among the cement producers,
for which Model No. 3 is clearly the best model. Dividend pattern does not
seem to be very conclusive evidence, however, because Model No. 3 performed
better for at least two textile producers with very irregular dividend pay ment records. Another possible factor is the variability of the price re
latives being predicted. The average price relative variance for the tex
tile industry was .1367 while it was .1131 for the cement industry. These
variances seem too similar to conclude that the difference in the vari
ability of the price relatives in the two industries is the basis for the
difference in prediction model performance. Because the ranking in the
textile industry is clearly significant at the 99.9% confidence level, we
cannot ignore the evidence and suggest that the superior performance of
Model No. 1 in predicting textile industry stock price relatives is merely
a chance event. In view of the evidence, it is concluded that individual
industries may differ sufficiently to imply that models of stock price or models of stock price relatives must be unique to each industry. Similar
industries might result in similar stock price relative models, but two different models may be necessary to best represent the stock price rela tives of two dissimilar industries. More evidence on the effects of in dustry classifications is presented below.
The cement producer’s industry results are also of interest. Model 129
No. 3 was uniformly the best prediction model for all cement producers and
the overall agreement as measured by the Kendall Coefficient of Concordance
is 70%, which is high for this type of measure. These results suggest a
high degree of homogeneity in this industry, which is unmatched by any
other industry in this study.
The mean-squared prediction error can be decomposed into three parts;
constant bias, proportional bias and validity. The nature of this decom
position is reviewed here prior to a ranking of the three prediction models with respect to each error element.
The goal of prediction model formulation is to attain predictions such
that
A (5 .7) Rit = Rit + e±t, where
R^t is the price relative of security i for period t,
A R^t is the predicted price relative of security i for
period t made as of the end of period t-1 , and
e^t is a normally distributed random variable with zero
mean, the smallest variance possible, and for which
cov (e^t» e^fc j) =0.0, j < t. Other desirable con-
A ditions are cov (e^t, **0.0 and cov (e^t» ^t^ =
0.0.
It is a goal because if e^t is at a minimum, then any other prediction model will result in greater prediction errors. In reality, predictions are related to actual price relatives by means of the equation
A (5.8) R^t = a + b R^_ + e , where all terms are defined in 5.7, 130
except the constants a and b. The constant of regression, a, arises be
cause the average predicted relative is unequal to the average observed
relative, and the term b arises because the regression slope deviates from
one. 126 The mean-squared prediction error may be decomposed as follows:
(5.9) I(R - R)2/n = (R - R)2 + (Sp - rSA )2 + (1 - r2)SA2, where
R is the mean price relative,
R is the mean predicted price relative,
S is the standard deviation of the price relative predictions, P
S A is the standard deviation of the actual price relatives A
r is the product-moment correlation coefficient of predic
ted and actual price relatives.
The first two terms on the right side of equation 5.9 are equal to zero if
a and b in 5.8 are equal to zero and one, respectively. The portion of the
“ _ 2 mean-squared error measured by (R - R) is labeled constant bias, and the 2 term (S - rS.) is due to proportional bias. The remaining portion of the p A 2 2 mean-squared error is equal to (1 - r )SA . This part of the squared-error
can be reduced by finding a prediction which is more highly correlated with 2 2 actual price relatives. For perfect correlation r = 1 and 1 - r = 0 .
This part of the mean-squared error is referred to as the correlation or
validity portion.
The relative performance of the prediction models with respect to
"^Henri Theil, p. 33. 131
constant bias are summarized in Table 8 . One Important constraint must be
emphasized, however. In all cases, the smoothing parameters were selected
on the minimum mean-squared error criterion. As a result, it is possible
that constant bias may be eliminated for any particular prediction model, but only by increasing the mean-squared error, which is undesirable. This
constraint is necessary, but it must be recognized in evaluating the rela
tive performance of the three prediction models with respect to constant
bias, proportional bias, and validity.
Use of Model No. 3 results in the smallest constant bias, relative to
that of Models No. 1 and 2, for the sixty-one companies tested. The rank
ings and the Kendall Coefficient of Concordance are both significant at the
99.9% confidence level. Although statistically significant, the Coeffi
cient of Concordance is very low, indicating a low level of agreement among
the rankings. Model No. 3, using a fixed smoothing constant for each com
pany, is uniformly the best prediction model with respect to constant bias,
except for the textile industry. It has previously been argued that the
• *\ textile industry may have some characteristics which suggest that a differ
ent model should be used for this industry.
Bankings at the industry level indicate that the results for beverage producers, textile producers and tire and rubber goods producers were quite mixed, and not statistically significant.
