2009:069 MASTER'S THESIS

Modeling CLV for Financial Service Providers - Case of Karafarin Bank

Roudabeh Gharaee

Luleå University of Technology Master Thesis, Continuation Courses Marketing and e-commerce Department of Business Administration and Social Sciences Division of Industrial marketing and e-commerce

2009:069 - ISSN: 1653-0187 - ISRN: LTU-PB-EX--09/069--SE

MASTER'S THESIS

Modeling CLV for Financial Service Providers: Case of Karafarin Bank

Supervisors: Dr. Amir Albadvi Dr. Deon Nel Referee: Dr. Keramati, Dr. Mohammadi, Dr. Kazemzadeh

Prepared by: Roudabeh Gharaee

Tarbiat Modares University Faculty of Engineering Department of Industrial Engineering

Lulea University of Technology Division of Industrial Marketing and E-Commerce

MSc PROGRAM IN MARKETING AND ELECTRONIC COMMERCE Joint

2009

1 Abstract

Today companies are urged to acquire and retain long-term relationships with customers rather than having discrete transactions with them, and this fact has led managers to wonder how they could measure the long-term value of a customer in a way that would represent both the relationship benefits and accounting profits of a customer.

These conceptual developments along with growing competitiveness of Iranian financial services industry, have urged top management of this industry to seek more accurate ways on how to better manage their relationships with customers and prevent losses by distinguishing high value and low value customers.

Although there are many value measurement tools developed by researchers in different industry settings, there is no such model specified to financial services which brings into account the risk of relationship. This research has focused on developing a value measurement model by which a bank’s management could better decide whether a business credit facility applicant is worth the resources s/he receives. This model is developed based on “customer lifetime value” concept and is risk-adjusted to overcome this common flaw of such models. The risk is modeled by both Delphi method (qualitative risk) and logistic regression method (quantitative risk) to partially resolve the unreliability issues concerned with financial statements in Iran. The model is then tested through a case study- Karafarin Bank (KB)- and new applications such as loan pricing and collateral setting based on this model are suggested to be worked out by KB’s management.

Keywords: Customer lifetime value, Delphi method, Qualitative risk, Case study, Mathematical modeling

2 Acknowledgement:

This thesis is dedicated to those who have supported me in different aspects; academically, psychologically, and of course financially!! Among these people, my husband and my mother helped me in all aspects and without them I wouldn’t have been able to accomplish this dissertation. I owe this achievement to them in first place. I would also like to convey my sincere appreciation to my supervisor, Dr. Albadvi, who dedicated enough time to my piece of work so that I could get my desired research quality. Many thanks to Karafarin Bank and all Banking experts on our Delphi panel who devoted great amount of time and expertise for completion of this thesis. Hope the results of this research can improve the marketing performance at Iranian financial services industry.

3 Table of Content Chapter 1 ...... 10 Introduction ...... 10 1. Introduction ...... 10 1.1 Problem Definition ...... 11 1.2 Importance of CLV Measurement for Organizations ...... 12 1.3 Research Motivations ...... 13 1.4 Research Objectives ...... 15 1.5 Research Design ...... 15 1.6 Model Development Phase ...... 18 1.6.1 Qualitative Risk Score Modeling ...... 19 1.6.1.1 Delphi Method Steps ...... 23 1.6.2 Quantitative Risk Score Modeling ...... 25 1.7 Model Fine-tuning Phase ...... 25 1.7.1 Case Study ...... 25 1.7.1.1 KB’s Background ...... 28 1.7.1.2 Credit Issuance Process at KB ...... 28 1.7.2 Data Collection Method for the Case Study ...... 30 1.7.2.1 Secondary Data ...... 30 1.7.2.2 Types of secondary data ...... 30 1.7.2.3 Evaluating secondary data ...... 32 1.8 Expected Contribution of the Research ...... 33 1.9 Abbreviations and expressions ...... 34 1.10 Structure of the Report ...... 34 1.11 Summary of the Chapter ...... 35 Chapter 2 ...... 36 Literature Review on Customer Lifetime ...... 36 Value Applications and Models ...... 36 2. Customer Lifetime Value ...... 36 2.1 Value Measurement Background ...... 37 2.1.1 Customer Relationship Management ...... 37 2.1.2 Relationship Marketing ...... 39 2.2 Historic Data Analysis Methods ...... 39 2.2.1 Top-down Approach ...... 39 2.2.2 Bottom-up Approach ...... 41 2.2.3 Limitations of Historic Data Analysis Methods ...... 42 2.3 CLV Definition ...... 44 2.3.1 Benefits of CLV Measurement ...... 46 2.3.2 Limitations for CLV measurement ...... 47 2.4 CLV Applications ...... 48 2.4.1 Segmentation/ Product Recommendation/ Retention ...... 49 2.4.2 Resource Allocation ...... 49 2.4.3 Merger and Acquisition ...... 50 2.4.4 Customer Equity ...... 50 2.4.5 Marketing Campaign ...... 52 2.5 CLV Calculation Models ...... 52

4 2.5.1 Models by Blattberg and Deighton (1996); Berger and Nasr (1998); Gupta, Lehmann (2003); Blattberg, Getz and Thomas (2001); and Rust, Lemon and Zeithaml (2000) ...... 53 2.5.1.1 BD model ...... 54 2.5.1.2 BN model ...... 55 2.5.1.3 GL Model ...... 55 2.5.1.4 BGT Model ...... 56 2.5.1.5 RLZ Model ...... 58 2.6 Models by Stahl (2003) and Wu (2005) ...... 59 2.6.1 Framework developed by Stahl (2003) ...... 60 2.6.2 Model by Wu (2005): ...... 62 2.7 Models by Hwang (2004) and Kim (2006) ...... 64 2.8 Models by Venkatesan (2004) and Kumar (2006) ...... 69 2.8.1 Modeling the Purchase Frequency for Each Customer ...... 71 2.8.2 Modeling the Contribution Margin for Each Customer ...... 72 2.8.3 Computing the Marketing Cost for Each Customer ...... 74 2.9 Models by Pfeifer (2000) and Haenlein (2007) ...... 75 2.10 Model by Fader (2007) ...... 77 2.11 Model by Crowder (2007): ...... 79 2.12 Summary of the Chapter ...... 81 Chapter 3 ...... 83 Model Development ...... 83 3. Model Development ...... 83 3.1 Relationship Risk Score Modeling ...... 84 3.1.1 Delphi Method Procedure ...... 85 3.1.1.1 Choosing Experts ...... 85 3.1.1.2 Brainstorming the Experts ...... 88 3.1.1.3 Narrowing Down Factors ...... 93 3.1.1.4 Grading Factors ...... 96 3.1.2 Qualitative Risk Score Estimation ...... 97 3.1.3 Quantitative Risk Score Estimation ...... 97 3.2 Probability of Default Modeling ...... 98 3.3 Revenue Modeling ...... 98 3.3.1 Repurchase Rate Estimation ...... 99 3.4 Cost Modeling ...... 99 3.5 WACC Modeling ...... 99 3.6 CLV Estimation Model ...... 100 3.7 Summary of the Chapter ...... 101 Chapter 4 ...... 102 Model Fine-tuning ...... 102 4. Model Fine-tuning ...... 102 4.1 Qualitative Risk Score Computations ...... 103 4.2 Quantitative Risk Score Computations ...... 108 4.3 Real PD Calculations ...... 109 4.4 Lifetime Estimation ...... 112 4.5 Loan Demand Growth Estimation ...... 113

5 4.6 Repurchase Rate Estimation ...... 114 4.6 WACC, Cost of Money, and Administration Cost Estimations ...... 114 4.7 CLV Calculations ...... 114 4.8 Fine-tuning the Model ...... 119 4.9 Summary of the Chapter ...... 124 Chapter 5 ...... 125 Conclusion and Future Research ...... 125 5. Conclusion and Future Research ...... 125 5.1 Over view of the Research ...... 126 5.1.1 Research Design ...... 129 5.2 Research Outcomes ...... 131 5.4 Research Applications for Financial Service Providers ...... 136 5.5 Research Limitations ...... 138 5.5.1 Limitations of Delphi Method Procedure ...... 138 5.5.2 Limitations of Scoring ...... 138 5.5.3 Limitations of CLV Modeling ...... 139 5.6 Future Research Directions ...... 139

6 List of Tables Table 1.1. Comparison of Delphi method and traditional survey adapted from Okoli & Pawlowski (2004) ...... 20 Table 1.2. Lending techniques adapted from Baas and Schrooten (2006) ...... 29 Table 1.3. Percentage of customers who have had defaults at KB ...... 33 Table 1.4. Total non-performing yearly loan to total yearly loan ratios ...... 33

Table 2.1 Definitions of Lifetime Value ...... 45 Table 2.2. Variables included for predicting purchase frequency and/or gross contribution ...... 73

Table 3.1. Initial list of experts’ factors including their reasons ...... 89 Table 3.2. Categorized list of attributes ...... 92 Table 3.3. Extracted Delphi attributes that met the consensus criteria ...... 95 Table 3.4. Final extracted attributes of Delphi process and their mean weights ...... 96

Table 4.1. Qualitative risk scores and performances ...... 104 Table 4.2. Strength of the differentiators ...... 107 Table 4.3. Probability of default of customers for each quadrant ...... 111 Table 4.4. Probability of Default of each customer ...... 111 Table 4.5. Average lifetime of different economic sectors ...... 112 Table 4.6. Rounded estimates of lifetimes ...... 113 Table 4.7. Current value and future value of customers ...... 115 Table 4.8. Total CLVs of customers ...... 117

Table 5.1. CLV Applications ...... 126 Table 5.2. CLV model characteristics, assumptions, and industry usage ...... 127 Table 5.3. CLVs sorted based on lowest current value to the highest ...... 133

7 List of Figures Figure 1.1. Flowchart of the research ...... 18 Figure 1.2. Flowchart of Delphi Method adapted from Okoli and Pawlowski (2004) ..... 24

Figure 2.1. The four components of CLV adapted from Stahl (2003) ...... 60 Figure 2.2. Measurement of CLV adapted from Stahl (2003) ...... 61 Figure 2.3. Conceptual framework for measuring and using CLV ...... 62 Figure 2.4. Conceptual framework of Hwang ...... 65 Figure 2.5. Segmentation by using only LTV values ...... 66 Figure 2.6. Segmentation by using LTV components ...... 66 Figure 2.7. Segmentation by considering both LTV values and other information ...... 67

Figure 3.1. Flowchart of expert selection for Delphi Method adapted from Okoli & Pawlowski (2004) ...... 86 Figure 3.2. Risk experts’ age and work experience ...... 87 Figure 3.3. Risk experts’ educational background ...... 87 Figure 3.4. Credit experts’ age and work experience ...... 88 Figure 3.5. Credit experts’ educational background ...... 88 Figure 3.6. Factors’ votes of first round ...... 94 Figure 3.7. Factors’ votes of second round ...... 94 Figure 3.8. Flowchart of qualitative risk score modeling ...... 97

Figure 4.1. Risk scores’ distribution of customers regarding quantitative and qualitative scores ...... 110 Figure 4.2. Sensitivity percentage of Future Value regarding remaining years of lifetime ...... 120 Figure 4.3. Sensitivity percentage of Future Value regarding repurchase rate ...... 121 Figure 4.4. Sensitivity percentage of Future Value regarding different growth rates .... 122 Figure 4.5. Sensitivity of Future Value regarding different PDs ...... 123 Figure 4.6. Future Value for different growth rates and different PDs ...... 123

Figure 5.1. Flowchart of the research ...... 130 Figure 5.2. The risk score matrix ...... 131 Figure 5.3. Customer values based on two value dimensions ...... 135 Figure 5.4. Segmentation based on CLVs ...... 136

8 List of Appendices Appendix 3.1. Questionnaire 1 of Delphi process ...... 144 Appendix 3.2. Questionnaire 2 of Delphi process ...... 145 Appendix 3.3. Questionnaire 3 of Delphi process ...... 149 Appendix 3.4. Experts’ Votes of the two rounds ...... 151 Appendix 3.5. Questionnaire 4 of Delphi process ...... 152 Appendix 3.6 . Risk Attribute Weights Assigned by Delphi Experts ...... 154 Appendix 4.1. Qualitative scores of the firms ...... 155 Appendix 4.2. Quantitative scores of the firms ...... 157 Appendix 4.3. Geometric average of four consecutive year loan demand ...... 159 Appendix 4.4. Repurchase rate of customers ...... 160 Appendix 4.5. Future value function for CLV calculation ...... 163

9

Chapter 1

Introduction

1. Introduction

Preface: For many years, customer accounting profitability had been used as the criteria for uncovering unprofitable customers. By formation of the new concept of CRM in 1980’s and its revolutionary applications for marketers, new age of marketing started to grow by which, making a sale was just the beginning of a relationship with customer, and not the end. This phenomena, called relationship marketing, then urged companies to acquire and retain long-term relationships with customers rather than having discrete transactions with them. This substitution of transaction marketing by relationship marketing led managers to wonder how they could measure the economic value of a

10 customer in a way that would consider both the relationship benefits and accounting profits that the customer brings into the organization. Customer Lifetime Value (CLV), is the ``excess of a customer's revenues over time over the company costs of attracting, selling, and servicing that customer'' (Kotler and Fox, 1995) which is said to be a more comprehensive metric for profitability measurement that could to some extent satisfy this consideration. This chapter will define this new concept briefly, discusses the motivations and goals of choosing this subject, and overviews the research design of this project.

1.1 Problem Definition

Ryals (2002) states that single period customer profitability analysis often uncovers unprofitable customers. For example, 20% of a retail bank’s customers may account for more than 100% of its profit. Sometimes these are new customers, buying infrequently and in small amounts, who will become profitable as the relationship develops and sales increase and/or costs reduce. This kind of unprofitable customer is not an issue; the rationale for continuing to serve them is based on future profit potential. However, there is another type of unprofitable customer. This is the customer who is unprofitable in the current period and also in the future, perhaps because customers are becoming more demanding, have higher expectations, more knowledge, more buying power, and are more sophisticated in their buying behavior (Ryals, 2002). This raises the central issue of: why do banks and other retailers continue to do business with unprofitable customers? Is there something wrong with customer profitability calculations? Or are there other benefits that retailers can obtain from their customer relationships, over and above the economic benefits? We will argue in the next chapter that, in fact, both are true. Not only are there problems with measuring customer profitability; there are also other benefits that retailers can obtain from their customer relationships. This means that profitability may not be the true measure of the value a customer creates for an organization. Instead, measuring future cash flows of the customer over his/her lifetime and discounting it to present, considering the risk, would be a better metric for the value created by customer. This new metric then is called Customer Lifetime Value which according to Rosset et al. (2003) is usually defined as the total net income a company can expect from a customer. In other words, it is the

11 discounted value, or present value, of the projected net cash flows that a firm expects to receive from the customer over time (Berger and Bechwati, 2001).

It is argued by Stahl (2003) that accounting-based profitability measures do not adequately reflect the value of a firm the main reasons being that (1) accounting methods differ widely, (2) risks are not adequately taken into account, (3) investment requirements are ignored, (4) dividend policy is not reflected and (5) the time value of money is ignored. It is generally agreed that the market value of a firm comes from the net present value (NPV) of future cash flows generated by the firm’s assets, discounted at an appropriate interest rate and adjusted for inflation and risk (Stahl et al., 2003).

Therefore, customer profitability which is recognized as the most common measure of customer value in Iranian banking industry might not be the best metric on which to base decisions about customer strategies due to mentioned flaws.

1.2 Importance of CLV Measurement for Organizations

Customer lifetime value (CLV) is rapidly gaining acceptance as a metric to acquire, grow, and retain the “right” customers in customer relationship management (CRM) (Venkatesan and Kumar, 2004).

How do firms decide the timing of an offering to a customer? How do firms decide which prospect will make a better customer in the future and is therefore worthwhile to acquire now? Having got the customer to transact with the firm, what kind of sales and service resources should the firm allocate to conduct future business with that customer? How should firms monitor customer activity, in order to readjust the form and intensity of their marketing initiatives? Firms have used many techniques and methods to make these critical decisions. The notion of “customer value” is a common thread across these methods, even if it is not explicitly stated. What firms do is develop a measure of what they consider to be the best indicator of the total profits that a customer is likely to provide the firm and use that indicator to base their marketing decisions (Kumar et al., 2004).

12 The lifetime value of a customer is an important and useful concept in interactive marketing. (Courtheaux, 1986)) as cited by Pfeifer and Carraway (2000) illustrates its usefulness for a number of managerial problems—the most obvious if not the most important being the budgeting of marketing expenditures for customer acquisition. It can also be used to help allocate spending across media (mail vs. telephone vs. television), vehicles (list A vs. list B), and programs (free gift vs. special price), as well as to inform decisions with respect to retaining existing customers. (Jackson, 1996) as cited by (Pfeifer and Carraway, 2000) even argues that its use helps firms to achieve a strategic competitive advantage.

A basic belief of relationship marketing is that firms benefit more from maintaining long-term customer relationships than short term customer relationships (Reinartz and Kumar, 2000). The move toward a customer-centric approach to marketing, coupled with the increasing availability of customer transaction data, has led to an interest in both the notion and the calculation of customer lifetime value (CLV) (Fader et al., 2005).

1.3 Research Motivations

During the past decade, the retail banking industry started to face a set of radically new challenges that had an overall negative impact on industry margins and profitability. For the major part, these challenges have been caused by advances in modern information and telecommunication technologies, which ultimately have resulted in higher cost transparency and brand switching behavior. The resulting increase in competitive intensity has led to a commoditization of basic banking products, such as deposit taking, mortgages and credit extensions. This has further been fueled by an ever rising number of new entrants in the retail banking sector coming from industries as diverse as insurance and automobile production (Haenlein et al., 2007).

Iranian banks are not an exception to these challenges. Banking industry in Iran is getting more and more competitive by the growth of private banks both in number and size, so banks are urged to manipulate practices by which they could gain competitive

13 advantage over contestants. One way to gain this advantage is becoming more and more customer-centric rather than being product-centric, and value measurement would be a tool to achieve this goal (Pfeifer and Carraway, 2000, Aravindakshan et al., 2004, Jain and Singh, 2002). By value measurement banks could recognize relationships that are more profitable in long term for them and prerequisite of this practice would then be identification of risk factors in relationship. The financial ratios have been used for the past 6 years by Iranian banks for risk score estimation of customers, but since there is no accredited credit history available for customers in Iran and financial statements are unreliable, the error of such computations is on average 35% (Sabzevari et al., 2007) and makes these financial scores useless in decision making. Currently a firm with fake good standing financial statements could get a low risk score and would take advantage of being cross-sold or up-sold during its relationship with the bank and the bank might not be able to prevent losses on time. So there has recently been an urge from the management of some Iranian banks to have tools by which they could gain reliable risk rating method for their customers to accompany the existing financial scores. So this research basically extracts the most important qualitative factors that would affect the relationship borrowers’ risk in Iranian banking industry and would then compute qualitative risk score for them.