In summary, Model No. 3 resulted in the smallest amounts of constant bias for.all but the textile industry. The other model rankings were mixed and provided little additional information.
Using Model No. 2 minimizes the amount of proportional bias of the predictions for sixty-one companies sampled, as is reported in Table 9. 132
The ranking across all companies, and the Kendall Coefficient of Concor
dance, are significant at the 99.9% confidence level, but the amount of
agreement is only 0.18 of a possible 1.0. Proportional bias results from
a failure to pick up trends in the stock price relatives.
These results imply that some trends may exist in common stock price
relatives, but not to the extent that a trend model such as Model No. 2
actually outperforms the constant models (1 and 3) with respect to total
prediction errors.
The industry rankings indicate somewhat conflicting judgments. Of the
five industry classifications, Model No. 2 performed best in three classi
fications and Model No. 3 performed best in two classifications. Model
No. 3 performed best in the Cement Industry and the results were clearly
significant at the 99.9% confidence level. For the Tire and Rubber Goods
Industry, Model No. 3 performed best, but the results were not statisti
cally significant at the 95% confidence level, so the ranking may have been
caused by chance factors alone. Model No. 2 performed best for the other
three industry groups and the results were significant at the 95% confi
dence level or higher.
In summary, the generally superior performance of the linear trend model in reducing proportional bias indicates that some short runs or
trends probably exist in the price relatives. However, the superior per
formance of the constant models with respect to mean-squared error indi
cates that such trends are not sufficiently long or pronounced to be of value in predicting future price relatives.
Using Model No. 1, with shifting smoothing constants, results in the least amount of correlation error when compared to Models No. 2 and 3. 133
Although this ranking and the Kendall Coefficient of Concordance are sig nificant at the 99% confidence level, the degree of agreement is so low
that it is of little use.
Except for the cement industry, none of the industry rankings are
significant at the 95% confidence level. Given this low level of agreement at the industry level, few conclusions can be drawn.
One result of interest is the generally sqperior performance of the shifting smoothing constants on an overall basis and on an industry by in dustry basis. In the three previous comparisons, the fixed smoothing con stants outperformed shifting smoothing constants on a model by model basis.
These results suggest that the predictions formulated by using shifting smoothing constants are more highly correlated with actual price relatives than the predictions formulated by using fixed smoothing constants. How ever, due to the resulting constant and proportional bias, the use of shifting smoothing constants is not preferred to fixed smoothing constants in minimizing mean-squared prediction error.
The influence of smoothing constant selection on the mean-squared prediction error.— Prior to the testing of the prediction models using equation constants estimated by means of exponential smoothing, the two methods of selecting smoothing constants were expected to yield very simi lar results. This expectation was based on the belief that after the ini tial ten year period 1948 through 1957, a particular smoothing constant would perform best thereafter. If this condition were met, then the opti mal smoothing constant for the prior ten years should be about equal to the smoothing constant for the entire prediction period. In other words, we expect stability in the process which generates the price relative 134
series, which means the best smoothing constant for any ten year period within the 1948 through 1967 period will be equal to, or nearly equal to,
the best smoothing constant for the entire period.
The test results indicate that the method used to select smoothing
constants for each of the three prediction models has a significant in
fluence on the resulting measure of uncertainty, the mean-squared predic
tion error. Using the mean-squared prediction error resulting from the use
of shifting smoothing constants as the base measure, uncertainty can be
reduced by 47.9%, 65.9% and 58.6% for Models No. 1, 2, and 3 respectively.
In addition to being large, these differences are statistically signifi
cant at the 99.9% confidence level and are not considered to be the result
of chance.
The 58.6% reduction in the average mean-squared error associated with
Model No. 3 is especially important because it is the best overall predic
tion model. The difference is even more significant when expressed in
terms of the average mean-squared error resulting from the use of a fixed smoothing constant for Model No. 3. The uncertainty associated with the use of shifting smoothing constants is more than 240% of the uncertainty associated with the use of a fixed smoothing constant for each company.
As reported in Table 13, similar results were found when comparisons are made of mean absolute errors.
The calculation of the prediction errors associated with the single best smoothing constant for the period 1948 through 1967 provides informa
tion about the best model of stock price relative movements over that pe riod, as well as indicating the maximum prediction efficiency associated with the smoothing models. The resulting mean-squared prediction errors 135
are a standard by which we can evaluate our efforts to select an optimal smoothing constant on a year by year basis.
Model No. 3 was judged the best naive prediction model after the series of years for which predictions were being made. An investor would not have known which naive model was going to perform best, and so we have eliminated an element of the uncertainty which an investor would have faced over the years 1958 through 1967. However, Model No. 3 is a simple model and it was expected to perform well as compared to Models No. 1 and
No. 2. Model No. 3 was not much better than Model No. 1, so the results are relatively insensitive to the choice between these two models.
Predictions based on prior means.— If the price relatives are mean- reverting, then a sample estimate of the mean is the best prediction one has. Accordingly, the fourth naive method of predicting is
10 (5.10) R. = I R. J 9 xt ^ it-j
R^t is the price relative prediction for company i
(i = 1, 2, ..., 61) and year t (t = 1958, ..., 1967)
The results are different than those of the constant exponent model, thus supporting the hypothesis that the price relatives are mean-reverting
(Table 41). However, the difference was not statistically significant.
The average mean-squared prediction method is significantly greater than the average variance (Table 37) thus supporting hypothesis 2, which is that the variance does not completely reflect total uncertainty.
Performance of the Growth Model
The general form of the postulated growth model of stock price rela- 136
tives is presented in equation 4.35. In order to implement the model, specific measures of each variable must be made for each company. For example, corporate growth rate may be expressed as an average annual rate of sales growth, operating earnings growth, or cash flow growth, determined over the last two years, three years, or any period believed best.
A generalized stepwise regression program is used to determine which methods of measuring each variable are best. That measure which has the highest partial correlation with the past price relatives is the best mea sure available for prediction purposes.
The empirical models tested.— The actual variables selected in the stepwise regression tests for 1948 through 1957 varied from company to com pany and from industry to industry. To select the set of variables for each company on an individualized basis was not possible because of the cost and difficulty of doing so. Individual runs of the regression tests would require much more computer time, and thus more cost. Also, the data handling problems would be much greater for individualized processing.
In order to compensate for intercompany differences, more variables were included in the prediction equation and the elements of the variable set were allowed to vary from industry to industry. To illustrate, in the beverage industry, for many companies the best measure of groxfth was the average growth rate in operating income plus depreciation, and for other companies the best measure was the two-year growth rate in operating in come plus depreciation. Forcing both variables into the prediction equa tion insures that the most significant variables for all of the companies are included in the equation. If a variable has little relevance for any particular company, then the coefficient for that variable should not vary 137
significantly from zero, and it will not affect the predictions very much.
Six variables were selected for each industry classification based on
the frequency and order in which these variables entered the stepwise re
gression equation for the years 1948 through 1957, for each company in the
industry. Once an industry variable set was selected, it remained fixed
throughout the test period 1958 through 1967.
For the beverage industry, the most significant variables are the
difference between the last annual price relative and the average price
relative for the past five years, the standard deviation of the price rela
tives over the past five years, the variance of changes in operating income about the average change for the last five years, the average growth rate
in operating earnings over the past six years, the annual growth rate ex perienced in the past two years in operating income plus depreciation, and
the rate of growth of gross national product expressed in current dollars.
In the formulation o f (4.35) it was expected that the most recent price relative would be partially correlated with the current relative. The empirical tests indicate that the difference between the most recent price relative and the average price relative is more significantly related to the current price relative than the most recent price relative alone. This finding is consistent with the earlier finding that the smoothing constants tended to be very small for the naive prediction models. The conclusion drawn is that the price relatives tend to be mean-reverting for many com panies. This statement cannot be generalized to all companies because for several companies the optimum smoothing constant was equal to 1 .0 , indicat ing a random walk in the price relatives, and thus no mean-reversion. The same set of variables was most significant for the tire and rubber goods 138 industry, except that the growth rate of annualized fourth-quarter gross national product was used instead of annual gross national product.
The petroleum and cement industries exhibited some similarity with respect to the variables found the most highly correlated with the price relatives. The common variables were the difference between the last an nual price relative and the average price relative for the past five years, the average proportional price range over the past two years, the variance of changes in operating income about the mean change for the past five years, and the average growth rate in operating income plus depreciation for the past six years. The remaining two variables used in the petroleum industry are the average growth rate in net sales over the prior two years and the annual growth rate of gross national personal income over the pre vious year, expressed in current dollars. For the cement industry, the other two variables are the average growth rate in net sales over the prior six years and the growth rate of annualized fourth quarter gross national personal income. These are actually different measures of the same basic variables, so that these two industries are very similar. Average propor tional price range is a measure of uncertainty which is equal to the price range, divided by the beginning price in order to make the measure insen sitive to the absolute level or the share price.