Another reason for choosing this subject is that value measurement has gained very little attention in Iran considering risk and shareholder value and there is an urge from managers, specifically in financial services industry, for doing such measurement. This way they could distinguish between a low value and high value customers and would prevent losses due to wrong selection of customers. The nature of the banking industry is risky by itself due to uncertainties about the ability of customers in being able to pay back their loans or not, so it is crucial for a financial service provider to recognize which relationships are hazardous and should be avoided. Thus a risk adjusted value measurement tool would be vital for banking management.

Very few academic researches have been specified to this subject (CLV measurement) in financial services industry which also brings into account the total risk

14 of customer as well as the shareholder value, and this fact would also add to the novelty of this project.

1.4 Research Objectives

There are several objectives involved in this research. According to mentioned importance of CLV measurement and motivations for conducting this research, the aims of this research would therefore be:

• Generation of a model that could best indicate the economic profitability of a lender’s business customer along with relationship benefits that s/he brings into the institute over her/his lifetime. • Our CLV model would bring into account the risk of relationship with the customer which has been mentioned as a pitfall of current value measurement models. This risk would include both the quantitative and qualitative risk factors. • Our risk adjusted CLV model would help the management of financial services to quantify their relationships. It could actually help managers of Iranian banks to estimate the relationship value they have with a customer considering the risk of relationship, and then decide how they can improve a customer’s position from high risk to low risk.

My research question according to motivations and objectives of this research would then be: “ What CLV model best fits the characteristics of Iranian financial services industry and their B2B borrowers?”

To answer our research question, we have designed a constructive research approach which will be overviewed in the following sections.

1.5 Research Design

As Haenlein (2007) mentioned, models used to value customers differ substantially in terms of loyalty and we should consider the customer migration model and the customer retention model. It therefore makes intuitive sense that models to determine CLV should, at least to a certain extent, be adapted to specific industry characteristics. Also the exact mathematical definition and its calculation method depend on many factors, such as whether customers are “subscribers” (as in most

15 telecommunications products) or “visitors”, as in direct marketing or e-business (Rosset et al., 2003).

Since our case is chosen to be a bank in which credit facilities (guarantees, loans, L/C’s) are recognized as the most profitable services in banking industry, we have considered a B2B lending relationship as the unit of measurement. This would make the industry setting a contractual one.

To calculate the lifetime value of a customer, three types of information are needed: the anticipated lifetime of the customer relationship in months or years; the profit in each future period adjusted for any customer-specific capital costs, such as marketing and customized services; and an agreed discount rate (Berger and Nasr, 1998).

In our research, the customer-specific cost and the lifetime will be estimated by the available information from our case.

The discount rate is preferably substituted by Weighted Average Cost of Capital (WACC) in order to consider the shareholder value and will be estimated by its calculation formula and the data available from our case. A company's cost of capital is the weighted average of its cost of debt and its cost of equity (or WACC) determined by the return its investors require on the money they invest in the business (Ryals and Knox, 2006). The capital used by a business comes either from debt or from equity funding; thus the overall cost of capital for a business is the weighted average of the cost of its equity and the cost of its debt. It is only if the return on capital exceeds its cost of capital that an investment creates shareholder value. Thus, using WACC as a basis for calculating the present value of future returns from customer relationships moves the CLV concept closer to direct shareholder value analysis.

The next step in our model development would be modeling the relationship risk which would be consisted of both qualitative and quantitative risks. The reason for not depending solely on financial score of customers is that in Iran the financial statement of customers are not reliable and so are the financial scores of them.

16 For the purpose of modeling the qualitative risk of customers, constructs that would affect the duration of lender-borrower relationship should be identified. This is done by an expert survey like the Delphi method which would ask the experts to solicit and grade attributes that in their opinion would bring higher risk in their lending relationship with a business customer. The model will consider the extracted experts’ factors to estimate the qualitative risk of each customer which would be added to the estimated quantitative risk of customers. This quantitative risk score is then estimated by statistical methods like logistic regression and this total risk score would be applied for estimation of revenue generation of each customer.

The first developed model will be tested on real data of data providers to be fine tuned. The developed model would certainly have flaws which could be modified after the case study is done. Then the new model would better be tested again through another case study for final justifications. So the flowchart of this research would then look like Figure 1.1.

The contribution of this research would then be a CLV model which considers the qualitative and quantitative aspects of risk of relationship for credit line clients of a bank. Since no specific model has been denoted and tested for this purpose, the contribution is both theoretical and empirical.

As shown in Figure 1.1, the research is divided into two phases: 1) model development phase and 2) fine-tuning phase. In the following sections, research methods for each phase of the research are discussed in details.

17 Problem Definition: Lack of risk considerations in CLV models

Model Development Phase

Relationship Risk Modeling 1-a) Modeling qualitative risk by Delphi method 1-b) Modeling quantitative risk by logit regression

Revenue Modeling

Initial risk adjusted CLV model

Model Fine- tuning Phase

Explanatory Case Study Model Modification

Barriers and Limitations

Fine-tuned Model

Further Theoretical Research Directions Empirical Contribution Contribution

Figure 1.1. Flowchart of the research

1.6 Model Development Phase

Since one of the objectives of our research is to bring the risk of relationship into account, the risk should be modeled first to be used in CLV calculations. There is an old

18 tradition of modeling financial risk of customers in financial services with different statistical methods. The roadmap to model development is to model qualitative risk of customers first, then merging it with quantitative risk to have an all-inclusive model for probability of default of customers. Then according to definitions of CLV in literature and characteristics of financial services industry, we will propose a mathematical model in chapter 3 which will be tested on our case, Karafarin Bank, in chapter 4.

1.6.1 Qualitative Risk Score Modeling

The qualitative factors of risk evaluation are rarely brought into calculation since they are complicated to be measured. In our research, since the emphasis is on bringing the risk into account, it is tried to solicit qualitative factors and use them in qualitative risk scoring.

The research method for estimation of qualitative risk score is chosen to be Delphi method. This method is an expert survey that has proven a popular tool in information systems research for identifying and prioritizing (ranking) complex issues for managerial decision-making. Surveys are described as a system for collecting information to describe, compare, or explain knowledge, attitudes and behavior. Though many surveys go further than this, looking for associations between social, economic and psychological variables and behavior (Gray (2004) cited by Saunders et al. (2000)).

The objective of Delphi method is to obtain the most reliable consensus of a group of experts (Okoli and Pawlowski, 2004).

The question here might be why we have chosen Delphi method over traditional survey method. The main differences between Delphi method and traditional survey method are summarized in table 1.1. In light of this comparison, we select the Delphi method for the following reasons:

1. This study is an investigation of factors that would affect the risk of relationship with a business customer in financial services industry. This complex issue requires knowledge from people who understand the different economic, social, and

19 political issues there. Thus, a Delphi study answers the study questions more appropriately.

2. A panel study most appropriately answers the question of “what qualitative factors affect risk of lending for business borrowers in Iran”, rather than any individual expert’s responses. Delphi is an appropriate group method. Among other high-performing group decision analysis methods (such as nominal group technique and social judgment analysis), Delphi is desirable in that it does not require the experts to meet physically, which could be impractical for nationwide experts.

3. Although there may be a relatively limited number of experts with knowledge about our risk factors question, the Delphi panel size requirements are modest and we could hopefully manage to have different panels of 10-18 experts.

4. The Delphi study is flexible in its design, and it is possible to follow-up interviews. This permits the collection of richer data leading to a deeper understanding of the fundamental question for our risk modeling.

Table 1.1. Comparison of Delphi method and traditional survey adapted from Okoli & Pawlowski (2004) Evaluation Traditional Survey Delphi Method

Criteria The questions that a Delphi study investigates are those of high uncertainty and speculation. Thus, a general Using statistical sampling population, or even a narrow techniques, the researchers subset of a general population, Representativene randomly select a sample that is might not be sufficiently ss of samples representative of the population of knowledgeable to answer the interest. questions accurately. A Delphi study could be considered a type of virtual meeting or a group decision technique, though it appears to be a complicated survey. Because the goal is to generalize The Delphi group size does not

20 Evaluation Traditional Survey Delphi Method

Criteria results to a larger population, the depend on statistical power, Sample size for researchers need to select a sample but rather on group dynamics statistical power size that is large enough to detect for arriving at consensus and significant statistically significant effects in the among experts. Thus, the findings population. Power analysis is literature recommends 10–18 required to determine an experts on a Delphi panel. appropriate sample size. Studies have consistently shown that for questions The researchers average out requiring expert judgment, the Individual vs. individuals’ responses to determine average of individual responses group response the average response for the is inferior to the averages sample, which they generalize to produced by group decision the relevant population. processes; research has explicitly shown that the Delphi method bears this out. Pretesting is also an important An important criterion for reliability assurance for the Reliability and evaluating surveys is the reliability Delphi method. However, test- response revision of the measures. Researchers retest reliability is not relevant, typically assure this by pretesting- since researchers expect retesting to assure reliability. respondents to revise their responses. In addition to what is required of a survey, the Delphi method can employ further construct validation by asking experts to Construct validity is assured by validate the researcher’s Construct careful survey design and by interpretation and

21 Evaluation Traditional Survey Delphi Method

Criteria validity pretesting. categorization of the variables. The fact that Delphi is not anonymous (to the researcher) permits this validation step. Respondents are always anonymous to each other, but Respondents are almost always never anonymous to the Anonymity anonymous to each other, and often researcher. This gives the anonymous to the researcher. researchers more opportunity to follow up for clarifications and further qualitative data. Researchers need to investigate the Non-response is typically very Non-response possibility of non-response bias to low since researchers have issues ensure that the sample remains personally obtained assurances representative of the population. of participation. For single surveys, attrition Similar to non-response, (participant drop-out) is a non- attrition tends to be low in Attrition effects issue. For multi-step repeated Delphi studies, and the survey studies, researchers should researchers usually can easily investigate attrition to assure that ascertain the cause by talking it is random and non-systematic. with the dropouts. The richness of data depends on the In addition to the richness form and depth of the questions, issues of traditional surveys, and on the possibility of follow-up, Delphi studies inherently Richness of data such as interviews. Follow-up is provide richer data because of often limited when the researchers their multiple iterations and are unable to track respondents. their response revision due to feedback.

22 There are two major applications of Delphi method in literature, one being forecasting and issue identification/prioritization, the other one concept/ framework development. The former represents one type of application of the method. The majority of the Delphi efforts during the first decade were for pure forecasting, including both short- and long-range forecasts. While most forecasting studies use Delphi to surface a consensus opinion, others emphasize differences of opinion in order to develop a set of alternative future scenarios. In both cases, the design involves a two-step process beginning with identification/elaboration of a set of concepts followed by classification/taxonomy development (Okoli and Pawlowski, 2004).

In our research, we also use Delphi method for its first type of application. Although we are not using Delphi for forecasting, but we actually need the consensus of experts for issue identification/prioritization (identification of most important risk factors in lending relationship in Iran).

1.6.1.1 Delphi Method Steps

After the researchers design the questionnaire, they select an appropriate group of experts who are qualified to answer the questions. This phase is called the “expert selection” phase which is very crucial in Delphi method.

Next phase is “data collection and analysis” phase. In this phase the researchers administer the survey and analyze the responses. Next, they design another survey based on the responses to the first one, asking respondents to revise their original responses and/or answer other questions based on group feedback from the first survey. The researchers reiterate this process until the respondents reach a satisfactory degree of consensus. The respondents are kept anonymous to each other (though not to the researcher) throughout the process. The flowchart of Delphi method is shown in figure 1.2.

Details of this phase are summarized bellow from Okoli and Pawlowski (2004) research: 1. Brainstorming: For this phase only, treat experts as individuals, not panels • Questionnaire 1: Ask experts to list relevant factors (not in any order)

23 Selection of the Delphi methodology

Selecting experts

Questionnaire 1: initial collection of Data collection and analysis method factors Brainstorming for important factors Questionnaire 2: validation of categorized list of factors Narrowing down factors Questionnaire 3: choosing most important factors Ranking relevant factors Questionnaire 4: ranking the chosen factors

No 1) Kendall ’ s W > 0.7 2) Iteration = 3 3) MeanRank l = MeanRank l+1

Yes

Practical Recommendation

Figure 1.2. Flowchart of Delphi Method adapted from Okoli and Pawlowski (2004)

• Consolidate these two lists from all experts, regardless of panel • Remove exact duplicates, and unify terminology • Questionnaire 2: Send consolidated lists to experts for validation • Refine final version of consolidated lists 2. Narrowing down: Henceforth treat experts as distinct panels • Questionnaire 3: Send lists to each expert • Each expert selects (not ranks) at least ten factors on each list • For each distinct panel, retain factors selected by over 50% of experts 3. Ranking

24 • Questionnaire 4: Ask experts to rank factors on each of their panel’s pared-down lists • Calculate mean rank for each item • Assess consensus for each list within each panel using Kendall’s W • Share feedback with each panelist and ask them to re-rank each list • Reiterate until panelists reach consensus or consensus plateaus

1.6.2 Quantitative Risk Score Modeling

In our research, as shown in Figure 1.1, we also intend to estimate the quantitative risk score of customers. This score should be estimated since the qualitative factors are not sole representatives of the risk of a customer. We will then bring an amalgamation of the two risk scores (qualitative and quantitative) as a total risk score for PD calculations to be embedded in our final CLV model for revenue generation considerations.

1.7 Model Fine-tuning Phase

In order to find out the limitations and barriers of the developed model, we would need to fine-tune the initial developed model. For this purpose, we will conduct a case study and in addition to stated purposes, will find out the reasons behind the error terms.

1.7.1 Case Study

Case study is defined as the development of detailed, intensive knowledge about a single case, or a small number of related cases (Robson, 1993). This strategy will be appropriate if the researcher wish to gain a rich understanding of the context of the research and the processes being enacted (Morris & Wood, 1991).

The case study approach also has considerable ability to generate answers to the question “why” as well as “what?” and “how?” questions (Robson, 1993), although “what” and “how” questions tend to be more the concern of the survey method (Saunders, Lewis, & Thornhill, 2000). The data collection methods employed in this kind of research may include questionnaires, interviews, observations and documentary analysis

25 (Saunders, Lewis, & Thornhill, 2000), but the main reliance is on qualitative data (Cooper & Schindler, 2003).

In spite of the unscientific feel this kind of research has, a simple well-constructed case study can enable the researcher to challenge an existing theory and also provide a source of new hypotheses (Cooper & Schindler, 2003; Saunders, Lewis, & Thornhill, 2000) and constructs simultaneously and have a significant scientific role (Cooper & Schindler, 2003).

A case study selects a small geographical area or a very limited number of individuals as the subject of study. Case studies, in their true essence, explore and investigate contemporary real life phenomena through detailed contextual analysis of a limited number of events or conditions, and their relationships. This method observes the data at the micro level. It allows the exploration and understanding of complex issues. It can be considered a robust research method particularly when a holistic, in-depth investigation is required. Through case study methods, a researcher is able to go beyond the quantitative statistical results and understand the behavioral conditions through the actor’s perspective (Burn and Ash, 2005).

According to Saunders et al. (2000) research purpose is classified in three folds as: Exploratory, Descriptive and Explanatory.

Exploratory studies are valuable means of finding out “what is happening; to seek new insights; to ask questions and to assess phenomena in a new light”. It is a particularly useful approach if the researcher wishes to clarify his/her understanding of the problem. There are three principal ways of conducting exploratory research:

• A search of the literature.

• Talking to experts in the subject.

• Conducting focus group interviews.

Exploratory research could be likened to the activity of the traveler or explorer. Its great advantage is that it is flexible and adaptable to change. That is, the researcher can

26 change the direction as the result of new data which appears and new insights which occur to him (Saunders, Lewis, & Thornhill, 2000). Adams and Schvaneveldt (1991) reinforce this point by arguing that the flexibility inherent in exploratory research those not mean absence of direction to the enquiry. What it does mean is that the focus is initially broad and becomes progressively narrower as the research progresses.

The object of descriptive research is “to portray an accurate profile of persons, events or situations” (Robson, 1993). This may be an extension of, or forerunner to, a piece of exploratory research. It is necessary to have a clear picture of phenomena on which you wish to collect data prior to the collection of the data. Although description in management and business research has a very clear place, scholars tend to go further and to draw conclusions from data. They use to develop the skills of evaluating and synthesizing data and ideas which are higher order skills than those of accurate description. Descriptive methods should be thought of as a means to an end rather than an end in itself (Saunders, Lewis, & Thornhill, 2000).

Studies which establish causal relationships between variables may be termed explanatory studies. The emphasis here is on studying a situation or problem in order to explain the relationships between variables (Saunders, Lewis, & Thornhill, 2000). Such studies usually apply quantitative analyses in order to test the relationship between variables and explain them. Explanatory case study is set to examine the data closely both at a surface and deep level in order to explain the phenomena in the data, so answers the “why” research question. It is used for theory building.

For the purpose of fine-tuning our developed model, we have chosen to run an explanatory case study because we have a model in hand which should be refined by finding out the pitfalls it faces during empirical testing. Case study is suitable since developing a model that considers many factors such as customer behavior, risk of relationship and industry setting is a very complex issue and this method helps understand why some constructs in the developed model should be more profoundly highlighted or omitted.

27 Since the lending relationship of Karafarin Bank (KB) has been the subject of the case study in our research, we review this bank’s background, its lending technique, and current losses this bank faces due to mal-functioning of current technique. Following sections would also clarify our incentive for having KB as our case study.

1.7.1.1 KB’s Background

Karafarin Bank was initially established as Karafarin Credit Institute in 1999. The Institution was officially converted into a bank on January 1, 2001 as the first privately- owned bank in operation (www.karafarinbank.com January, 2008).

With new opportunities in the business environment available in Iran and an expansion of the Iranian market beyond the oil and gas industries, Karafarin is going to formulate a new strategy to meet the growing demands of both the retail customers and the business community. Their mission is to offer specialized financial services to develop business units that offer strong potential in terms of growth and profitability. The capital of Karafarin now amounts to 2 trillion Rials right now.