The textile industry was unique in some respects, which is consistent with earlier findings with regard to the naive prediction models. The most significant variables on an industry basis are the most recent price rela tives, the relative price range over the prior five years, the variance of changes in operating income about the average change for the prior five years, the average growth rate in operating earnings plus depreciation for 139
the past six years, the average growth rate in net sales over the past two
years, and the growth rate in annualized fourth quarter gross national pro
duct. This is the only industry for which the most recent price relative
was significantly related to the current price relative, in all others,
it was the difference between the most recent price relative and the aver
age price relative for the prior five years. These results are not con
sistent with the fact that the optimal smoothing constant for most of the
textile companies was 0 .0 , thus using the average price relatives of an
earlier period to predict current price relatives as opposed to using last
year's price relative to predict the current year's relative. At the pre
sent time we do not have an adequate explanation of these somewhat con
flicting results for the textile industry.
Bounded optimism and pessimism.— Preliminary trials of the postulated
growth model indicated that occasionally an extreme prediction was being
made; one that would exceed all reasonable expectations by a knowledgeable
investor. What might be called a programmed judgment factor was added to
the price relative predictions in order to prevent these extreme predic
tions .
The programmed judgment factor took the form of an upper and lower bound for each prediction. The first stage of the bounding process in
volves finding the maximum and the minimum price relatives over the prior
five years for each common stock. A visual examination of these upper and
lower bounds indicated that a wide range of values might appear during any
five year period, and that the range of these limits was still very great.
The second stage of the bounding process is designed to further reduce
the extreme values that may be predicted. An arbitrary maximum upper 140 bound of 0.80 were established after a visual examination of the price rela tives for the sixty-one companies during 1948 through 1957. Price rela tives of 1.50 were not unusual, but they seldom occurred in consecutive periods or within a five year period. The same is true of very low price relatives.
The net result of these two parts of the bounding process is the set ting of upper and lower limits on the predicted price relatives. The upper limit is equal to the maximum price relative for that company over the prior five years, or 1.50, whichever is smaller. The lower bound is equal to the minimum price relative observed for that company over the prior five years, or 0.80, whichever is larger. These programmed judgment factors were important with respect to the predictions of six companies. The pre diction equation coefficients for these six companies had large standard errors and tended to fluctuate greatly from year to year.
Comparative predictive performance.— The relative mean-squared errors and absolute errors are compared for sixty-one companies to determine if the postulated growth model resulted in smaller prediction errors than re sulted from use of Model No. 3, with shifting smoothing constants. The average results for these two models can be seen in Table 32, and Table 33.
Tests of significance are summarized in Table 38.
It is safe to conclude that Model No. 3 resulted in smaller prediction errors than resulted from the growth model. The average mean-squared error for the growth model was 0,034, or about 20%, larger than resulted from
Model No. 3. These results were significant at the 99.9% confidence level for the Wilcoxen matched-pairs, signed-ranks test, and are also significant at the 99% confidence level for the t-test. The cement industry results 141 were clearly mixed with insignificant differences. However, for the other industries and for all companies combined, the null hypothesis of no dif ferences is rejected, and the results accepted as evidence of the superior performance of Model No. 3.
The average absolute error resulting from use of the growth model was
0.0432 or about 14% larger than the average absolute error resulting from the use of Model No. 3. As was true for the mean-squared error, these results are significant at the 99.9% confidence level for the Wilcoxen matched-pairs, signed-ranks test, and are also significant at the 99% confidence level for the parametric t-test.
The growth model prediction errors contained less constant bias than the smoothing model but greater prediction error due to proportional bias and low correlation. Eliminating proportional bias may be the most promising correction factor available to reduce the growth model predic tion errors.
Linear correction factors, the distribution characteristics of the prediction errors, and possible methods of improvement are discussed in the subsequent three sections.
Optimal linear correction parameters.— One additional method of re viewing the prediction results is to examine the linear correction factors associated with the set of actual price relatives and the set of pre dicted price relatives. The results of the determination of the correction constants are summarized in Table 29. Apparently, the correlation was low between the actual price relatives and the predicted price relatives. 142
This observation is supported by the low value of b (see Table 29)
and the large value of a. Ideally, a = 0 and b = 1. The results of the
simple analysis of variance summarized in Table 27 are similar. The with
in group regression is not significant at the 95% confidence level, even
though the group effects are significant. These two tests suggest that
the prediction performance of the growth model is very poor.
However, when each prediction was adjusted for the difference
between the average price relative and the average predicted price rela
tive, the relationship of each predicted relative to the relevant actual
price relative was quite good (see Table 28). The regression of R^t on
A R^t was forced through the origin. The regression coefficient was 0.96149
(b = 1.0 is ideal) and the squared multiple correlation coefficient was
0.8861 out of a possible 1.0, which is relatively good. If the amount
of constant bias (a^) is relatively stable over time, then there is some promise that the use of linear correction factors may be used to improve
the prediction results of subsequent studies.
Prediction error distribution tests.— The nature of the distributions which represent the behavior of economic variables have long been of in terest and have received much attention in recent research. Benoit 127 Mandlebrot has renewed this interest because he has argued that price changes for various economic time series tend to be more leptokurtic
(long-tailed) than would be expected for the Gaussian distribution (normal).
127 Benoit Mandlebrot, "The Variation of Some Other Speculative Prices," The Journal of Business, Vol. XL, No. 4 (October, 1967), pp. 393-413. 143
Various authors have explained the apparant non-normal behavior of stock price changes in various ways. Mandlebrot has argued that price changes are represented by the general class of distributions called Stable
Paretian, of which the Gaussian Distribution is a special case. A most significant feature of this class of distributions is that of an un defined variance. Regression and many other statistical models are based on the existance of a finite variance measure, which only exists for one special case of the Paretian Distribution, which is the Gaussian Distri bution.
The distribution of price relatives and the distribution of price relative prediction errors are both studied to determine if these vari ables can be described by the Gaussian Distribution, and if not, how they deviate from the Gaussian Distribution. The prediction errors which result from the use of Model No. 3, using shifting smoothing constants, are not normally distributed according to the Kolmogorov-Smirnov test, as explained in Table 15. When plotted, the errors appear to be nearly bell-shaped, but that is not the case.
The most interesting aspect of this test is the nature of the devia tion from normality. It is not the tails that are more heavily weighted than the normal or Gaussian Distribution, it is the concentration at the peak that is greater than expected. These results are very different from those expected because of Mandlebrot*s work. Because there is little evidence of a very large or undefined prediction error variance, the use of mean-squared error instead of average absolute errors is preferred 144 because of the utility function assumed.
Similar results were found for the predicted price relatives made by Model No. 3 using shifting smoothing constants.
As reported in Table 14, the predicted price relatives tended to be more concentrated about the mean predicted price relative than one would expect for a normally distributed random variable. As a result, the variance of the predicted price relatives tends to have a finite variance, and there is no evidence that the statistical models used are not applicable to this data.
Similar results are obtained for the prediction errors that result from the use of Model No. 3 with a fixed smoothing constant for each company. The distribution is more heavily concentrated about the mean prediction error than would be expected for a normally distributed random variable (see Table 17). The predicted price relatives for
Model No. 3, using a fixed smoothing constant for each company, are similarly distributed (see Table 16).
The errors associated with price relative predictions made by the growth model for the sixty-one companies for 1958 through 1967 were not normally distributed. The errors were more concentrated about the mean error than one would expect for a normal distribution, as was also the case for the prediction errors from Model No. 3 (see Table 23).
The predicted price relatives for all sixty-one companies from 1958 through 1967 were less concentrated about the mean predicted relative than one would expect for a normally distributed random variable. These 145 results do not parallel the results obtained for prediction Model No. 3.
Greater weight is included in the tails of this distribution because the predictions are bounded, so that there are numerous predictions equal to
0.80 but none less than 0.80, and numerous predictions equal to 1.50, but none are larger. The distribution of these predictions is thus nearly uniform, with more weight at the ends than one would expect for a uniform distribution.
The 610 price relatives observed over the ten year test period were tested for normality. The relatives deviate significantly from normality in at least two ways. First, the price relatives are more heavily con centrated about the mean price relative then one would expect if the price relatives were normally distributed. Second, the actual price relatives are skewed toward the larger relatives.
The finding that prediction errors show no sign of excessively large variances is significant, and it justifies the use of mean-squared pre diction error and variance as measures of uncertainty.
Implications for Measuring Uncertainty
The mean-squared prediction error for a sample of predictions and the variance of price relatives about the mean sample relative are two surrogate measures of uncertainty that have been discussed in this study.
Using mean-squared error as the more logical measure of uncertainty, the variance of price relatives, for each corporation, is evaluated in terms of its appropriateness as a measure of the uncertainty inherent in price relative predictions. 146
Components of uncertainty.— The investor faces two problems when pre dicting price relatives and evaluating the uncertainty of those predictions.
The movement of price relatives over time can be represented by various stochastic models. One element of uncertainty facing any investor is the purely random element of a price relative's changes. This element of un certainty is inherent in the market process and cannot be eliminated.
The usual measure of uncertainty associated with this component of total uncertainty is the variance, or some equivalent but more elaborate measure, when the process is very complex. The second element of uncertainty is determining which stochastic process best fits the time series of prices or price relatives. Many models of stock price changes have been pos tulated and tested without conclusive results. The same results are to be expected for the study of price relatives. Likewise, there are many theoretically sound price relative prediction models, which vary in their effectiveness in predicting price relatives. Total uncertainty is thus a combination of the stochastic nature of a price relative, and man's inability to determine or fabricate the best possible model of stock price relative behavior, or the best possible predictions of the stock price relatives.