1.7.1.2 Credit Issuance Process at KB

When a loan application is filed by a firm at KB, the credit committee of the bank will decide on the amount of loan that could be granted to the firm. This committee, which differs in number and expertise of members from one bank to another, includes four members at KB which decides on the basis of financial and non-financial criteria and credit policies that the board of directors sets for the bank. These criteria include profitability, solvency, and efficiency of the firm through financial statements and also the business environment and the status of the firm’s industry. These criteria are checked either by objective factors or by subjective knowledge of the committee members about the firm and its industry. The problem with this process is its inefficiency due to lack of precision of self judgments based on unstructured financial and non-financial factors which also makes this process a prolonged practice.

28 According to Baas and Schrooten (2006), there are four types of lending in financial services in which the first is based on soft information and the other three based on hard information (Baas and Schrooten, 2006). These techniques are summarized in Table 1.2.

Table 1.2. Lending techniques adapted from Baas and Schrooten (2006) Relationship lending is based on the experience of a given bank with a specific borrower and therefore on soft information collected over time. So if financial data is limited, relationship banking is the technique of choice

Financial statement lending is based on evaluating information from the firm’s financial statements. The decision to lend depends largely on the strength of the balance sheet and income statements.

Asset-based lending is principally based on the quality of the available collateral. This type of lending causes high monitoring costs and requires high-quality receivables and inventory available to pledge (Berger and Udell, 1995, 1998; Boot, 2000). That is why it is generally used as a substitute for relationship lending if the term of the relationship is short.

Small business credit scoring is an adaptation of statistical techniques used in consumer lending. In addition to information about the financial statements, the creditworthiness and history of the owner is heavily weighted (Frame et al., 2001).

In Iranian financial services including KB, traditional asset-based lending is applied in which hard assets, such as real property, equipment, and inventories are pledged. In such an application, the bank’s experts determine the value of the borrowing firm’s assets, and if the total value of these assets is higher than the credit amount, the bank lends the money. If a firm fails to repay its debts, which is very common in Iran, the bank takes over pledged assets through a lengthy and a bureaucratic process and tries to sell those assets to the highest bidder in an auction. The lengthy process of collateral evaluation besides the lengthy process of collateral liquidation through judicial system (in case of default of customers), makes the lending process unfavorable both to borrowers

29 and lenders in Iran and has caused some banks like KB to think of new banking concepts like relationship lending.

1.7.2 Data Collection Method for the Case Study

Research strategies can apply different research tactics or in the other terms, data collection methods. Appropriate data collection method will be chosen according to the research strategy and research questions. According to Saunders et al. (2000) there are four data collection methods, including: secondary data, observation, interviews and questionnaires.

In the following sections, the secondary data as the data collection method of our case study will be discussed.

1.7.2.1 Secondary Data

When considering how to answer the research questions or to meet their objectives, few researchers consider the possibility of re-analyzing data that have already been collected for some other purpose (Hakim, 1982). Such data are known as secondary data. Although most researchers think in terms of collecting new (primary) data, secondary data can provide a useful source for research questions to be answered or at least they can be considered as a beginning to answer research questions. Most research questions are answered using some combination of secondary and primary data. Where limited appropriate secondary data are available, the researcher will have to rely mainly on data he/she collects himself/herself (Saunders et al., 2000).

1.7.2.2 Types of secondary data

Secondary data includes both quantitative and qualitative data and can be used in both descriptive and explanatory research. Saunders et al. (2000) have built three main sub groups of secondary data: Documentary data, survey-based data and those compiled from multiple sources.

30 Documentary data are often used in research projects that also use primary data collection methods. Documentary data includes written documents and non-written documents. Written documents include notices, correspondents, reports to share-holders, transcripts of speeches, administrative and public records, books, journals and magazine articles, and newspapers. These documents can provide qualitative data such as managers’ reasons for decisions or they can be used to generate statistical measures of profitability derived from company as instance (Bryman, (1989) cited by Saunders et al. (2000)). Non-written documents include tapes, video recordings, pictures, drawings, films and television programs (Robson, (1993) cited by Saunders et al. (2000)).

Documentary data can be analyzed both quantitatively and qualitatively. In addition, they can be used to help to triangulate findings based on other data such as written documents and primary data collected through observations, interviews or questionnaires (Saunders et al., 2000).

Survey-based data refers usually to data collected by questionnaires which have already been analyzed for their original purpose. Such data can refer to organizations, people or households (Hakim, 1982). These data have been collected through one of three distinct types of survey: Censuses, Continuous/Regular surveys or Ad hoc surveys (Saunders et al., 2000).

Multiple source data can be based entirely on documentary or on survey data or can be a mixture of the two. The key factor is that different data sets have been combined to form another data set prior to the researcher assessing the data. This category includes: Area-based and Time-series based data (Saunders et al., 2000). While time-series based data are combination of some secondary data related to the same sample over time, area- based data are combination of different sources that have the same geographical base (Hakim, (1982) cited by Saunders et al. (2000)).

In our case study, we have used secondary sources of data in the form of written documentary materials. These documents include the financial data of KB’s customers.

31 1.7.2.3 Evaluating secondary data

Secondary data must be viewed with the same caution as any primary data that researcher collects (Saunders et al., 2000). Most authors suggest a range of validity and reliability against which the researcher can evaluate potential secondary data. Saunders et al. (2000), based on previous works, have suggested the following criteria to evaluate secondary data.

• Measurement validity: One of the most important criteria for the suitability of secondary data sets is measurement validity. Secondary data that fail to provide the information that researcher needs to answer his/her questions or to meet his objectives will result in invalid answers. This concept can be applied both for quantitative and qualitative data. For quantitative data, as an example, a manufacturing organization may record monthly sales whereas the researcher is interested in monthly orders. In another example, for qualitative data, a researcher may be using minutes of company meetings as a proxy for what actually happened in those meetings. These are likely to reflect a particular interpretation of what happened, the events being recorded from a specific view point, often the chair person’s. The researcher therefore needs to be cautious before accepting such records at face value (Saunders et al., 2000)

• Coverage and unmeasured variables: The researcher needs to be sure that the secondary data cover the population about which you need data, for the time period he/she needs, and contains data variables that would enable him/her to answer his/her research questions.

• Reliability and validity: The reliability and validity of the secondary data is a function of the method by which the data were collected and the source. If the source of secondary data is well-known and reliable the data is likely to be trustworthy (Saunders et al., 2000). In addition, relying on their procedures for their collecting and compiling the data is eligible.

Validity and reliability of collection methods for survey data will be easier to assess where the researcher can find a clear explanation of the methodology used for the data collection. This needs to include the clear explanation of any sampling techniques used and response rates as well as a copy of the survey instrument (questionnaire). Secondary data collected through a survey with a high response rate are also likely to be more reliable than from that with a low response rate (Saunders et al., 2000).

32 1.8 Expected Contribution of the Research

The theoretical contribution of this research is a mathematical CLV model that expressively considers qualitative and quantitative risk score of business customers of the bank. This model estimates the value of loan applicants who are the most profitable customers of the bank. Such a specific model has not been developed before which would make this aspect of our contribution a novel one. We will also test this developed model on a case, KB, and will have an empirical contribution.

Currently at KB, the management is concerned about losses due to default of borrowers. Table 1.3 shows the percentage of customers who have had defaults of more than 3 months.

Table 1.3. Percentage of customers who have had defaults at KB ò   /   b 5  5                              

It is too complicated to estimate the experienced loss due to default of customers in the bank, but we could estimate the percentage of non-performing loans. The results are shown in Table 1.4.

Table 1.4. Total non-performing yearly loan to total yearly loan ratios

Mar-06 Mar-07 Mar-08 Total non-performing net 839,774,981,498.00 1,176,894,721,476.00 1,191,804,549,390 Total loans 7,310,063,411,300.00 11,312,508,780,619.00 17,354,093,919,352 Percentage 11.49% 10.40% 6.87%

Although the percentage has been decreasing for the past three years, the management of this bank seeks tools by which they could get early signals for risky customers who would possibly cause this percentage.

33 Our contribution is to equip KB with such a tool. This tool, a risk adjusted CLV model for loan applicants, would not only help KB to become aware of risky customers but also to be able to recognize value creating customers and to try to lend more to them. This means this tool would empirically in long term increase the denominator of the above ratio and at the same time decrease the numerator.

1.9 Abbreviations and expressions

Expressions and abbreviations used in this report are summarized in this section.

CE: Customer Equity is defined as the total of the discounted lifetime values summed over all of firm ’s current and potential customers ((Rust et al., 2004) cited by Kumar and Goerge, 2007).

CLV: Customer Lifetime Value same as CLTV (Customer Lifetime Value) and LTV (Lifetime Value) is the present value of a customer’s future purchases (Ryals, 2002).

CRM: Customer Relationship Management is ‘Managerial efforts to manage business interactions with customers by combining business processes and technologies that seek to understand a company’s customers’ (Hwang et al., 2004).

NPV: Net Present Value is the sum of , where is the profit contribution of customer i at period ti and is the interest rate factor (Hwang et al., 2004).

WACC: Weighted Average Cost of Capital is defined as (cost of debt×proportion of debt) plus (cost of equity×proportion of equity) (Ryals and Knox, 2006).

1.10 Structure of the Report

The rest of this report is organized as follows. Next chapter briefly reviews the literature on customer lifetime value, its applications, and ends by covering some major

34 CLV models, assumptions and characteristics of each model and the industry each has been tested on (if applicable). Chapter three then reviews the steps of undertaken research methods in details for the first phase of the research which is the model development. Chapter four brings up the data and calculations for the case study phase. This project report is concluded in chapter five by conclusion, limitations of the research, and future research directions.

1.11 Summary of the Chapter

This chapter basically defined the problem as that profitability measurement not being an appropriate measure on its own for decision making and appraises the new metric of CLV by which it is tried to consider the relationship value with customers over their lifetime. The problem with current CLV models is that they do not consider the risk of relationship in their models except in a few cases by which it is tried to bring in the risk into consideration to some extent.

My personal motivations of choosing this subject were brought up which were basically the novelty of this subject in Iran and the growing competitiveness among Iranian banks which urges them to seek ways of gaining competitive advantage. Our goal would then be developing a model which could help managers of banks estimate the value of their customers considering the risk of relationship and then decide how to better manage their relationships with them.

The research approach was also discussed in details in this chapter. The general strategy of this research was an expert survey for extraction of qualitative risk factors which would ultimately be used for PD modeling of customers to be used in CLV model that will be generated specifically for banking industry, in addition to the case study strategy which would need the secondary data collected from our case, KB. Next chapter will define CLV concept and its applications in details.

35

Chapter 2

Literature Review on Customer Lifetime

Value Applications and Models

2. Customer Lifetime Value

Preface: The old concept of lifetime value, LTV, has been revived by CRM. It usually refers to the net value of an individual consumer’s purchases over his or her lifetime, sometimes widened to the whole family, even to both private and professional consumption (Bayon and Gustsch, 2002). This chapter defines CLV concept and basically reviews pros and cons of CLV models in general and brings up its applications.

36 2.1 Value Measurement Background

The interest the Marketing discipline has recently been paying to CLV and the related subject of customer relationship management has its roots in an evolution that started in the mid 1980’s (Haenlein et al., 2007). Researchers highlight that Marketing, which has historically focused on the analysis of single transactions, should start paying attention to the relationship aspect of buyer–seller behavior. In this section we will review the literature on the history of value measurement concept.

2.1.1 Customer Relationship Management

Since the early 1980s, the concept of relationship management in marketing area has gained its importance (Ryals, 2002). Customer relationship management (CRM) is part of marketing’s new dominant logic. Customer relationship management (CRM) has become one of the leading business strategies in the new millennium. It is difficult to find out a totally approved definition of CRM. (Hwang et al., 2004), however, describe it as ‘Managerial efforts to manage business interactions with customers by combining business processes and technologies that seek to understand a company’s customers’ i.e. structuring and managing the relationships with customers. CRM covers all the processes related to customer acquisition, customer cultivation, and customer retention. Even though we put aside the existing studies, which assert that it costs more to acquire new customers than to retain the existing customers, we can imagine that customer cultivation and retention are more important than customer acquisition because lack of information on new customers makes it difficult to select target customers and this will cause inefficient marketing efforts. Acquiring and retaining the most profitable customers are serious concerns of a company to perform more targeted marketing campaigns. For effective customer relationship management, it is important to gather information on customer value (Hwang et al., 2004).

In order to manage relationships as assets, companies need to know which are their most valuable and which are their least valuable relationship assets so that appropriate marketing strategies can be put in place. The most valuable customer assets

37 have to receive priority and be defended from poaching by the competition. Less valuable customer relationships have to be scrutinized to see how returns can be improved. For this reason, interest in the concept of customer profitability is very strong (Ryals, 2002).

Retained customers are known to be more profitable because they tend to buy more, are relatively cheaper to serve and may be less price-sensitive. However, developing relationships with customers may in addition give retailers access to relationship benefits. These relationship benefits may, in some cases, create greater value for a retailer than is obtained from the stream of customer profits; as cited by Ryals (2002), (Kalwani and Narayandas, 1995) find that long-term customer relationships do pay off for suppliers, partly through increased profitability but largely through relationship benefits. The existence of significant relationship benefits may even explain why companies continue to deal with unprofitable customers. If the relationship benefits are great enough, and are managed well enough, the retailer can create value for shareholders by dealing with certain customers, even if these customers are not profitable in an accounting or shareholder value creating sense. Focusing on the economic value of the customer will enable organizations to manage both the returns and the risks in their customer relationships in a far more sophisticated way than looking only at profit. However, there is one more aspect of customer relationship management that retailers can use. This is the notion that customer relationships have a worth to the organization that is not fully captured through calculations of profit or even of value. This notion, which has its roots in relationship marketing, suggests that relationships carry additional relationship benefits. If so, the value a company obtains from its customer relationships does not just come from the stream of returns it makes from that customer over the relationship lifetime. By developing successful relationships with customers, retailers may also access other sources of value. These ‘relationship benefits’ can have a powerful impact not just on the value of a single customer but also on the value of the business as a whole. Some relationship benefits affect process efficiency. Some customers, for example, help to evaluate new product concepts. Other relationship benefits from their existing customers make it easier for organizations to acquire new customers; advocate customers may refer

38 non-customers to their favorite suppliers or big name customer firms may agree to give an endorsement to their suppliers (Ryals, 2002).

2.1.2 Relationship Marketing

It is the age of relationship marketing, an age in which making a sale is just the beginning, rather than the end, of a company-customer relationship. At the core of relationship marketing is the development and maintenance of long-term relationships with customers, rather than simply a series of discrete transactions. One consequence of relationship marketing is, therefore, a major directional change in the criterion variable that should guide managerial decisions (Jain and Singh, 2002).

Relationship marketing has often been contrasted to transaction marketing which is about developing, selling and delivering products by means of short-term, discrete economic transactions. Because the lifetime value of the customer is not taken into account, customer attraction but not customer retention is at the heart of transaction marketing exchanges (Lindgreen and Crawford, 1999). However, it is now proposed that closer attention is paid to the long-term financial benefits, and other benefits, of retained customers the main reason being that competition in the marketplace has intensified. To achieve growth, it is argued, organizations must change their paradigm to that of relationship marketing.

2.2 Historic Data Analysis Methods

There are two broad approaches to the calculation of customer profitability—top down, and bottom up. The top-down approach begins with total profits, takes the customer base as a whole and tries to determine the profitability of customer segments. The bottom-up approach aims to identify the profitability of individual customers.

2.2.1 Top-down Approach

Arriving at an accurate picture of customer profitability via a top-down approach is problematic, particularly where product costs are a relatively low proportion of total

39 costs. The higher the proportion of indirect costs, the more misleading a simple proportional allocation of such costs could be. Different customers use a company’s resources very differently; for example, inventory holding and delivery requirements, payment terms, order entry, customer and sales support, may all vary considerably from customer to customer. Allocating such costs proportionate to volume, as is often done, may well fail to reflect the true pattern of the customer’s usage of the retailer’s resources. Where indirect costs are significant, an incorrect allocation can lead to a seriously misleading picture of customer profitability. Some customers are just more costly to serve than others, often as a result of their behavior. Aggressive customers can be more costly to serve and may demand lower prices than passive customers; they may demand special packaging, delivery and service as well as being tough negotiators on price. The balance of power between customer and supplier is likely to affect the profitability of the relationship; the stronger the customer, the more concessions they can wring from their suppliers and the less profitable that relationship may be (Ryals, 2002).

A more accurate picture of customer profitability can be obtained using IT and applying new analysis techniques. Customer data can be collected in data warehouses and then analyzed. Newer financial tools, particularly activity-based costing (ABC), can help retailers understand true customer profitability. Where unallocated overheads are substantial and customers are very different in their purchasing behaviour (particularly where the demands that customers make are not proportional to the quantities they buy), ABC is appropriate. Activity-based costing enables organizations to apportion more accurately the real costs of serving individual customers or customer segments.

It does this by examining the actual time spent on specific activities or processes supporting individual customers or a customer segment. The costs per hour of the activity can be worked out and then multiplied by the actual time spent per customer or segment. Because these costs are activity based rather than allocated, ABC enables managers to understand which customers or segments are more demanding of marketing or support time and hence more costly to look after, and which customers or segments are less demanding. For a given amount of sales revenue, a customer or segment that takes up more marketing or support time is not as profitable as a customer or segment that uses

40 less of the retailer’s marketing or support resources. This important point is obscured by cost allocation but revealed through ABC. Because ABC can demonstrate not just the total costs by customer or segment, but also which specific activities or processes by customer or segment are the most costly, managers can look at ways to reduce costs, streamline processes or reallocate resources. This gives a better basis for managerial action than profitability analysis using cost allocation because it enables managers to focus their attention and energy on improving activities that will have the biggest impact on the bottom line (Ryals, 2002). For retailers working with a top-down approach to customer profitability analysis, ABC will give a more accurate picture of customer profitability. No cost allocation system is perfect, however, and ABC is no exception. The rules of information search still apply to customer cost information derived via ABC; the benefit of having the information must exceed the cost of acquiring it. This means that, for certain less significant cost elements, old-fashioned apportionment might be the most cost-effective strategy. Sometimes, new software tools can reduce the cost of collecting cost information. Such software can collect and allocate even relatively small cost items to customers both cheaply and accurately. For example, software products are now available which monitor the length and cost of telephone calls by telephone number. This can then be used as the basis for an extremely accurate allocation of call centre or help desk costs to a customer profitability account (Ryals, 2002).

Taking a top-down approach to customer profitability measurement can give some indication of the value of the customer to the retailer but, because of the difficulties mentioned above, may have limited usefulness.

2.2.2 Bottom-up Approach

Where retailers wish to determine the value of their customers more precisely, a bottom-up approach is usually preferred. There are a number of ways of doing this approach. One is to use historic transactional data. Retailers are often interested in three dimensions of historic customer purchasing behaviour: recency, frequency and amount purchased (RFA). The results may be used to identify key events in the customer’s purchase cycle at which he or she is more likely to buy from the retailer, enabling tailored

41 marketing campaigns to be put in place. A wine-buying customer, for example, might receive a special offer on a case of wine just before Christmas (Ryals, 2002).

The concept of customer loyalty has always been at the forefront of retailers’ quest to retain customers. To identify loyal customers, retailers have typically analyzed customer behavior with respect to the following:

(a) For how long has the customer been active?

(b) How regularly does the customer buy?

(c) What is the RFM 1 score of my customer? (Kumar et al., 2006)

Liu and Shih (2005b) believe that measuring RFM is an important method for assessing customer lifetime value. Bult and Wansbeek (1995) as cited by Liu and Shih (2005) defined the terms as: (1) R (Recency): period since the last purchase; a lower value corresponds to a higher probability of the customer’s making a repeat purchase; (2) F (Frequency): number of purchases made within a certain period; higher frequency indicates greater loyalty; (3) M (Monetary): the money spent during a certain period; a higher value indicates that the company should focus more on that customer (Liu and Shih, 2005b).

2.2.3 Limitations of Historic Data Analysis Methods

Based on these metrics, retailers often take future investment decisions for customer relationship management. For example, if a customer is deemed to be a long- life customer who transacted frequently with the firm in the past, then the retailer may allocate greater resources for that customer with the underlying expectation of increased profitability from that customer in the future. Intuitive as it may sound, a serious problem with this reasoning is that a backward-looking metric is employed to take a forward- looking decision. Further, some researchers in the past have cautioned against a weak relationship between behavioral loyalty (such as measures (a)–(c) above) and profitability. Hence, it is worth exploring whether the traditional loyalty metrics should

1 RFM score for a customer is derived from the simple or weighted combination of the customer’s recency, frequency and the monetary value of purchase.

42 be treated as reliable indicators for furthering the relationship with a customer in future (Kumar et al., 2006).

The major disadvantage of the RFM analysis as Ryals (2002) mentions is that it focuses on revenue rather than cost and therefore does not capture the real profitability of a customer relationship. It is possible for a retail company to have customers with identical RFM profiles but whose demands are very different. When buying a holiday, some customers spend a lot of time with the travel representative whilst others decide quickly. The latter are efficient purchasers (from the point of view of the travel agent) and the former are inefficient and more costly to serve. This is not reflected in the RFM data. RFM analysis also suffers from the key drawback of all historic data; it may not be reliable as a guide to the future. This is true whether the attempt to value the customer begins with the whole of the customer base and works down, or begins from the individual customer and works up. The approaches to customer profitability analysis that have been discussed so far are historic, where retailers are working with data about what customers used to do in the past. Of far more interest for the development of marketing strategies is what customers will do in the future. The problem with using historic data is so significant that it repays some additional consideration.

A fundamental issue which applies to all historic data analyses is that current and historic transaction and profitability data is not necessarily reliable as a guide to the future. Changes in a customer’s life circumstances or even changes in preferences can alter purchasing behavior from one period to the next. Retail managers have to consider whether single period customer profitability calculations, however accurate, can and should be used as a guide to managerial action. There is a danger that, in focusing on the fine detail of the profitability of the customer day-to-day, managers may lose sight of the big picture—the value of the relationship as a whole. The profits from a customer may vary significantly from one period to the next, and decisions made on the basis of the profitability of the customer in 1 year might look rather unfortunate in the next. Basing marketing strategies on historic or current customer profitability could be damaging to longer-term value creation. Marketing strategies about which customers to retain and which to invest in require a different perspective. Retailers need to examine the future

43 potential of their customer relationships. As companies move towards one-to-one marketing, they need a longer-term view of the value of their customer relationships than current data can provide. Effectively, they need to predict the future purchasing behavior of their customers to arrive at customer lifetime value (Ryals, 2002).

2.3 CLV Definition

As mentioned in previous sections, relationship marketing constitutes a major shift in marketing theory and practice. Rather than focusing on discrete transactions, it emphasizes the establishment, development and maintenance of long-term exchanges which are thought to be more profitable than short-term relationships as a result of exchange efficiencies. This is especially true of customer relationships. It is argued that customer relationships are viewed as investment decisions and customers as generators of revenue streams, but customer relationships also generate costs. Hence, in order to measure the customer lifetime value, all revenues and costs pertaining to a customer relationship must be assessed (Stahl et al., 2003).

The customer lifetime value is the present value of a customer’s future purchases. Usually, the CLTV calculation is based on the expected purchases of a single customer and adjusted back to the present day using a discount rate. Knowing the present value of the future relationship is helpful in developing marketing strategies. The CLTV analysis demonstrates that the value of the relationship with this customer can be increased either by increasing the amount of profit (by increasing the revenue from the customer and/or decreasing costs to serve) or by extending the relationship lifetime (Ryals, 2002). Different definitions of CLV are brought in table 2.1.

In fact, companies can trade off the costs of increasing service levels to this customer against the likelihood of gains from extending the relationship. It is not just short-term fluctuations in customer profitability that affect the marketing strategy that a supplier adopts; the stage of the customer lifecycle can also be crucial. The lifetime value of consumers at the beginning of their relationship lifetime is greater than that of a customer at the declining stage of their relationship lifetime.

44

Table 2.1 Definitions of Lifetime Value Definition of LTV Article The present value of all future profits generated from a customer Gupta and Lehmann (2003) The net profit or loss to the firm from a customer over the entire Berger and Nasr life of transactions of that customer with the firm (1998) Expected profits from customers exclusive of costs related to Balttberg and customer management Deighton (1996) The total discounted net profit that a customer generates during Bitran and her life on the house list Mondschein (1996) The net present value of the stream of contributions to profit that Pearson (1996) result from customer transactions and contacts with the company The net present value of a future stream of contributions to Jackson (1994) overheads and profit expected from the customer The net present value of all future contributions to overhead and Roberts and Berger profit (1989) The net present value of all future contributions to profit and Courtheoux (1995) overhead expected from the customer

(Hwang et al., 2004)

Retail banks understand this principle well. They have identified students as a potentially high value segment over a lifetime, even though in the short term they may be unprofitable. The banks compete to offer special deals to these customers in the hope that they can attract and retain customers who will deliver profits in future periods. Customer lifetime value, the present value of a stream of future profits from that customer, is a better guide for customer strategy than current period profit (Ryals, 2002). The reliability of a CLTV calculation depends on the ability to accurately predict a customer’s future spending patterns and the cost of acquiring future sales to that customer. Data warehouses, which store customer data from a variety of sources, and data mining that looks for previously undiscovered patterns in customer transactions, are vital tools in the development of CLTV. They enable retail companies to monitor past and current transactions and add in other information about their customers such as lifestyle or lifecycle stage data, geographical and demographic information. Food retailers have been pioneers in the use of data warehouses to identify purchasing patterns and develop loyalty programs; now retail banks and insurance companies are using increasingly sophisticated

45 data warehouses to project future income and profit streams from individual customers (Ryals, 2002).

2.3.1 Benefits of CLV Measurement

Customer lifetime value (CLV), which appears elsewhere in management literature as customer equity_ and customer profitability_ helps firms quantify customer relationships. Customer lifetime value looks at what a retained customer is worth to the organization now, based on the predicted future transactions and costs (Berger and Nasr, 1998). It is therefore far more useful than historic customer analysis as a basis for developing marketing strategies to maximize shareholder value (Ryals 2002). Customer lifetime value (CLV) is typically used to identify profitable customers and to develop strategies to target customers (Liu and Shih, 2005b).

Also Haenlein (2007) mentions that central to the idea of CRM is the assumption that customers differ in their needs and the value they generate for the firm, and that the way customers are managed should reflect these differences. CRM is therefore not about offering every single customer the best possible service, but about treating customers differently depending on their CLV. Such appropriate treatment can have many faces, starting with offering loyalty programs to retain the most profitable customers through to the abandonment of unprofitable customer relationships (Haenlein et al., 2007). Yet, selecting between these strategies requires that the company knows the value its different customers generate (Haenlein et al., 2007).

Effectively, marketers need to predict the future purchasing behavior of customers to arrive at their Customer Lifetime Value (CLV) (Ryals, 2002). Looking forward to the value of future sales and costs (expressed as the present value of a stream of future profits) fits more comfortably with the development of CRM strategies than current period profits (Ryals and Knox, 2006). Also, forward-looking customer lifetime value measures are more consistent with the principles of shareholder value creation.

46 2.3.2 Limitations for CLV measurement

Unfortunately, even if CLV calculations show a positive return, it is still possible that a customer relationship could destroy shareholder value. One reason is that commonly-used profit measures do not necessarily reflect shareholder value creation because they do not take into account the true cost of capital (Ryals and Knox, 2006).

Also Hwang et al. (2004) mention that many researches have been performed to calculate customer value based on customer lifetime value (LTV), however, they all have some limitations because it is difficult to consider the defection of customers. Prediction models have focused mainly on expected future cash flow derived from customers’ past profit contribution.

Many researchers state that existing CLV models still have limitations in applicability for three reasons:

1. Customer behavior is the result of a complex interaction among factors including the level of marketing activity, the competitive environment, brand perception, the influence of new technologies, and individual needs. Therefore, current CLV models, which predict purchase behavior based on past customer spending patterns or demographic characteristics, are of limited use in predicting future behavior. In order to extend the basic CLV model and effectively apply it to a complicated open market, additional factors must be discussed and considered, including social effects, competitive effects, economic environment, product lifecycle, customer lifecycle, and customers’ purchasing habits, lifestyle, customer satisfaction, price sensitivity and brand loyalty. Without taking account of these factors, authors believe that CLV models are of limited use (Wang and Hong, 2006). Therefore, research is needed to investigate which factors determine customer profitability (K.Stahl et al., 2003), and which factors determine the distribution of profitability among consumers (Jain and Singh, 2002). However, in the framework of conventional CLV models, it would be too complicated to associate with these factors because of the structural and data differences between the factors. Aggregating these factors in order to capture the timing and stochastic nature of revenue flows is almost impossible in a CLV appraisal (Wang and Hong, 2006). Moreover, the probability-based CLV models only guarantee that the models predict well within the time horizon of the collected data, but there is no guarantee for forecasting values beyond that horizon. For volatile customer profitability, it is especially difficult for a forecasting model to predict dramatic upward or downward trends generated by, for instance, a new product or service provided by a competitor. Therefore, firms need a more flexible model, which would not only able to incorporate managers’ judgment calls and other uncertain factors, but also detect changes in customers’ behavior. As the result, firms

47 would be able to monitor and calibrate marketing action in response to unpredictable customers’ behavior. 2. Existing CLV models provide a static estimate of customer valuation for a given future period to segment customers into several levels of a firm’s customer pyramid, such as profitable, less profitable and unprofitable (Wang and Hong, 2006). However, dynamic markets require a more tactical view towards these measures. The possible directions in changes, and the possible volatility over customer profitability have been considered as the effective indices to trace the customer behavior. The direction of customer profitability is a reliable indicator of the customer’s status (defecting, upgrading or steady), and volatility represents the possible risk level of a customer’s profitability for a firm. Therefore, a mechanism to monitor possible directions and volatility of customer profitability would enable firms to dynamically adjust marketing activity towards their targeted customers. 3. There is a lack of practical discussion about how to incorporate customer profitability measures into marketing planning (Wang and Hong, 2006). Wang and Hong (2006) also mention that to develop a marketing activity, a firm needs to draw a picture of its customers by combining customer profitability with customer accessibility, needs and attitudes. Analysis of customer profitability is often used to indicate possible consumption patterns of the targeted customers. However, this index may not be sufficient for identifying the customers a firm truly wishes to acquire or retain through allocating additional marketing resources. Further information of customer accessibility and customer attitudes is needed. Customer accessibility represents the possibility of accepting a marketing package by a potential targeted customer or segment, and could be estimated through past marketing or contact records stored in database or given by account/ product managers based on their experiences and judgments. Moreover, customer attitudes, such as customers’ preferences and their levels of satisfaction, can provide the information for a firm to offer the right marketing packages to meet customers’ needs.

In summary, simply relying upon the measurement of CLV to determine Customer Relationship Management (CRM) success can be misleading because it ignores the dynamics of customers’ purchasing behavior.

2.4 CLV Applications

Customer lifetime value (CLV) is an important concept in segmenting, selecting, and retaining customers, and in customer management in general (Crowder et al., 2007). It describes the likely return on a customer over the course of their relationship with the company. There has been a wide range of applications for customer lifetime value. Some applications are mentioned here.

48 2.4.1 Segmentation/ Product Recommendation/ Retention

Precise evaluation of customer value and targeted customer segmentation must be critical parts for the success of CRM. (Hwang et al., 2004) and (Kim et al., 2006) for instance, have used CLV measures to segment customers of a wireless telecommunication industry in which competition is very stiff and customer churn is rapid.

Recommending products to attract customers and meet their needs is important in fiercely competitive environments. Recommender systems have emerged in e-commerce applications to support the recommendation of products. Liu and Shih (2005) propose two hybrid methods that exploit the merits of the WRFM-based method and the preference-based CF method to improve the quality of recommendations (Liu and Shih, 2005a).

Membership to customer loyalty initiatives provides members with rewards and additional value, making it popular among consumers. This has led to an increasingly competitive landscape with different companies within the same retailing industry vying with one another to woo the same set of customers. Consequently, consumers often enroll in loyalty programs of multiple companies within the same industry (Kumar and Shah,

2004). These authors propose a conceptual framework for building and sustaining loyalty and profitability simultaneously at individual customer level. A two-tiered rewards structure is presented as a means for marketers to operationalize the framework. The conceptual framework hopes to serve as a platform to understand the evolving dominant logic of loyalty programs for building and sustaining loyalty in the twenty first century.

2.4.2 Resource Allocation

By appropriately allocating marketing resources to targeted customers, firms are better positioned to increase profits (Wang and Hong, 2006). For instance, (Berger and Bechwati, 2001) offer a general approach to the organization of promotion budget allocation, where the objective function is to maximize customer equity. Another instance is brought by (Liu et al., 2007) where they propose a general model of resource allocation

49 by which service managers can shape their thought process in decision making. Authors apply customer lifetime value models to assess the overall value of the service encounter and to establish implications that such an assessment has for managing customer relationships under a fixed-size sales force. Using a specific relationship between customer servicing activities and the buying rhythms of customers, an analytical model for assessing the overall value of a service encounter is developed by them. Managerial questions that could be answered by their model are How much capacity should be committed to a given customer type? And, which customer type should be given priority in terms of committing capacity?

2.4.3 Merger and Acquisition

From a Marketing perspective, M&A transactions are nothing other than the acquisition of the customer base of one company by another one, usually based on the assumption that the acquiring firm can manage this customer base more profitably than the selling firm was able to (Haenlein et al., 2007).

As stated by (Kumar et al., 2004), independent analysts without access to confidential customer level data of a firm can use the average CLV approach to arrive at the market value of firms. Kumar et al. (2004) mentions that, (Gupta et al., 2004) demonstrate that for high growth companies, arriving at the firm’s aggregate CLV or customer equity could indeed be a good surrogate measure of the market worth of a firm. Thus, computing aggregate CLV of the target firm can assist in making merger and acquisition decisions, when traditional measures like P/E are not appropriate indicators.

2.4.4 Customer Equity

The ability of a company to acquire and retain attractive customers is ultimately crucial for that company’s success in a competitive market. Blattberg and Deighton (1996) have coined the term Customer Equity (CE) for this value potential. Using their definition as a basis, authors are interpreting this to mean the sum of the discounted cash surpluses generated by present and future customers (within a certain planning period) for

50 the duration of the time they remain loyal to a company, i.e. the sum of individual customer lifetime values (CLV) from the company’s point of view.

Customer equity is “a combination of a firm’s current customer assets and the value of the firm’s potential customer assets” (Kumar and Goerge, 2007) or in other words it is defined as the total of the discounted lifetime values summed over all of firm’s current and potential customers.

The customer equity approach to marketing management finds its origins in several overlapping research streams including direct marketing, service quality, relationship marketing and brand equity (Aravindakshan et al., 2004). Conventionally, firms evaluated the profitability of an organization in terms of product profitability. Thus, the activities and decisions of the firms were organized around products and their functions. However, in today’s business environment, sales success results from relationships more than from individual transactions. To ensure profitability, a good or service must not only be purchased by a number of consumers, it must be repurchased repeatedly which means the key challenge to firms is satisfying the customer. Thus the customer forms the key challenge to a firm’s profitability, thus attracting and maintaining customers is essential for firms to adopt a customer-centric focus. This customer-centric view leads to the concept of customer equity (Aravindakshan et al., 2004).

How a firm gains customer equity is dependent on three key drivers: value equity (i.e., the customer’s objective evaluation of the firm’s offerings, e.g., quality or price), brand equity (i.e., the customer’s subjective view of the firm and its offerings, e.g., brand awareness and attitude toward the brand), and retention equity (i.e., the customer’s view of the strength of the relationship between the customer and the firm, e.g., loyalty and affinity programs) (Rust et al., 2000).

According to (Aravindakshan et al., 2004) current interrelated trends that are shaping economic change in every developed economy make it inevitable that management will shift its focus from brand equity to customer equity. This shift is guided by the following trends:

51 1. The Shift from Goods to Services: The underlying basis for all of the trends is the dramatic long-term shift of developed economy from goods to services (Aravindakshan et al., 2004).

2. The Shift from Transactions to Relationships: A service economy facilitates a shift from a focus on customer transactions to a focus on long-term one to one customer relationships (Aravindakshan et al., 2004).

3. The Shift from Customer Attraction to Customer Relationship: For the firm, keeping or retaining customers is crucial to its success. While attracting customers is still important, customer retention is crucial to the firm’s long-term survival (Aravindakshan et al., 2004).

4. The Shift from Product Focus to Customer Focus: In general an increasing emphasis on customers and relationship management coincides with a decreasing emphasis on products (Aravindakshan et al., 2004).

Hence, it is easy to see that, for most firms, customer equity is bound to be the most important component of the value of the firm. Thus, understanding how to drive customer equity is central to the decision making of any firm and formulating a procedure to achieve this can give the firm an important competitive advantage (Aravindakshan et al., 2004). So the more reliable the proposed methods are for calculating this figure, the more important it will become as a criterion on which to make investment decisions (Bayon and Gustsch, 2002).