Four naive models of stock price relative behavior over time, and a growth model based on economic forces, are used to formulate predictions in this study. The resulting prediction errors are caused by the stochas tic nature of the pricing process as well as by the errors that result from incorrectly formulating the model. The variance of price relatives 147 about the mean price relative for some sequence of periods tends to under state the uncertainty associated with the formulation of predictions be cause the mean, or expected value, is known, for a particular sample, after the fact, and only then.
Comparisons of these measures of uncertainty are summarized in
Table 37. The average mean-squared prediction error resulting from use of the growth model is more than 33% greater than the price relative variance, using the former measure as the base value. When expressed in terms of the variance, the average mean-squared error is about 51% greater than the average variance. Less extreme, but similar results are found for Model No. 3, when the optimal smoothing constants are sel ected on a yearly basis. The average mean-squared error for Model No. 3 is 17.2% greater than the variance. Expressed in terms of the smaller base, the average mean-squared error is 121% of the average variance.
Unless better prediction models can be formulated, or better intuitive predictions can be made, these results indicate that the variance under states the amount of uncertainty inherent in price relative prediction.
It may be possible to find better models of the behavior of stock price relatives, and it may be possible to find much better predictions of price relatives. Progress may be made to the point that prediction errors are less than deviations from the mean, but the present evidence suggests that the variance estimate understates the level of uncertainty associated with price relative prediction.
Variance as related to mean-squared error.— If the variance under- 148
states true uncertainty by some constant amount or proportion, it is a
relatively simple matter to calculate the variance, and then apply some
correction factor in order to better estimate the level of uncertainty
associated with the price relative being predicted. Moreover, if any
functional relationship exists between the variance and the true, but unknown, uncertainty, then that measure of.true uncertainty can be es
timated by means of the variance estimate and the known relationship between total uncertainty and the variance.
At the present time, we have two surrogates for true uncertainty, one of which is the mean-squared prediction error of Model No. 3 using shifting smoothing constants. Using this mean-squared prediction error as one available surrogate for true uncertainty, we tested various relation ships to determine if mean-squared error could be estimated by means of the variance. These attempts were relatively unsuccessful. Linear regression models failed to produce any model in which the variations in variance measures were useful in explaining more than 20% of the variation in mean-squared error for the sixty-one companies. The residuals were linearly correlated with the value of the mean-squared error so that the smaller values of mean-squared error were being overstated by the model and larger values were being understated. We do not argue that no such relationship exists, we only argue that the relationship is not obvious, and that it is unknown at present.
In summary, no relationship was found between the variance of a company's price relatives over the ten year test period and the mean-squared 149 prediction errors for that period. The tentative conclusion is that no such functional relationship exists, for this measure of true uncertainty.
Risk class discrimination.— As discussed in Chapter II, the selec tion of investments involves a measure of uncertainty. Unless it can be shown that alternative measures of uncertainty are linear functions of one another, it is likely that the selection of uncertainty measures will influence the portfolio selected from a given set of opportunities.
Because the squared correlation coefficient is invariant under linear transformation, it is our belief that a linear function of the variance would result in the selection of the same portfolio as would result from the use of the variance as the measure of uncertainty. As mentioned above, no such linear transformation of the variance was found, assuming that the mean-squared error is a good measure of true uncertainty.
Moreover, Modigliani and Miller have discussed the decomposition of investments into different risk classes, which requires only ordinal . 128 measurement of uncertainty. There is some evidence that even the selection of risk classes using the variance as the classification cri terion will lead to different results than would be obtained if the mean-squared error were the classification criterion.
Evidence that use of the variance to measure uncertainty may result in one classification, while use of the mean-squared error may result in another classification, is provided by the simple analysis-of-variance
■^^Modigliani, Franco, and Miller, M. H. "The Cost of Capital, Cor poration Finance, and the Theory of Investment," American Economic Review, XLVIII (June, 1958), pp. 261-279. 150
test summarized in Tables 19, 20, 21, and 22. The mean-squared error is
the dependent variable, while the variance is the independent variable,
thus providing sixty-one paired observations. Industry mean-effects
were significant at the 99% confidence level for both the variance and
the mean-squared error, which indicates that the average mean-squared
error differed from industry to industry, ds did the average variance.