2.4.5 Marketing Campaign

One approach to measure customer value is to measure the value that a customer or a category of customers bring into an organization, and use this as a targeted marketing campaign. This can work in two ways: using the knowledge of high value customers to offer them additional information or incentive to maintain their loyalty, or offering incentives to lower value customers to try and move them into the high value category (Goerge Evans, work study, 2002).

2.5 CLV Calculation Models

The appropriate way to model lifetime value depends on the nature of the business and on the aims of the modeling (Crowder et al., 2007).

52 As corporations increasingly come to see customers as important assets, methods for estimating Customer Lifetime Value (the CLV model) have been developed as an important strategic marketing tool (Ryals, 2002).

There are a lot of researches on calculating customer value. The basic concept of these researches, however, follows

(2.1)

where i is the period of cash flow from customer transactions, Ri the revenue from the customer in period i; Ci the total cost of generating the revenue Ri in period i; and n is total number of periods of projected life of the customer under consideration. Therefore, the numerator is the net profit that has been obtained at each period while the denominator transforms the net profit value into the current value (Hwang et al., 2004).

In this section we will review some basic calculation models for Customer Lifetime Value, their assumptions and specifications, the data set they have been tested on (if applicable), and the purpose they have used the model for. We have tried to bring the widest range of mathematical models to get a broader view of different types of modeling. The models are tried to appear in the order of dates from the latest to the most recent ones.

2.5.1 Models by Blattberg and Deighton (1996); Berger and Nasr (1998); Gupta, Lehmann (2003); Blattberg, Getz and Thomas (2001); and Rust, Lemon and Zeithaml (2000)

Customer equity, the asset value of customers, can be measured using different aggregate- and disaggregate-level approaches. A comparison of different aggregate approaches shows that, while an emphasis on retention is a common feature across approaches, conceptual differences in terms of accounting for existing customers and

53 prospects, acquisition, and the projection period exist across the different approaches (Kumar and Goerge, 2007).

2.5.1.1 BD model

Blattberg and Deighton (1996) developed an aggregate-level approach in the context of using customer equity for balancing acquisition and retention. Under this approach, the optimal acquisition rate is calculated by identifying the shape of the acquisition curve and finding the acquisition rate at which the customer equity is maximized. The optimal retention rate is also calculated in a similar manner. Incorporating these two rates, the customer equity is expressed as:

(2.2)

where, r: r/(1+d)

a: the acquisition rate given a specific level of acquisition cost, A

m: contribution margin

A: Acquisition cost per prospect

R: Retention cost per customer

r: yearly retention rate

d: yearly discount rate

This is similar to the BGT approach in terms of how customer equity is expressed as the sum of returns from acquisition spending and retention spending. However, the emphasis in this framework is on the calculation of the optimal level of acquisition and retention spending (Kumar and Goerge, 2007).

54 2.5.1.2 BN model

The BN model, Berger and Nasr (1998), introduces the basic CLV model based on three main assumptions: (a) sales takes place once a year, (b) yearly retention spending (M) and retention rate (r) remain constant over time, and (c) yearly gross contribution margin (GC) remains the same (Kumar and Goerge, 2007). Under these assumptions CLV is computed as:

(2.3) where, n: length in years; d: yearly discount rate. Here, the promotional expense, M is expected to occur at the middle of the purchase cycle. When authors relax the assumptions about the constant retention rate, gross contribution margin, and promotional expenses, and assume that the purchase cycles can be longer or shorter than 1 year, the CLV equation gets modified as:

(2.4)

where (t) the profit per customer in year t, which can be estimated separately using the appropriate equation for the profit curve.

2.5.1.3 GL Model

The GL approach (Gupta and Lehmann 2003; Gupta et al. 2004) simplifies the

CLV formula, with certain assumptions such as constant average margins (m), constant retention rate (r), and infinite projection period. The lifetime value of a customer is calculated using the following equations:

Case 1: When the average margins are constant,

55

(2.5)

Case 2: When the margins grow at a constant rate g per period,

(2.6)

where, m: constant average margin

i: discount rate

r: constant retention rate.

Margin multiple is r/(1+i−r) when margins are constant and r/(1+i−r(1+g)) when margins grow at a constant growth rate. (Kumar and Goerge, 2007)

2.5.1.4 BGT Model

In the BGT approach (Blattberg, Getz and Thomas, 2001), customer equity is calculated as the sum of return on acquisition, return on retention and return on add-on selling across a firm’s entire customer portfolio (Blattberg et al. 2001 cited by Kumar and George, 2007). One component of the equation computes returns from acquisition as the contribution from newly acquired customers minus the cost of acquiring them. The other component of the equation calculates the expected profits from future sales to these newly acquired customers adjusted for retention rate and time value of money. This is expressed in a mathematical equation as in (2.7):

56

(2.7)

where, CE (t): the customer equity value for customers acquired at time t

Ni,t: the number of potential customers at time t for segment i

ai, t: the acquisition probability at time t for segment i

i, t: the retention probability at time t for a customer in segment i

Bi,a,t: the marketing cost per prospect (N) for acquiring customers at time t for segment i

Bi,r,t: the marketing costs in time period t for retained customers for segment i

Bi,AO,t: the marketing costs in time period t for add-on selling for segment i

d: discount rate

Si,t: sales of the product/services offered by the firm at time t for segment i

ci,t: cost of goods at time t for segment i

I: the number of segments

I: the segment designation

t0: the initial time period.

Under this approach, the computation of customer equity is for each segment or cohort rather than for individual customers. Since average acquisition probability and retention probability for the customer segment are used in the model, authors can only get an aggregate measure of customer equity (Kumar and Goerge, 2007).

57 2.5.1.5 RLZ Model

The RLZ approach (Rust, Lemon and Zeithmal, 2000) uses a CLV model, which incorporates customer-specific brand-switching matrices, although only for customers in the selected sample (Rust et al. 2000, 2004 cited by Kumar and George, 2007). This approach uses information about both the focal brand and the competing brands to model acquisition and retention of customers in the context of brand switching. Respondents in a selected sample provide information such as the brand purchased in the previous purchase occasion, the probability of purchasing different brands, and individual-specific customer equity driver ratings. The Markov switching matrix then models individual customers’ probability of switching from one brand to another based on individual-level utilities. The probability thus calculated is multiplied by the contribution per purchase to arrive at the customer’s expected contribution to each brand for each future purchase. Summation of expected contribution over a fixed time period after making adjustments for the time value of money (i.e. applying a discount factor) yields the CLV for the customer. The lifetime value, CLVij of customer i to brand j, is given as:

(2.8)

Tij: number of purchases customer i makes during the specified time period

dj: firm js discount rate

fi: average number of purchases customer i makes in a unit time (e.g., per year)

Vijt: customer i’s expected purchase volume of brand j in purchase t

Ǚ ijt: expected contribution margin per unit of brand j from customer i in purchase t

Bijt: probability that customer i buys brand j in purchase t

58 Customer equity of firm j, CEj is then calculated as;

(2.9)

meani(CLVij): the average lifetime value for firm j’s customers i across the sample

POP: total number of customers in the market across all the brands

From the above equation, it is clear that individual-level CLVs are calculated only for the customers in the sample.

The mean CLV calculated from the sample is taken as the average worth of a firm’s customers. Based on the classification of approaches given earlier, authors treat this approach as an aggregate-level approach because it computes only mean CLV, and not individual CLVs.

2.6 Models by Stahl (2003) and Wu (2005)

Referral or word-of-mouth effects, where existing customers recommend their favorite retailers to friends and family, are highly valued by potential customers (Murray 1991). Anton (1996) cites a study showing 61% of respondents felt that the opinions of their friends were ‘very useful’ in selecting which retail outlet to use. ‘Friends and family’ deals are relatively common in telecoms, mail order clothing and motor insurance. Investment in customer relationships can pay off with a stream of referrals and positive word of mouth over a period of time, underlining the need to treat customers as assets. First Direct is a well known example of a financial services organization that has generated about one-third of its customers through referral, as existing customers recommend the bank to their family and friends. Customers who come through a referral usually have lower acquisition costs than those who do not; moreover, new customers referred by existing loyal customers are themselves more loyal. The reverse also applies—negative word of mouth can be extremely costly in terms of future profits

59 (Ryals, 2002). Referencability is when a customer allows a company to tell others that it supplies what is usually a large or prestigious account. Wilson (1996) as stated by (Ryals, 2002) refers to endorsement (flagship customers that can help suppliers win additional business).

2.6.1 Framework developed by Stahl (2003)

A framework was developed to analyze the relationship between customer lifetime value and shareholder value. It is based on four processes that drive shareholder value as shown in Figure 2.1, namely (1) increasing cash flow, (2) accelerating cash flow, (3) reducing cash flow volatility and vulnerability and (4) increasing the residual value of the firm. It has been shown how the four components of customer lifetime value are related to these drivers. The authors of this paper provide a conceptual framework for linking customer lifetime value to shareholder value.

Figure 2.1 The four components of CLV adapted from Stahl (2003)

It is argued that customers have to be treated as assets that increase shareholder value by accelerating and enhancing cash flows, reducing cash flow volatility and vulnerability and increasing the residual value of the firm. It is generally agreed that the market value of a firm emanates from the net present value (NPV) of future cash flows

60 generated by the firm’s assets, discounted at an appropriate interest rate and adjusted for inflation and risk. Hence, proponents of the shareholder value approach claim that strategies and initiatives must be evaluated against the NPV of the cash flows they generate (Stahl et al., 2003).

Figure 2.2. Measurement of CLV adapted from Stahl (2003)

Stahl’s framework should be seen as a practical approach to shareholder-value- based customer valuation. It allows marketing practitioners to allocate resources in a way that is more shareholder value oriented than traditional methods. This analysis was done on the data of manufacturer of industrial goods mostly for analysis of the linkages rather than assessment of those linkages (Stahl et al., 2003).

Critique: In the conceptual framework proposed by Stahl , no practical solution is introduced to address the “word-of-mouth” effect contributed by customers (Wu et al., 2005).

61 2.6.2 Model by Wu (2005):

One notable feature of Wu’s model is that it combines the social influence among customers into the CLV estimation. He actually takes the social influence factor of customers into account and combines it into the estimation for potential value and network value. The framework developed by Wu (2005) is shown in figure 2.3.

(Wu et al., 2005)

Figure 2.3. Conceptual framework for measuring and using CLV

The calculation model that supports this framework proposed by Wu (2005) is:

(2.10)

CLVi lifetime value of customer i time index forecast horizon discount rate basic value of customer i at time t potential value of customer i at time t network value of customer i at time t retention rate of customer i at time t total cost of acquiring, developing or retaining customer i at time t

62 The basic value catches on the usage level of the consumed products or services that form the base of the customer relationship. Authors adopt time series modeling method to predict the monetary amount of a customer spending on the up-selling products in some future - time period. Given historical spending data of up-selling products, time series model can capture the fluctuant regularity of historical expenditures and use the regularity to infer the future monetary amount of the customer spending on them.

The potential value is the expected profits from the cross-selling to a customer. It reflects the number of different products or services bought by a customer from the same supplier and measures the breadth of the customer relationship. Authors have measured it by extending the probit model used by (Verhoef and Donkers, 2001) to predict a customer’s purchase decision on the cross-selling product or service.

The retention rate of customer i here is taken as a weight factor to tradeoff the benefit and risk, and the discount rate factor. Customer retention rate i is the likelihood of a customer staying in or terminating a relationship with a company in a certain period. It reflects the length of customer relationship. Authors adopt probabilistic neural networks (PNN) to calculate the customer retention rate. PNN can be used for classification problems. For a given input pattern characterized by a set of features, PNN classify it into one of k classes (e.g. retention class or churn class) with maximum probability of being correct.

The network value is the additional profit expected from the network influence of a customer exerting on other network members. It measures the extension of the customer relationship, and together with the social effect customers have on each other, models the dyadic processes of social influence. That is the buying decisions of a customer influence the choices of other customers, and the customer in turn are influenced by others.

To valuate the network value of a customer, the number of new customers acquired through his referrals and the average profitability of each have to be assessed. However, historical measurement of customer-to-customer referrals is a challenging task

63 since it is very difficult to get the data needed even if the prevalence of Internet has facilitated the communication between customers greatly.

Wu’s model has been tested on the data set of a bee product company in health care industry. The results from Wu’s study analysis suggest that proposed CLV model provides an effective assistant tool for the company to identify and select its most valuable customers.

2.7 Models by Hwang (2004) and Kim (2006)

LTV models evaluate the long-term value of customers focused on the entire lifetime of customers. The lifetime of customers describes the period that the customers are staying as customers. However, the long-term value does not fit for the industry having stiff competitions and rapid changes of market environments. Especially, it is not easy to evaluate the LTV of customers in the wireless communication industry, which are very sensitive to the external environments and the customer defections. Hence, the study of authors focuses on the short-term value of customers of a wireless communication industry. Since customer value cannot be evaluated at a time, it has to be regularly and continuously renewed for covering the disadvantages of the short-term based evaluation.

Furthermore, the evaluations of customer value in previous studies have treated prediction method with regression models simply based on profits from customers to calculate the future value of customers. That is to say, considering the changing profit contribution obtained from customers in the past, the existing models calculate the future worth and then define the LTV of customers with the projected value of the future worth. Therefore, the LTV model above is not capable of considering potential values of customers, not available from the past profit contribution, which would be able to be the profits of companies. Also these models do not consider the defection of customers. So possibility of defection should also be considered.

Hence, it is reasonable to consider the probability of individual customer’s churn rather than to consider only the total decreasing rate of whole customers. Verhoef and Donkers (2001) used two dimensions, current value and potential value, to segment the

64 customers of an insurance company (Verhoef and Donkers, 2001). In Hwang’s study, authors use three dimension, current value, potential value, and customer loyalty, to consider the customer defection. The conceptual framework of Hwang’s study is:

(Hwang et al., 2004)

Figure 2.4. Conceptual framework of Hwang

Customer segmentation methods using LTV can be classified into three categories: (1) segmentation by using only LTV values, (2) segmentation by using LTV components and (3) segmentation by considering both LTV values and other information.

In the first method, the list of customers’ LTV is sorted in descending order. The list is divided by its percentile. In this case, researchers segment customer list by only LTV however, other information like socio-demographic information or transaction analysis may be used together for a better marketing practice.

For instance, after segmenting a highly profitable customer group, a firm may recommend popular products to the targeted group at a discounted price. Figure 2.5 briefly depicts the concept of segmentation using only LTV.

65

(Kim et al., 2006) Figure 2.5. Segmentation by using only LTV values

The second method performs segmentation by considering components used in LTV calculation. Hwang et al. (2004) considered three factors: current value, potential value, and customer loyalty to calculate LTV and present the method to segment the three factors for customer segmentation. Figure 2.6 shows segmentation using factors in calculating LTV.

(Kim et al., 2006) Figure 2.6. Segmentation by using LTV components

The last method is to segment the customer list with LTV value and other managerial information. In this case, LTV is an axis of the segment in n-dimensional segment space and other information, such as socio-demographic information and transaction history become another axis. This approach is more meaningful for segmenting the customer list than the first method. Figure 2.7 shows a segmented customer list with LTV value and other managerial information.

66

(Kim et al., 2006) Figure 2.7. Segmentation by considering both LTV values and other information

Generally, three core works are necessary for the increase of customer value: up- selling, cross-selling, and customer retention ((Kim, 2000) cited by Hwang, 2004). Up- selling is selling the same kinds of products that a customer has already bought and cross- selling is selling what a customer have never bought, i.e. new kinds of products for the customer. Customer retention means the effort to keep our customers being stayed as ours, prohibiting them from changing their minds. It is reasonable to consider these three sides when considering customer value (Hwang et al., 2004).

The potential value represents a measure of cross-selling possibility while the customer loyalty denotes a measure of customer retention. Authors define current value as a profit contributed by a customer during a certain period (for six months), not as a cumulative value from the past to the current point.

Current value can be obtained from a simple calculation with the data fields. In his paper, authors have settled current value as the average amount of service charge asked to pay for a customer minus the average charge in arrears for a customer, regarding six months for calculation. As there is no field related to the average amount in arrears in the data, the average amount in arrears were calculated from the cumulative amount in arrears divided by the total period of use.

67 Customer Value =(Average amount asked to pay - Cumulative amount in arrears)/total service period (2.11)

As mentioned before, it is important to consider cross-selling and up-selling as well to calculate customer value (Kim and Kim, 1999). In particular, cross-selling opportunity needs to be considered to evaluate customer value in the wireless communication industry since many profitable optional services are available for customers. Authors define here potential value of customers as expected profits that can be obtained from a certain customer when a customer uses the additional services of a wireless communication company.

An approach that calculates the potential value of customers has been developed by (Verhoef and Donkers, 2001). Following their approach, the potential value of a customer i is calculated as:

(2.12)

Equation (2.12) evaluates potential values where Probij is the probability that customer i would use the service j among n-optional services. Profitij means the profit that a company can get from the customer i who uses the optional service j: In other words, the equation above means expected profits from a particular customer who uses optional services provided by a wireless communication company. The expected profits will become potential value authors need to evaluate. The probability that a customer owns a product is modeled using the probit (univariate or multivariate) model. Socio- demographic indicators are used to predict the ownership of a product or service. The probability thus obtained, is multiplied by the profit margin to arrive at the potential value of the product. This approach may be useful to predict the potential value of a customer in a contractual setting (e.g. life insurance, magazine subscription) where revenues can be projected fairly accurately and where it is important to know whether a customer is going to purchase a product or service in the future.

68 Customer loyalty can be defined as the index that customers would like to remain as customers of a company.

Customer Loyalty = 1- Churn rate

Churn describes the number or percentage of regular customers who abandon relationship with a service provider. Customer loyalty can be a measure of customer retention. Level of customer retention can be derived from churn rate. It is significant for customer cultivation and retention to consider the churn rates. In particular, negative reputations have critical influences on the brand image of a company in the wireless communication industry that includes most of people as customers.

The existing studies on customer value have not treated the churn rate yet, limiting themselves to predict the future profit change of customers with the past profit history. The effective evaluation of customer value, however, should comprehend the leaving probability of each customer.

(2.13)

The sum of represent NPV of the past profit contribution, where is the profit contribution of customer i at period ti and is the interest rate factor, which transforms the past profit into the present value. The future cash flow can be derived from the sum of the expected future profit and potential benefits during the expected service period of customer i.(Hwang et al., 2004)

2.8 Models by Venkatesan (2004) and Kumar (2006)

This model examines how customer lifetime value (CLV) can be computed at individual customer level in a retail setting to maximize profitability.

69 At the disaggregate level approach, the lifetime value of an individual customer is calculated as the sum of cumulated cash flows—discounted using the Weighted Average Cost of Capital (WACC)—of a customer over his or her entire lifetime with the company. The individual CLVs of all customers are then added together to arrive at the customer equity of the firm. The future contribution and thereby, CLV, depends on the customer’s future activity. Venkatesan and Kumar (2004) use the always-a-share approach for predicting future customer activity. They predict the frequency of a customer’s purchases given his/her previous purchases, and use these predicted frequencies to calculate CLV (Kumar and Goerge, 2007).

According to (Kumar et al., 2004), another source of variation among individual customers is the nature of response to messages received by them from different channels. Managers would find it useful to know the differential effects of commonly used marketing and communication channels like face-to-face contact, telesales, direct mail, and web portals on individual customers. Hence, optimizing the mix and frequency of these channels to suit every individual customer is likely to result in higher profitability. The CLV computation that needs to be carried out should now be sensitive to the response of each individual to each available channel. (Venkatesan and Kumar, 2004) as cited by Kumar (2007) suggest an approach that incorporates the cost of contacting customers through each channel and the number of contacts made through each channel. The individual CLV model in such a case is as follows:

(2.14)

where

CLV i _ customer lifetime value of customer i

GCi,t _ gross contribution from customer i in purchase occasion t

ci,m,l _ unit marketing cost, for customer i in channel m in time period l

xi,m,l _ number of contacts to customer i in channel m in time period l

70 frequency i _ predicted purchase frequency for customer i d _ the discount rate for money

n _ the number of periods to forecast

Ti _ the number of purchases made by customer i until the end of the planning period

Customer demographics, exchange variables, and time-varying covariates are used to predict the gross contribution (GC) and the purchase frequency (frequency i). These predictions are incorporated in to the overall model for CLV computation. The optimization process then indicates the number of contacts made through each channel to each customer that would maximize the CLV of each customer.(Kumar et al., 2004)

From Equation (2.14), authors can see that CLV consists of the following main components—(i) purchase frequency, (ii) contribution margin, and (iii) marketing cost. For accurate measurement of CLV, must estimate the purchase frequency, contribution margin and marketing cost for each customer using suitable models and then combine the predictions from the three models to arrive at a single value representing the lifetime value of the customer in dollar terms. Authors have modeled these three components and bring the result of their work here:

2.8.1 Modeling the Purchase Frequency for Each Customer

In a non-contractual setting, the purchases may be made at irregular time intervals and the customer may end the relationship at any time. Authors are, therefore, more interested in estimating the frequency of purchase and the contribution from future transactions and the marketing costs to be invested on the customer as modeled in the VK approach.

Authors have proposed a model for purchase frequency in which the specification of the model allows estimating individual customer level coefficients for the influence of the various covariates on the probability of a customer belonging to a particular subgroup, and hence the inter-purchase times.

71 2.8.2 Modeling the Contribution Margin for Each Customer

The gross contribution for each customer was modeled using panel-data regression methodologies. Hence, the gross contribution model can be represented as:

(2.15)

where GC i,j is the gross contribution for customer i in purchase occasion ‘ j ’ measured in dollars;

Xi,j −1, the independent variable relevant to customer i in purchase occasion j−1; ei,j , the error term;

n, the number of independent variables;

k, the index for the independent variable;

i, the index for the customer;
0,
j are the coefficients.

Authors define gross contribution margin as the revenue the customer provides to the firm, whenever they make a purchase, minus the cost of goods sold. While the cost of goods sold does not change very much over time, there is a possibility of large variance in the revenue a customer provides over time.

Authors found that the simple linear specification of Eq. (3.15) provided the best in-sample fit and predictive accuracy. To help select the appropriate customer-behavior related independent variables that could be tested for predicting purchase frequency and/or the gross contribution of each customer, authors referred to the previous research and theory as well as inputs from the business managers of the retailer (Kumar et al., 2006). Various supplier-specific factors (channel communication) and customer characteristics (involvement, switching costs, and previous behavior) are first identified as the antecedents of purchase frequency and contribution margin. Purchase frequency and contribution margin are then modeled separately using suitable models. In the framework developed by Venkatesan and Kumar (2004), a generalized gamma distribution is used to model interpurchase time, and panel-data regression methodologies

72 are employed to model the contribution margin. (Kumar and Goerge, 2007) call the above approach for measuring customer lifetime value as the VK approach.

The CLV model described above can be made use of to identify the responsiveness of customers to marketing communication through different channels. This forms the basis for optimal allocation of marketing resources across channels of contact for each customer in order to maximize his or her respective CLV. The CLV framework can also be used for formulating other customer-level strategies such as customer selection, purchase sequence analysis, and for targeting the right customers for acquisition (Kumar and Goerge, 2007). The variables tested and finally selected in the model are summarized in Table 2.2.

Table 2.2. Variables included for predicting purchase frequency and/or gross contribution

(Kumar et al., 2006)

73 2.8.3 Computing the Marketing Cost for Each Customer

There are several methods available to compute the marketing cost for each customer. For example, it may be calculated as an aggregate measure by dividing the total marketing budget by the number of customers. A more sophisticated approach would entail calculating the total marketing cost separately for each customer based on the various marketing channels expected to be used to interact with that customer. This can be denoted as shown in Equation (2.16).

(2.16)

where MC i is the total marketing cost for customer i; ci,m,l , the unit marketing cost, for customer i in channel m in time period l; xi,m,l , the incidence of marketing customer i in channel m in time period l ; m, the marketing channels (for this retailer it was web and catalog); r is the discount rate for money; l, the index for time; n, the number of years to forecast. Eq. (3.16) provides the retailer the means to compute the marketing cost not only by individual customer but also by individual marketing channels adopted for the same customer. This is imperative given the fact that direct marketing cost could vary widely across customers and across communication channels. For example, it would cost a retailer a fraction of the cost to send three different shopping catalogs to a customer by electronic mail as compared to sending one glossy catalog by regular mail to another customer. Further, with the increasing trend of multi-channel marketing to the same customer, computing marketing cost as shown in Equation (2.16) can help retailers accurately arrive at a fair estimate of the expected marketing cost per customer. The expected marketing cost is discounted by l years to arrive at the present value of marketing cost.

Since this study was specifically done with the objective of measuring and implementing the CLV metric for a retailer, authors followed the directives specified by the retailer. According to the retailer, the best prediction of future marketing cost for each

74 customer over the next 3 years could be taken as three-times the direct marketing cost in the most recent year (i.e., 2004). In other words, authors assumed the direct marketing cost for each customer to be the same for the next 3 years. This is a naive assumption. However, a quick look into the customer dataset did confirm the fact that historically the direct marketing cost per customer remained more or less constant for this retailer. Note that the marketing cost was available as the total marketing cost (across all channels) for each customer in the dataset.

The predictions from the three models (as explained in (a)–(c)) were integrated to calculate the CLV score for each customer as expressed by Equation (2.14). The CLV scores were then used to rank-order all customers in descending order.

The study finds that maximum positive impact to CLV occurs when the customer cross-purchases, shows multi-channel shopping behavior, stays longer with the firm, buys specific product categories and purchases more frequently with the firm. Interestingly, the CLV follows an inverted U relationship with increase in return of prior purchases (Kumar et al., 2006).

2.9 Models by Pfeifer (2000) and Haenlein (2007)

This is rather surprising as it has long been highlighted that customers may differ substantially across industries and that such differences should translate to the models used to value them. For example (Jackson, 1985) as cited by (Haenlein et al., 2007), stressed that customers can be grouped into different categories, depending on the level of commitment they show to a particular seller. On one end of the spectrum is the ‘‘always-a-share’’ model, which assumes that customers can easily switch part or all of their spending from one vendor to another. The opposite end of the behavior spectrum assumes that, due to high switching costs, the buyer is committed to only one vendor to satisfy his or her needs. Once the customer stops purchasing from this vendor and changes to another one s/he is ‘‘lost-for-good’’ and cannot return to the vendor easily. Although the category every customer can be allocated to depends to a certain extent on this specific customer’s preferences, it is also heavily influenced by the type of product

75 sold and, hence, the industry. Building on this categorization, (Dwyer, 1989) showed that models used to value customers in these two settings differ substantially and proposed two approaches to determining CLV: a customer migration model and a customer retention model.

Pfeifer and Carraway (2000) proposed that a general class of mathematical models called Markov Chain Models (MCM) are appropriate for modeling customer relationships. The authors believed that MCM models are very flexible and can address the situations depicted in models proposed by Berger and Nasr (1998) and Blattberg and Deighton (1996). MCM can be used to model both customer retention and customer migration situations. In most CLV models, when a customer stops being active, then the customer is treated as dead and returning customers are treated as new customers. In such models, there is no provision for a customer to be inactive for sometime while still being retained. Customer migration refers to such a situation where a customer might remain inactive for some periods and still be treated as a retained customer on his/her return. The probabilistic nature of MCM allows accounting for the inherent stochasticity in customer relationships (Jain and Singh, 2002). Pfeifer and Carraway demonstrate the use of MCM in various situations and show how to construct the key elements of MCM, that is, the transition probability matrix and the reward vector. Among the models described in this chapter, although Markov Chain based models are the most flexible, the models proposed have critical assumptions underlying them. For instance, in these models, time period for purchase by all the customers is again assumed to be same, and fixed. The calculation of transitions probabilities is critical to the success of such models and these probabilities are not easy to compute (Jain and Singh, 2002). Another application of MCM is available in Rust, Zeithaml, and Lemon (2000) as mentioned by (Jain and Singh, 2002).

(Haenlein et al., 2007) present a customer valuation model that is developed in cooperation with a leading German retail bank, which takes account of the specific requirements of this industry. Their model is based on a combination of first-order Markov chain modeling and CART (classification and regression tree) and can deal equally well with discrete one-time transactions as with continuous revenue streams.

76 Furthermore, it is based on the analysis of homogeneous groups instead of individual customers and is easy to understand and parsimonious in nature.

2.10 Model by Fader (2007)

The original equation for CLV calculation, Equation. 2.1, is not applicable in many business settings, particularly those that can be viewed as non-contractual. A defining characteristic of a non-contractual setting is that the time at which a customer becomes inactive is unobserved by the firm; customers do not notify the firm “when they stop being a customer. Instead they just silently attrite” (Mason, 2003) . This is in contrast to a contractual setting, where the time at which the customer becomes inactive is observed (e.g., when the customer fails to renew his or her subscription, or contacts the firm to cancel his or her contract). When the point at which the customer disappears is not observed, we cannot meaningfully utilize notions such as “retention rates” and therefore formulae along the lines of original equation are not appropriate. We can, however, capture the “silent attrition” phenomenon by using a probabilistic dropout process for each customer. We can define the survival probability,” S(t), for each customer at a given time t, (i.e., the probability that the customer is “alive” at t). This leads to the following definitional expression for expected CLV

(2.17)

where E[v(t)] is the expected value (or net cash flow) of the customer at time t (if active). The challenge is to operationalize (2.17) in any given setting (Fader et al., 2007).

As cited by (Jain and Singh, 2002), a model is proposed by (Schmittlein and Morrison, 1987), called the Pareto/NBD model, that calculates the probability that a customer is still active. The model requires the number and timing of customers’ previous transactions as input. Using this model, firms can identify and count the customers who are still active. The authors demonstrated that this model can be used to answer questions about the number of retail customers that a firm has, the growth of this customer base

77 over the past year, which individuals in the customer group most likely represent active and inactive customers, and what level of transactions should the firm expect next year by those on the list, both individually and collectively. The proposed model is as follows:

For ,

(2.18)

, ,

where are model parameters; t is the time since trial at which the most recent transaction occurred; T is the time since trial; F(a1, b1; c1; z) is the Gauss hypergeometric function; x is the number of purchases the customer makes in time period (0, T] with the last purchase coming at time t _ T. It is assumed that the customer is “alive” (active) at time 0 (Schmittlein and Morrison, 1987).

The Pareto/NBD model is applicable in contexts where the time when the customer becomes inactive is unknown to the analyst and the customer can make any number of purchases, at any time, and can become inactive any time. This model can be very useful to firms having few long-term customers.

Pareto/NBD type models proposed in CLV literature have limitations concerning the input data requirement for each model. Such models might give misleading results if a data string of transactions for more than two years for a customer is included as an input to the model. The implicit stationarity of Pareto/NBD model is strained by very long (purchase) histories (Schmittlein et al., 1987). In the calculation of CLV, the most critical part is the determination of the number of customers who are still active and the number

78 of customers who will be active in each future time period. This model provides a sophisticated way to get these probabilities of a customer being active in each time period. The probabilities thus obtained can then be used to calculate CLV.

One example of a non-contractual business setting is presented in the Tuscan Lifestyles case ((Mason, 2003) cited by (Fader et al., 2007)). This case provides a summary of repeat buying behavior for a group of 7,953 new customers over a 5-year period beginning immediately after their first-ever purchase (Fader et al., 2007). Fader et al., (2007) in their paper try to arrive at an estimate of CLV that includes the customer’s “life” beyond five years or are interested in, say, sorting out the purchasing process (while “alive”) from the attrition process, and they use the Pareto/NBD model originated by (Schmittlein and Morrison, 1987) as discussed earlier in this section, of buying behavior that can be applied on such data .

2.11 Model by Crowder (2007):

In this paper, a particular aspect, namely the expected income to the firm from a customer over his ‘lifetime’, and also determination of the length of the probationary period and the criterion for continuation is examined. The present model, considers making a single terminate/continue decision (extension to multiple decision points is possible), based on the customer’s value record up to the decision time. Such a model is appropriate for a wide variety of business situations, including rolling credit operations, loan top-ups, telecoms contracts, etc. Within this framework, it finds the value threshold and review time that optimize overall CLV.

Crowder (2007) states that Rosset et al. (2003) described an approach to CLV in which, for a given customer, the rate of accrual of profit is v(t), where t is time; this value function can be appropriately discounted. The length of time for which a customer stays with the company, his ‘tenure’, is T, a random variable with distribution function F(t). Thus, the expected lifetime value of a customer, over his tenure, is shown in equation 2.19 below:

79 (2.19)

Suppose that a customer opens an account with a company at time 0. Over the subsequent period (0,t) the company receives accumulated income V1(t) from the customer and the cost to the company of managing the account over this period is C1(t).

Consider the following company strategy. The customer account is reviewed at time r1 (i.e. when its age is r1), assuming that T 1 > r1. If , where v1 is some specified income-threshold level, the account is terminated by the company.

Otherwise, if , the customer is encouraged to renew his account by the company’s offering a new agreement, perhaps on more favorable terms. The new arrangement entails a new income process, V2, a new cost process, C2, and a new account lifetime, T2.

Under the single-review strategy the income to the company over the lifetime of the customer’s account, the CLV, is given by

(2.20)

Authors do not assume that the processes V1 and V2 are independent. On the contrary, the whole point of the probationary period (0, r1) is to observe the individual customer’s value with a view to predicting his future potential value. Specifically, authors will replace V1 and V2 by ZU1 and ZU2, where Z is a random effect representing a personal value-factor of the individual, and the processes U1 and U2 have upward trends. Thus, a ‘good’ customer will have a large, positive Z-value, yielding a high rate of income accrual; a ‘bad’ one will have a small positive, or even negative, Z. The expression (2.20) for CLV becomes

80

(2.21)

The paper then describes how to maximize the CLV equation above by taking partial derivative on r1 and v1. In fact the whole point of the paper is finding out the best time and amount of revenue (v1) at which the company decides whether to continue business with that particular customer or terminates its relationship. The paper in the end has tested the model on some real financial data from a company to show the practicality of the methodology (Crowder et al., 2007).

2.12 Summary of the Chapter

This chapter first reviewed the development of Customer Lifetime Value (present value of a customer’s future purchases) from more basic marketing concepts such as CRM and relationship marketing. The limitations of historic data analysis were also brought up to reach the pros and cons of CLV measurement. Some applications of CLV were also discussed in this chapter. The most important benefit of CLV versus the profitability measurement is that it is a forward looking method versus the historic data analysis method. It also can quantify relationships. It is therefore far more useful than historic customer analysis as a basis for developing marketing strategies to maximize shareholder value. On the other hand, this chapter brought up some limitations of CLV modeling among which is the inability of CLV models to consider many constructs of profitability such as social effects, competitive effects, economic environment, product lifecycle, customer lifecycle, and customers’ purchasing habits, lifestyle, customer satisfaction, price sensitivity and brand loyalty. Among other most recognized applications of CLV are segmentation, product recommendation, merger and acquisition, and decision support (resource allocation decision making).

This chapter also reviewed some models for CLV calculations. It started from the most basic ones which were Berger and Nasr (1998) and ended with Crowder’s (2007) model which is mostly an abstract model. Characteristics of each model (due to industry

81 setting, data requirement, customer relationship) such as contractual or non-contractual setting, aggregate or disaggregate level data, always a share or lost for good relationship, were brought up for each model and for the models which have been tested on real data, the results were mentioned. Models like that of Hwang (2004) or Wu (2006) emphasize on evaluating the dimensions of CLV like potential value, current value and loyalty, while others like Kumar (2006) state that the CLV model should be sensitive to the response of each individual to each available channel of communication like telesales, direct mail, and web portals. On the whole, the industry for which a model is developed should be recognized as contractual, or non-contractual, which would impact some constructs of the model such as the revenue of the customer.

82

Chapter 3

Model Development

3. Model Development

In this chapter we will go through the tasks undertaken to develop a risk adjusted CLV model for a financial service provider as stated in the introduction chapter of this report. The focus would be on modeling the relationship risk which we consider to be consisted of both qualitative and quantitative constructs and each would have a different approach for estimation.

83 The only main CLV model which is risk oriented will be discussed in this chapter and the modeling structure of basic elements of our CLV model is also discussed. Since our case is a bank and our focus is on lending relationship with business customers, our CLV model would be specified to loan applicants of a commercial bank.

3.1 Relationship Risk Score Modeling

Ryals and Knox (2006) in their research prepared a relationship scorecard for business customers of an insurance company according to nine main factors they had extracted. Their factors were extracted by semi-structured interviews with KAM’s (Key Account Management) team of an insurance company. These factors included number of customer relationships within the company, number of products bought by the customer, longevity of relationship, and how good is the company’s understanding of customer’s company and industry. This relationship risk scorecard was then used to analyze the 10 key accounts for which full data were available. The probability adjusted CLVs were calculated for these key accounts meaning that the risk factors were given probabilities by the KAM team to show the retention probability of each key customer on the basis of extracted factors. These retention probabilities were then applied to future revenues of the selected customers.