The variance is then used as a control variable to determine if the mean-effects for the mean-squared errors can be explained by differences
in the variances. Referring to Table 19, the mean effects for the ob
served measures of the mean-squared errors can be explained by differences
in the variances. Referring to Table 19, the mean effects for the ob served measures of the mean-squared error cannot be explained by dif
ferences in the variance, so the null hypothesis must be rejected at the
99% confidence level. The implication of these findings is that the variance may not be a sufficiently complete measure of uncertainty for purposes of defining risk classes and assigning companies to those risk classes. At this point we cannot make any normative statements, we can only raise the problem and suggest that different classifications prob ably will result. Mathematical proofs of these statements and empirical tests of their validity are beyond the scope of this project.
Potential Improvements in the Growth Model Predictions
The prediction performance of the growth model is disappointing.
Most authors do not test their valuation models by actually making price relative predictions, so the performance of our model cannot be compared 151 to the performance of the other models that appear in the literature.
Nevertheless, the model is not useful in its present form. There are several approaches that one might use to improve the price relative pre dictions of an economic model. Most of these techniques are apparent in retrospect, even though they were not obvious at the start of the study.
Sample size increases.— The large standard errors of the regression coefficients undoubtedly contribute significantly to the average mean- squared prediction error resulting from use of the growth model. Increasing the sample size may reduce the standard error of each coefficient and thus reduce year-to-year fluctuations in these coefficients.
One method of increasing sample sizes is to combine the observations for all companies that are in one industry classification, thus obtain ing one set of equation coefficients per industry per year. Within indus try differences will still be significant because growth rates for the various companies will still differ, as will the measure of uncertainty.
For homogeneous industries this method should work well, but it probably is of little value in the case of conglomerates, and other loosely defined industry groups.
A second method of increasing sample size is to increase the number of years of data in the formulation of the regression model coefficients.
Two problems are created by the use of more data years, and these may de crease the performance of the growth model. One is the probable instability of relationships over long periods of time. If the relationships of indi vidual variables to the price relatives and to each other change over time, 152
then adding more observations may have a detrimental affect on the standard
errors of the coefficients. If the relationships are stable, then one
might expect the standard errors to decline, and prediction performance
to increase. It is difficult to predict the success of this method of
increasing sample size.
Variable measurement changes.— One of the more apparent opportunities
to improve the results of the approach used in this study is the continued
search for better measures of the variables used in the prediction model,
and the inclusion of some industry variables in the model. The particular measure of sales growth rate used in this study, or that of the uncertainty
of operating earnings, are not optimal in any true sense of the word. Much more research must be done in sales predictions, growth rate prediction
and other related measures before we can confidently state that the best measures of each variable have been included in the prediction model.
This type of searching never ends, but should result in moderately im proved results.
The more significant improvement of prediction results probably will be the result of including an industry activity variable in the prediction equation. For those industries in which annual industry predictions are made, in quantified form, the reduction in uncertainty should be much more than can be expected when such data or predictions are lacking. Here again, the more homogeneous industry classifications are the more desirable starting points for such efforts. The collection and processing of industry 153 information is a costly and time-consuming process, and the incremental cost may not be justified economically, even though predictions may improve.
Learning factors.— The most interesting source of potential im provement in the performance of the postulated growth model is that of learning. First, through an examination of price relative changes we may be able to set more scientifically determined upper and lower limits on the predicted price relatives. For example, if it is determined that price relative increases of 0.6 or more have a 0.5 probability, we establish the rule that the predicted price relative may not exceed last period's relative plus 0.6. Any larger change is considered highly unlikely, and thus should not be predicted. In this case we are incorporating know ledge about the distribution of price relative changes.
The inclusion of past prediction errors in the regression model and thus in the prediction model is a more direct application of the concept of learning. In this case we are admitting that the regression model does not fit the data quite as well as is postulated, because the errors are of no predictive value in theory. However, as a practical matter, the regression coefficients are based on data and relationships from the previous ten years, and they may not adjust to changing conditions as fast as is desired. Some additional information may be provided in prior prediction errors, and this method is an attempt to capture some of that information for prediction purposes.
A second method of incorporating the relationship of past price re lative predictions to past price relatives is to formulate predictions 154
as was done in this study, and then adjust these predictions for the
historical amount of bias associated with prior use of the growth model
predictions. The optimal linear correction of predictions was discussed
above, and this linear correction factor can be applied to each prediction,
as it is being made, in order to formulate an adjusted prediction. The
constants in the correction factor are estimated each year from the his
torical relationship of price relatives and predicted price relatives.
More esoteric approaches may exist, but the latter two approaches
mentioned are straightforward methods of incorporating information about
past prediction errors in current predictions. The true test of any one
of the suggested improvements is its ability to reduce expected uncertainty.