In our research, we too emphasize on risk considerations mainly for revenue generation of business customers, and as mentioned in chapter 1, the B2B lending relationship is the focus of our research which makes our industry setting for CLV modeling a contractual one. However instead of using the risk factors for retention probability estimation as done by Ryals (2006), we use the extracted factors for probability of default estimation of customers which in a sense means the retention probability of our target customers: credit facility applicants.

The methodology for extraction of these attributes is the Delphi method instead of semi-structured interviews conducted in Ryals (2006). Delphi method was discussed in details in chapter one. In this section, Delphi process for extraction of qualitative risk

84 attributes that affect the lender-borrower lending relationship is discussed, and then we discuss how it could be used for estimation of qualitative risk.

3.1.1 Delphi Method Procedure

For the purpose of gathering attributes that are most important in the continuation of relationship lending with a business client of a bank in Iran, Delphi method was used. In other words we wanted to solicit information from banking experts on relationship lending qualitative risk factors. The following sections will discuss the steps of undertaken Delphi method along with the results of each step.

3.1.1.1 Choosing Experts

The first step of Delphi method is expert selection. Expert selection, has its own steps according to Okoli and Pawlowski (2004) which includes preparing the knowledge resource nomination worksheet (KRNW) by which we identify relevant organizations and academics or practitioners who could be on our experts list. The purpose of the Knowledge Resource Nomination Worksheet is to help categorize the experts before identifying them, in order to prevent overlooking any important class of experts. Then we have to populate our KRNW by filling out as many names as we can, then we nominate our additional experts by asking the experts to nominate other experts. After that we rank experts for each category and then start inviting them. The details of expert selection steps for Delphi method is shown in Figure 3.1.

According to Okoli & Pawlowski (2004), we should divide experts into panels. Their size and constitution depends on the nature of the research question and the dimensions along which the experts will probably vary.

We chose the experts for our Delphi panel from two main categories of risk and credit . The reason for choosing these two panels was the purpose of our research. The credit experts had to be chosen since the target customers of our research are credit facility applicants. The risk experts had to be on panel since we wanted to solicit qualitative factors that determine the risk of lending relationship.

85

• Identify relevant discipline or skills: academics, Step 1: practitioners, government officials of NGOs Prepare KRNW • Identify relevant organizations

• Identify relevant academic and practitioner literature Step 2: • Write in names of individuals in relevant disciplines or skills Populate KRNW • Write in names of individuals in relevant organizations with names • Write in names of individuals from academic and practitioner literature

Step 3: • Contact experts listed in KRNW Nominate • Ask contacts to nominate other experts additional experts

Step 4: • Create four sub-lists, one for each discipline

Rank experts • Categorize experts according to appropriate list • Rank experts within each list based on their qualifications

• Invite experts for each panel, with the panels

Step 5: corresponding to each discipline Invite experts • Invite experts in the order of their ranking within their discipline sub list

• Target size is 10-18 • Stop soliciting experts when each panel size is reached Figure 3.1. Flowchart of expert selection for Delphi Method adapted from Okoli & Pawlowski (2004) The experts were listed according to their work experience in banking sector and were nominated either by CEOs, head of branches, or managers of the six Iranian private banks. They were asked to introduce their risk and/or credit experts that they had in the bank or knew outside the bank. From 28 panelists of the two categories that we contacted, 23 accepted to participate and remained till the last questionnaire. Of these twenty three experts, seven were risk experts and the rest were credit experts.

The risk experts’ average age was 39; they had an average of fifteen years of working experience in financial services, and were graduated at B.S or M.S levels in one of the following majors: Economics, Finance, management or engineering. One of the risk panelists had PHD in finance and around 38 years of working experience in financial services. The general information of risk panelists is provided in Figure 3.2 and Figure 3.3.

86 Risk Experts' Backgrounds

70 60 50

40 Age 30 Work Experience 20 10 0 RE1 RE2 RE3 RE4 RE5 RE6 RE7

Figure 3.2. Risk experts’ age and work experience

Degree Distribution of Risk Experts

PHD, 14% BS, 28% PHD MS BS

MS, 57%

Figure 3.3. Risk experts’ educational background

On the other hand, the credit experts’ average age was 50; they were mostly experienced in banking with average of twenty six years of working experience. Their educational background was B.S in accounting, management, CS, and banking. Two of them had no official degree from any accredited college or institute past high school diploma. One of them was a masters student majoring in banking. The general information of credit panelists is provided in Figure 3.4 and Figure 3.5.

87 Credit Experts' Background

70

60

50

40 Age 30 Working Experience

20

10

0 CE1 CE3 CE5 CE7 CE9 CE11 CE13 CE15

Figure 3.4. Credit experts’ age and work experience

Diploma, 12%

MS, 6%

BS MS Diploma

BS, 83%

Figure 3.5. Credit experts’ educational background

3.1.1.2 Brainstorming the Experts

Next phase in Delphi is Data Collection and Analysis which includes brainstorming of the experts and asking them to list relevant factors. As a researcher, in this stage we should treat the experts as individuals regardless of which category they belong to. Two questionnaires will be distributed in this phase by which we gather initial factors and then validate them by giving back the original questionnaire of each expert plus other factors and asking them to confirm their own list.

88 In our first questionnaire, we asked panelists to bring up as many qualitative factors as they could that were, in their opinion and due to their experience, influential in the continuation of relationship with a relationship borrower. They were also asked to give a brief description of each factor they had mentioned, to help in categorization of the factors. Questionnaire 1 is brought in Appendix 3.1.

There were 26 factors gathered from this questionnaire by the panelists. These factors along with the brief description of each as mentioned by experts, is brought in Table 3.1.

Table 3.1. Initial list of experts’ factors including their reasons Risk Factors Summarized description by the experts Different economic sectors like agriculture, Economic sector of the firm construction, … have different risks Number of competitors More competitors bring more risk due to loss of job The less under governmental regulations, the less Impact of government risky since the governmental regulations are not regulations on firm's activities stable Growth rate of the firm's The higher the domestic growth, the less the risk industry Impact of inflation on firm's The higher the impact, the higher the risk activity Impact of imports on firm's The higher the impact, the higher the risk activity

Market share of the firm The higher the market share, the lower the risk

Performance of the firm in the The better the performance, the lower the risk banking system Relationship of the firm and The better the relationship the lower the risk its clients Bargain power of suppliers If the number and variety of suppliers is high, the

89 Risk Factors Summarized description by the experts risk is lower If the number and variety of buyers is high, the risk Bargain power of buyers is lower Extent of word of mouth the The more the WOM, the less the risk of losing client can bring customer The more the experience of the management in the Management quality of the field, the higher the related education, the lower the firm risk Sales fluctuations of the firm The higher the sales fluctuations, the higher the risk Ratio of variable cost/ fixed If high, means technology usage and/or internal cost management is weak, so the risk is high Liquidation capability and type of promissory Type/amount of collateral and notes, stocks, residential property, account reliability of cosigners receivable as collateral reduces the risk The higher the compatibility of permits with Activity permits of the firm activity, the lower the risk Production capacity of the The higher the volume of production, the lower the Firm risk Five options identified from low risk to high risk: Reliability of the firm's Audited financial statements, un-audited financial financial statements statements, tax statements, balance sheet only, no financial statement Credit facilities’ usage If the usage is beneficial and matches the firm’s purpose activities, the risk is lower The longer you know the customer, the lower the Longevity of the relationship information asymmetry, the lower the risk Reliability of the referee The more reliable the referee, the lower the risk Number of bought services by The higher the number of bought products, the the firm lower the risk

90 Risk Factors Summarized description by the experts Concentration of checking account activities in the Extent of checking account bank means more reliability and control so lowers activity within the Bank the risk Extent of credit activities Concentration of credit activities in the bank means within the Bank more reliability and control so lowers the risk Growth rate of the firm's If the growth is high and matches the firm’s activity credit activity with the bank growth, the risk is lower

In the second questionnaire we categorized the 26 extracted factors of the first questionnaire’s results into three groups of internal factors of the firm, environmental factors, and firm’s relationship with the lending bank and mentioned reasons for selection of the factors by experts (if applicable) and sent the categorized factors, along with a copy of the expert’s response to first questionnaire, to each expert. We then asked experts to verify the interpretation of their reasons, relevance of attributes to the research question and also to verify the categorization of their factors. According to (Schmidt, 1997), ‘‘without this step, there is no basis to claim that a valid, consolidated list has been produced.’’ Questionnaire 2 is shown in Appendix 3.2.

The criteria we set for omitting a factor in this stage was that if at least 80% of the experts mentioned a factor should be omitted, the factor was omitted. The categorization was approved by all our experts, but their own lists had major changes in some cases such as adding the factors from the consolidated list to their own list and mentioning that they had forgotten to bring up those factors in their first questionnaire. Also three of the factors were identified as “influential factors in commencement of relationship lending with a new customer” by 19 experts (82%). These factors were 1) relationship of the firm with its clients, 2) the performance of the firm in the banking system, and 3) the credibility of the referee of that firm to the lending firm, and were all omitted from the list. The other omitted factor was extent of word of mouth the firm could bring for the financial service provider which was described as “non-related to the subject” or “value creating indicator” by 22 experts (95%) and so was omitted from the list.

91 So the output of this phase of our Delphi was a verified categorized list of 22 factors. These factors are shown in Table 3.2.

Table 3.2. Categorized list of attributes Risk Factors Environmental Factors 1. Economic sector of the firm 2. Number of competitors 3. Impact of government regulations on firm's activities 4. Growth rate of the firm's industry 5. Impact of inflation on firm's activity 6. Impact of imports on firm's activity 7. Market share of the firm 8. Bargain power of suppliers 9. Bargain power of buyers Internal Factors 1. Management quality of the firm 2. Sales fluctuations of the firm 3. Ratio of variable cost/ fixed Cost 4. Type of collateral and reliability of cosigners 5. Activity permits of the firm 6. Production capacity of the Firm 7. Reliability of the firm's financial statements 8. Credit facilities’ usage purpose Lender- borrower Relationship 1. Longevity of the relationship 2. Number of bought services by the firm 3. Extent of checking account activity within the Bank 4. Extent of credit activities within the Bank 5. Growth rate of the firm's credit activity with the bank

92 3.1.1.3 Narrowing Down Factors

In the second phase that narrows down the lists of factors, the goal will be to understand the rating of importance of the factors based on the differing perspectives of various stakeholder groups.

So, in our third questionnaire we asked experts to identify (and not rank) at least 10 factors (from all three categories) that they thought were the most important factors in continuation of relationship lending. Questionnaire 3 is brought in Appendix 3.3. We repeated this step for two consecutive iterations and the final attributes were those which met the following criteria:

1) Attributes with more than 80% of total votes (>= 19 votes) in a single round were selected

2) Attributes with more than 85% of the votes from a panel that have also got more than 50% of the other panel’s vote in a single round, were selected

3) Attributes with >= 50% of total votes in both rounds that have also received >= 50% of each panel’s votes in a single round were selected

There are no firm rules for establishing when a consensus is reached but the reliability of Delphi results increase with the size of the group and the number of rounds (Fink, Kosecoff, Chassin, & Brook, 1984).

We got total of 13 attributes at this stage. The reason for omitting some attributes is that we want to find the attributes that experts have consensus on their importance, while we know that the number of risk experts is less than half the number of credit experts. Figure 3.6 and Figure 3.7 show the number of the votes to each factor in each round.

93 Total Votes Vs. CE's and RE's Votes Round 1

25 20 15 10 5 Number of Vote Numberof 0 Sales Market Type of Total Round1 Inflation Checking Degree of Degree Economic Number of Number Production RE Round1 Longevity of Growth Rate Growth Factor CE Round1

Figure 3.6. Factors’ votes of first round

Total Votes Vs. CE's and RE's Votes Round 2

25 20 15 10 5

Numbe of Vote 0 Sales Market Type of Total Round2 Inflation Checking Degree Degree of Economic Number of Production RE Round2 Longevity of Growth Rate Factor CE Round2

Figure 3.7. Factors’ votes of second round

94 We got total of 13 attributes that met our criteria of consensus in two rounds. The output of the two rounds are brought up in Appendix 3.4. So the finalized list of experts’ attributes of this phase was determined and is shown in Table 3.3.

Table 3.3. Extracted Delphi attributes that met the consensus criteria Environmental factors

1. Competence capability

2. Degree of deregulation

3. Independency from imports

4. Client’s market share

5. Domestic growth of firm’s industry

6. Number of buyers

7. Number of suppliers

Internal factors

8. Management quality of the firm

9. Type of collateral and credibility of cosigners

10. Firm’s production/sales capacity

11. Reliability of firm’s financial statements

Lender-borrower Relationship

12. Longevity of relationship with lender

13. Firm’s checking account activity within the bank

Comparing the results of Table 3.3 to the total factors of Table 3.2 we can see that from the nine environmental factors, “economic sector of the firm” and “impact of imports” do not seem to have as many votes as required, hence omitted from the category. In the internal factors category three factors which were the “impact of inflation”, “variable cost/ fixed cost ratio”, “Credit facility usage purpose”, and “sales fluctuations” were omitted. In the lender-borrower relationship category also three factors

95 were omitted which include “number of bought services”, “extent of credit activities within the bank”, and “growth rate of credit activities with the bank”.

3.1.1.4 Grading Factors

The last step of a ranking-type Delphi method according to (Okoli and Pawlowski, 2004) is ranking the factors that experts of each category have brought up. In our research, since the purpose of attribute solicitation from the experts would be its application in scoring customers, the ranking of the attributes wouldn't be of any use while the degree of importance of each attribute is essential to be known. So in our last questionnaire (questionnaire 4) we asked the experts to indicate the importance of each extracted factor. In order to be able to give a grade according to the importance, we considered the scale of 1 (indicating very poor) through 7 (very strong) and we set the criteria of consensus to be average weight of more than or equal to 4. The mean weight of each attribute was then calculated and the final results are shown in Table 3.4.

Table 3.4. Final extracted attributes of Delphi process and their mean weights

Factor Mean weight 1.Competence capability 4.35 2. Degree of deregulation 5.13 3. Independency from imports 5.17 4.Client’s market share 5.74 5. Domestic growth of firm’s industry 5.13 6. Number of buyers 5.83 7. Number of suppliers 4.61 8. Management quality of the firm 6.39 9.Type of collateral and credibility of cosigners 5.13 10.Firm’s production/sales capacity 5 11.Reliability of firm’s financial statements 6.13 12.Longevity of relationship with bank 5.83 13.Firm’s account activity within the bank 6.09

96 Questionnaire 4 is shown in Appendix 3.5 and the results are brought in Appendix 3.6.

3.1.2 Qualitative Risk Score Estimation

For the purpose of estimating risk score on the basis of qualitative factors, the extracted attributes should be given to the lending company and then the customers being graded regarding these attributes. The multiplication of the weight of each attribute by the given grade will be the score of that firm in that attribute. Summing up the total scores for each firm derives the qualitative risk score of the firm.

The flowchart of qualitative risk score modeling would then look like Figure 3.2.

Selection of the Delphi methodology

Delphi method steps Questionnaire 1: Initial collection Selecting experts of factors Brainstorming for important factors Questionnaire 2: Validation of categorized list of factors

Narrowing down factors Questionnaire 3: Choosing most important factors

Assigning weight to factors Questionnaire 4: Grading the extracted attributes factors

Grading business customers according to Delphi output

Finding qualitative risk score for business customers

Figure 3.8. Flowchart of qualitative risk score modeling

3.1.3 Quantitative Risk Score Estimation

There are different risks concerned with financial services’ industry including liquidity risk, interest rate risk, and the credit risk. Among the different types of risks,

97 credit risk scoring is among the most important ones so that when the bank for instance is faced with a major financial crisis, it is due to excessive credit risk because of defaults of the customers for paying back the loans.

Nonlinear regression models are used where the response outcomes are discrete and error terms are normally distributed like the case where the response variable is qualitative with two possible outcomes like the financial status of a firm (sound status, headed toward insolvency) (Kutner et al., 2005) .

In our research, we also need to estimate the quantitative risk score of customers to bring it into our final model. In order to do so, we use logistic regression by which the probability of default of customers is estimated according to their financial records.

3.2 Probability of Default Modeling

For the purpose of modeling the real probability of default (PD) of a customer according to their risk, we should put customers into different classes according to their risk scores and then check to see in each class how many customers have past dues, or more precisely have defaults. This way the probability of default of each class of customer is obtained having their risk scores.

3.3 Revenue Modeling

The revenue of a loan applicant could roughly be calculated from the following equation:

(1-PD)(L* rL) (3.1) In this equation the probability of default (PD) is obtained based on total risk score that we have estimated for each customer and so “1-PD” is the probability that the customer pays back the amount with no default of more than 3 months. “L” is the amount of loan and “ rL” is the loan interest rate.

98 3.3.1 Repurchase Rate Estimation

In Iran the switching cost of business customers who apply for credit facilities is very high since the number of banks who would provide customers with these facilities is very limited and so is convincing these banks to lend required amount to customers. Therefore the assumption is that customer remains with the lender unless s/he defaults.

We have modeled the probability of default of the customer as the risk of relationship, so the repurchase probability is the next important factor to be estimated for the future revenue generation estimation.

3.4 Cost Modeling

The cost of each customer is consisted of cost of money and the administration cost. The total cost would roughly then be estimated by (L* rC) + C A where rC is the average cost of money and C A is the administration cost which includes the cost of marketing for acquisition and retention of customers, cost of employees, and other administrative costs.

3.5 WACC Modeling

WACC in the financial services is equivalent of the cost of equity since the cost of money is calculated in cost estimations and the shareholder value is considered in cost of equity. The cost of equity is then estimated from R e = (D 1/P 0) + g (Dermine and Bissada, 2002).

The current stock price, P0, is known and the dividend of the next period, D1, can be predicted. The growth rate which is usually the challenge, could be estimated from the relationship between the growth rate, the retention rate, and the return on equity where the relationship is shown by: g = (1-D/EPS) ROE in which D/EPS represents the assumed stable dividend payout ratio and ROE is the historical return on equity. So 1-D/EPS is the company’s earnings retention rate (Dermine and Bissada, 2002).