Extension of theory.— Our economic model of stock price relatives
includes measures of growth and stability, a prediction of the economic
environment next year, and an estimate of the probable short-run change
in the price relative. These are the same factors evaluated by the Value
Line Investment Survey, even though the variables are measured differently.
If use or usefulness is any criteria of the empirical validity of a con- * * » struct, then our approach to price relative prediction has empirical validity because investors and stock brokers purchase the Value Line pro jections. However, the value line approach lacks the derivation and the
analysis that should be characteristic of a formal model of stock price
relatives.
A more complete theory of the economic forces causing price relatives
to vary with changes in expectations, changes in the supply of money, or 155 other forces, may lead to improved price relative predictions. Given the heterogeneous nature of investors in common stock, there is absolutely no assurance that a more formal conceptual model of stock price relatives will result in better predictions. 156
APPENDIX A
A series of predicted price relatives can be compared with the series of those price relatives subsequently observed in order to calculate a series of prediction errors. Let R represent the common stock price JL U> A relative for security i during period t, and R^ represent the predicted value of The prediction error, denoted is the difference be tween the actual price relative and its predicted value where
(A-l) dlt-Rlt-Rlt
The i subscript will be dropped for convenience since we are discussing a general measure of uncertainty, not the specific value of a statistic for a particular company. For any n-year period, the mean prediction error is given by n (A-2) d -trx dt/n, and the variance of the prediction errors is given by
(A-3) S ^ « . J 1 (dt - d ) 2/(n-l)
Suppose the prediction errors associated with the price relatives of a particular security can be described by a random variable D, where the 2 true, but unknown, mean difference is and the expected variance is o^.
The mean-squared rror is defined as
(A-4) a2 - E(D-O)2 - E(D2) e
By definition, 157
(A-5) a2 = E(D - yD)2 = E(D2 - 2DyD + y2)
(A-6) cr2 = E(D2) - y2
Note that:
(A-7) a2 + y2 = E(D2) - y2 + y2
(A-8) a2 + y2 = E(D2) = a2
The relationship defined in A-8 is true by definition of the terms. 2 2 2 A sample estimate of a , denoted S , is an unbiased estimate of a r e e e 2 2 if and only if E(Sg) = ae * also true that E(A + B) = E(A) + E(B)
for two random variables A and B. Our problem is to find an unbiased es- 2 timate of a . e We know that
(A-9) E(d - yQ)2 = o2/n
By manipulating the terms of A-9, we have that
(A-10) E(d2) = 02/n + y2
129 Another standard relationship in mathematical statistics is
(A-ll) E(S2) = o2
2 130 Define the biased estimate of as
(A-12) s2 = (n - 1) S2/n, so that d d
(A-13) E(ns2/(n - 1)) = E(S2) = a2
129 Hogg, p. 39.
130Hoel, p. 198. 158
2 2 The sample estimate of 0g is denoted Se> and can be decomposed as follows:
CA-14) s2 - trx d2/n - ( dt/„)2 + j 1 (dt - d)2 n
n (d. - d)2 d 2 _ n d because E------E - 2d E- — + d , t=l n n t-i n
n (d-d)2 n d 2 __ and thus E------= E. — - d t=l n t=l n
Because the expected value of a sum is equal to the sum of the expected values,
(A-15) E(S2) = E(d2) + E(s2)
(A-16) E(S2) = a2/n + y2 + E[(n - 1) S2/n]
(A-17) E(S2) = a2/n + y2 + (1 - £> E(S2)
(A-18) E(S2) = a2/n + y2 + a2 - a2/n
2 2 2 2 (A-19) E(Se) = aD + yjj» which is equal to a so we have shown that
2 2 2 2 (A-20) E(S ) = o , so S is an unbiased estimate of a . e e e e 2 It should be noted that adding the unbiased estimate of to the square of the unbiased estimate of y^ does not provide an unbiased estimate 2 2 of (yQ + oD), which is to say
(A-21) E(d2 + S2) * a2 + y2
The left side of expression A-21 does equal
(A-22) E(d + S2) ■ cr2/n + y2 + a2 which is biased by the amount a2/n. 159
APPENDIX B
Let R^t be the price relative for security i (i = 1, 2 ..... 61) for time period t (t = 1948, 1949, ..., 1967). Further, assume that the price relatives of each company can be represented by a random variable
- ]_oi R^, with a sample mean value of R^. The difference between each price relative and R is itself represented by a random variable, X. Let
(B—1) xt = Rt - R
The x^'s form a purely random series; one of random deviations from a norm.^^ The expected value of X is zero, because E(R - R) =0. Similar ly, let
(B-2) yfc = xfc+1 - xt, where
(8-5) *t+1 - xt = (Rt+1 - R) -