99 3.6 CLV Estimation Model

According to what is said, our final estimation model would look like equation 3.2 bellow:

Ni Ni −ti CLV i =ƒ[](rL(ti ) −rc(ti ) ) −Ca(ti ) Lti 1( +WACC ) + ti&=0)))))))) ')))))))) ( Term 1

Ni +Mi ([]1( − PD )( r −r ) −C L (Rate )( 1+ g)Ni +Mi −ti ) ƒ i L(ti ) c(ti ) a(ti ) ti repuri 1( +WACC )ti −(Ni +Mi ) ti&=Ni)+1))))))))))) ')))))))))))) ( Term 2 (3.2)

In this equation, “N i” is the number of years that customer i has had relationship with the service provider so far. So the first term in this equation is the current value of each customer being calculated from the financial records of the customers. The second term of the equation which is the future value of customers is consisted of base value (up-selling potential) and retention value (loyalty) and both are discounted to present. In term2 “g” is the growth rate of customer’s request for credit facilities which could be estimated according to historical behavior of customers, “C a” is the administration cost for acquiring and retaining customers, “r c” is cost of money, “Ni+Mi” is the lifetime of customer i, and “WACC” is the discount rate which is substituted by the cost of equity as explained in section 3.5.

The assumptions of the above formula are:

• The discount rate or WACC is fixed over lifetime of customer • PD is fixed over remaining relationship time

In our mathematical modeling we have actually considered two value dimensions defined in literature as up-selling dimension (current value) and loyalty dimension

100 (retention value). There is another commonly used dimension related to cross-selling potential (potential value) of customers, but as mentioned in chapter 1, we focus on most profitable activities of financial service providers which are providing credit facilities and these facilities do not have much variety to be cross-sold, so potential value is not brought up in our CLV model.

3.7 Summary of the Chapter

This chapter basically brought up the details of undertaken steps for modeling qualitative risk score of business customers, and then a basic mathematical model that could be used for CLV estimation in financial services industry was discussed. The methodology for extraction of qualitative attributes was Delphi which led to extraction of thirteen attributes. These attributes will then be used for grading customers of a financial service provider and that grade would be the qualitative risk score of that customer. The quantitative risk score is most commonly determined by logistic regression method as is offered for this research too. What is then needed in calculation of CLV is the probability of default of customers which we propose to be considered for future revenue estimation. This probability could be determined by classifying customers regarding their risk scores and then by checking the number of real defaults in each class, the real PD is calculated.

The mathematical model is consisted of current value and predicted future value. By current value we calculated the value of customer from the beginning of its relationship with the bank, and for the future value we estimated the up-selling potential and retention value of customers for their remaining lifetime. The potential value of customers as defined in literature is not considered in this model since the focus of this research is on estimation of CLV for credit facility applicants and cross-selling is very limited for this stream of business customers.

101

Chapter 4

Model Fine-tuning

4. Model Fine-tuning

In this chapter we will go through the tasks undertaken to fine-tune our developed model of chapter 3. As mentioned in chapter 1, the case that we have chosen is Karafarin bank (KB) which is a financial service provider. The financial records of this bank’s 101 business customers are considered between year 2001 and 2008. In this chapter the steps for the case study are brought up in details and the data analysis for each step of CLV measurement of these 101 customers is explained.

102 4.1 Qualitative Risk Score Computations

For the purpose of gaining a score for the qualitative risk, we gave the list of extracted Delphi attributes (13 attributes) to credit committee of KB (four members) and asked them to grade a list of 101 firms on basis of each solicited attribute, again on the scale of 1 (very low) through 7 (very high). From the members of the committee only one member was completely familiar with the firms that we had chosen, so the grades were assigned by him. He was included in our Delphi panel and so did not entail any objection to the extracted Delphi attributes. For 25 firms, the head of the branch by which the credit facility was granted also graded the firms and the average of the two opinions was considered the final grade.

The 101 firms were selected on the basis of having relationship lending with KB. The reason for this selection criterion was that since we weren’t going to consider the switching cost in our modeling, we wanted to increase the probability of having retained business customers.

Scores of each firm was calculated by sum-product of mean grades of each attribute by the mean weight of it then divided by the maximum possible score which is the sum-product of maximum possible grade of each attribute and its weight. For instance Customer #1 has got the following grades for its 13 attributes: A1=1, A2=4, A3=7, A4=4, A5=2, A6=7, A7=5, A8=3, A9=5, A10=4, A11=3, A12=7, A13=3. Then each of these grades is multiplied by its corresponding mean weight indicated in Table 3.4 and then added together which yields 302.67 for this customer. The results are shown in Table 4.1. The “performance” column of Table 4.1 indicates the performance of each firm where 0 means past dues of less than three months, and 1 means past dues of more than three months. The performances of the firms are filled out from their files at KB.

The scores are then divided by 481.45 which is the sum-product of mean weight of each attribute and their maximum possible value to scale them all down to a figure between 0 and 1 which can then be compatible with financial scores of this range. So for instance 302.67/481.45 would yield 0.6286 which is the standardized score of this

103 customer. The closer the score is to 1, the lower the qualitative risk of relationship with that firm. The results are shown in Appendix 4.1.

Table 4.1. Qualitative risk scores and performances

Customer ID Qual Score Performance Customer ID Qual Score Performance 1. 302.67 0 51 241.02 0 2. 298.68 0 52 256.97 0 3. 323.08 0 53 294.65 0 4. 283.12 0 54 378.37 0 5. 349.41 0 55 279.85 0 6. 296.43 0 56 277.78 0 7. 302.20 0 57 338.74 0 8. 305.88 0 58 362.79 0 9. 318.01 0 59 341.81 0 10. 297.57 0 60 282.05 0 11. 287.95 0 61 346.67 0 12. 328.50 0 62 345.47 0 13. 329.32 0 63 297.34 0 14. 263.12 0 64 327.46 0 15. 369.57 0 65 332.80 0 16. 308.29 0 66 233.25 1 17. 282.41 0 67 292.25 1 18. 312.45 0 68 234.67 1 19. 298.79 0 69 238.73 1 20. 363.77 0 70 242.27 1 21. 355.97 0 71 229.28 1 22. 295.45 0 72 233.89 1 23. 349.50 0 73 256.37 1 24. 369.79 0 74 245.41 1 25. 267.97 0 75 277.03 1 26. 264.91 0 76 227.54 1 27. 290.18 0 77 211.38 1 28. 312.54 0 78 202.88 1 29. 293.43 0 79 246.03 1 30. 321.37 0 80 250.67 1 31. 325.42 0 81 242.95 1

104 Customer ID Qual Score Performance Customer ID Qual Score Performance 32. 289.14 0 82 209.96 1 33. 308.20 0 83 240.93 1 34. 323.44 0 84 238.54 1 35. 322.55 0 85 230.96 1 36. 312.44 0 86 225.51 1 37. 356.95 0 87 230.07 1 38. 239.15 0 88 243.31 1 39. 273.12 0 89 231.78 1 40. 251.58 0 90 200.43 1 41. 227.37 0 91 235.84 1 42. 342.33 0 92 252.24 1 43. 328.07 0 93 167.13 1 44. 289.09 0 94 247.13 1 45. 359.07 0 95 226.01 1 46. 341.53 0 96 222.77 1 47. 335.04 0 97 241.68 1 48. 291.40 0 98 231.99 1 49. 249.34 0 99 235.37 1 50. 255.85 0 100 212.13 1 101 233.25 1

The normality of the scores was tested by Chi-square goodness of fit test for both risky and reliable population groups.

H = 0 p = 0.2109 stats = chi2stat: 7.1330 df: 5 edges: [1x9 double] O: [7 3 6 15 6 11 7 11] E: [5.0095 5.1745 7.8484 10.0448 10.8482 9.8862 7.6024 9.5860] [H,p,stats]=chi2gof(Y) H = 0 p = 0.0524 stats = chi2stat: 5.8967 df: 2 edges: [167.1300 217.1780 229.6900 242.2020 254.7140 292.2500]

105 O: [6 5 13 8 3] E: [7.6095 7.1371 7.9860 6.4255 5.8419]

According to Chi 2 test above in Matlab, reliable customers (66 members) got the 99.5% confidence level to be a normal population, while the risky customers (35 members) got the 95% confidence level which could be due to smaller sample size of this group. The mean and STD for risky customers was respectively 233.95 and 21.48 whereas for reliable customers they were 309.71 and 36.38. Considering both populations have normal distributions, we tested for the following sets of hypotheses to validate our results:

A) H0: the two population means are equal

H1: the two population means are significantly different

B) H0: the two population variances are equal

H1: the two population variances are significantly different

To test the mentioned hypotheses we ran the T-test for set A and F-test for set B. For the T-test, the degrees of freedom yields 99 (rounded) and since we have a two-tailed test, t(99,0.025) = 1.984. The value for t yields 2.28 having the STD and sample sizes. The acceptance region would be -1.984 < t < 1.984, so we reject the null hypotheses of set A with 95% confidence and would accept H1 which suggests the means are significantly different for the two populations. The MATLAB output of this T-test was:

[H,p,ci,stats]=ttest2(X,Y,0.05,'both') H = 1 p = 1.6174e-019 ci = 62.4587 89.0581 stats = tstat: 11.3026 df: 99 sd: 32.0551

Since H=1, we reject the null hypotheses.

For the F- test, the degrees of freedom are 64 and 34 respectively for reliable and 2 risky customers. On the other hand F 65,34 = [STD(reliable) / STD (risky)] = 2.86 and since the critical value for right-hand tail area equal to 0.005 is 2.30, we reject the null

106 hypotheses with 99.5% confidence which means we agree that the variances of the two populations are significantly different. The MATLAB output for the F-test was as follows:

[H,p]=vartest2(X,Y,0.05,'both') H = 1 p = 0.0012 ci = 1.5390 5.0563 stats = fstat: 2.8685 df1: 65 df2: 34

Again since H=1, we reject that variances are the same.

For each group, we consider 95% area under the normal distribution which is the µ±2STD. For the reliable customers, this range is 237 till 382 vs. this range for the risky customers which is 191 till 277. So the range of scores between 237 till 277 is where the two groups’ scores overlap and in our sample we have 20% of the customers in this range.

To further analyze the results, we got the score for difference of averages of each attribute for reliable customers and risky customers to indicate the most important differentiators of these two groups. The results are shown in Table 4.2. We found out that A8, and A9 are respectively the most influential factors that cause the difference between risky and reliable customers. These factors according to Table 3.2 are “management quality of the firm” and “type of collateral”. The former, A8, has also got the highest importance weight from our Delphi experts. So although experts have verified this attribute as an important factor, it is still the most influential discrepancy factor at KB. This could imply that KB would better set more lucid indicators for A8 so that the committee can more accurately evaluate this characteristic of the firm. The latter, A9, which came out as a significant discrepancy factor in our research, is also emphasized in literature by Jimenz et al., (2006) as a way to decrease the risk of relationship lending.

Table 4.2. Strength of the differentiators Factor Reliables Riskies Dif Weight Strength A1 3.11 2.63 0.48 6.39 2.08 A2 4.2 4.09 0.11 5.13 0.57

107 Factor Reliables Riskies Dif Weight Strength A3 4.15 4.14 0.01 5 0.04 A4 3.73 2.63 1.1 6.13 6.31 A5 4.11 2.76 1.34 6.39 6.9 A6 5.08 4.21 0.87 5.13 5.05 A7 4.59 4.37 0.22 5 1.01 A8 6.02 2.96 3.05 6.13 19.5 A9 5.65 3.45 2.2 6.39 11.3 A10 4.56 3.23 1.33 5.13 6.66 A11 3.45 3.09 0.37 5 2.26 A12 5.14 4.43 0.71 6.13 4.13 A13 3.06 1.43 1.63 6.39 9.94

On the other hand, A3 and A2 respectively got the lowest strengths as discrepancy factors. A3, “independency from imports”, could have been a discrepancy factor because the political restrictions for imports in Iran would cause firms to face roughly the same conditions for this attribute. A4, “degree of deregulation”, could also be justified as government not only heavily regulates firms in Iran, but also holds ownership of large firms in many industries and this could cause firms not to differ much regarding this attribute. The “reliability of the firm’s financial statements” that has got high importance weight according to experts of our Delphi panels, is the fifth lowest discrepancy factor. This attunes the fact that financial statements are not trustworthy enough in Iran to be used for risk scoring of customers.

4.2 Quantitative Risk Score Computations

For the purpose of having the quantitative risk score into consideration, we ran the logit regression for 500 business customers of KB including our 101 customers. The 1 quantitative ratios in 1+ e−ωχ are identified at KB and what we did was running the logit regression in STATA and getting the quantitative scores.

Some of the variables used for quantitative modeling of risk score at KB are as follows:

108 company’s economic sector (construction, governmental or non-governmental, mining and industrial), company’s year in business minus loan start time, opening account time minus loan start time, number of loans given in a year, years of experience of company’s CEO, no default, default, debt to asset, equity to asset, long-term debt to asset, banking system debt to asset, banking system debt, banking system debt to total debt, long-term debt to equity, long-term debt, liquidity ratio, etc.

Among these variables, some are similar to the factors that we had in our Delphi categorized list if considered from quantitative aspect. For instance “the years of experience of the company’s CEO” together with “company’s year in business minus loan start time” are quantitative indicators of “management quality of the firm” that we had in our Delphi’s final 13 attribute. Even “number of loans given in a year” is comparable to “number of bought products” factor that we had in our initial Delphi list.

Although our Delphi experts didn’t come to any consensus on “company’s economic sector” and “number of bought products” factors, this repetition confirms the importance of these risk factors and adds to the value of considering both qualitative and quantitative risk scores as proposed for this research.

The scores are shown in Appendix 4.2. Again the closer the score is to one, the lower the risk of that customer.

4.3 Real PD Calculations

After we estimate the quantitative and qualitative scores, we need to calculate the real probability of default for customers to be embedded in our CLV mathematical model. The distribution of 101 customers regarding their quantitative and qualitative scores is shown in Figure 4.1.

109 Risk Scores' Distribution

90

80

70

60

50

Qual Risk Score % Score Risk Qual 40

30 0 20 40 60 80 100 Quant Risk Score %

Figure 4.1. Risk scores’ distribution of customers regarding quantitative and qualitative scores

In order to find the probability of default, we had to put our customers into different classes. So first we divided quantitative and qualitative scores into two groups of above and below 0.55. This division then yielded a matrix with four quarters. This cutoff point (0.55) was estimated in excel with trial and error to see for which segmentation we get better results.

We then counted the number of customers in each quarter and by having the performance codes (default= 1, no default = 0) for customers of each quarter, we calculated the probability of default for each quarter. For instance we had 12 customers who had both qualitative and quantitative scores below 0.55 and from these 12 customers all 12 of them have default code of 0. So the probability of default for this quarter is 12/12= 1 or in other words the customers in this quarter will have defaults with 100% probability. On the other hand those with qualitative and quantitative scores above 0.55 will have PD of 2.17% which is a reasonable result. The result for PD of customers for each identified quadrant is brought in Table 4.3.

110 Table 4.3. Probability of default of customers for each quadrant Quantitative score Qualitative score Below 0.55 Above 0.55

Below 0.55 12/12 = 100% 21/29 = 72.41%

Above 0.55 1/14 = 7.14% 1/46 = 2.17%

In the above table, PDs are calculated for each quadrant of the matrix so now we can have the PDs assigned to each of our 101 customers. The PD of customers is shown in Table 4.4.

Table 4.4. Probability of Default of each customer Customer Customer Customer Real Real PD Real PD ID ID ID PD 1 0.02  0.02  0.72 2 0.02  0.02  0.99 3 0.02  0.02  0.72 4 0.02  0.72  0.72 5 0.72  0.02  0.72 6 0.02  0.72  0.72 7 0.02  0.02  0.07 8 0.02  0.02  0.02 9 0.02  0.02  0.02 10 0.02 0.02  0.99 11 0.02  0.02  0.72 12 0.02  0.02  0.07 13 0.02  0.72  0.72 14 0.07  0.02  0.72 15 0.02  0.99  0.02 16 0.72  0.72  0.02 17 0.72  0.02  0.72 18 0.72  0.72  0.99 19 0.02  0.02  0.02 20 0.02  0.99  0.07 21 0.02  0.72  0.02 22 0.07  0.72  0.72 23 0.02  0.99  0.99 24 0.02  0.72  0.99

111 Customer Customer Customer Real Real PD Real PD ID ID ID PD 25 0.02  0.72  0.72  0.02  0.02  0.99  0.02  0.02  0.72  0.02  0.07  0.07  0.72  0.99  0.07  0.72  0.02  0.07  0.72  0.07  0.07  0.07  0.99  0.72  0.72  0.02  0.02 34 0.07  

The real probabilities of default of Table 4.4 indicate the probability that each customer would not pay back its loan amount. For instance Customer#1 has 2% probability of default which means it is 98% possible that this customer pays back its loan amount on time in future, while Customer#57 has 99% probability of default which means it is only 1% possibility of repayment from this customer. These PDs are much more reliable than PDs based solely on financial score of customers because as mentioned in chapter 1, the financial statements are not reliable in Iran to be used for risk scoring of customers.

4.4 Lifetime Estimation

Lifetime of customers was estimated for different economic sectors of customers at KB. We selected six economic sectors of business, services, governmental and non- governmental institutes, construction, mining and industry, and agriculture. For each sector, we calculated the establishment year of the company till now and then averaged it for the whole customers. The results are shown in table 4.5.

Table 4.5. Average lifetime of different economic sectors Number of Firms Average Economic Sector Active in this Sector Lifetime Business 326 10.66 Services 115 10.69 Governmental & non governmental institute 17 8.125 Construction 620 19.17

112 Number of Firms Average Economic Sector Active in this Sector Lifetime Mining & Industry 642 14.9 Agriculture 29 15.86

Since the Business and Services lifetimes are pretty close, we consider them the same for lifetime substitution in CLV model. This case is also true for Mining & Industry and Agriculture sectors. The rounded estimations for lifetimes would then be as shown in Table 4.6.

Table 4.6. Rounded estimates of lifetimes

Economic Sector Average Lifetime Business and Services 11 Governmental &non governmental institutes 8 Construction 19 Mining & Industry and Agriculture 15

4.5 Loan Demand Growth Estimation

In order to roughly estimate the growth rate of loan demand, we considered customers between years 2004 and 2007 who had applied for loans for four consecutive years. We then took the geometric average of each customer’s four-year loan demand and then we took the arithmetic average of the whole results which ranged from negative amounts to 110% in some recent cases. The geometric average is determined from following formula:

1/3 Gi =[(1+g1)(1+g2)(1+g3)] where g= Ln[this year/last year] and i is the customer.

The results are shown in Appendix 4.1.

The arithmetic average of these G is then yielded 29% to be used in our model which is a reasonable answer because the inflation is around 27% so the demand in average should at least grow by 27%.

113 4.6 Repurchase Rate Estimation

In order to estimate the repurchase rate of customers, we counted how many times a firm has been granted a loan during its past relationship with the bank and divided it by the total number of past relationship years.

So for substitution in our CLV’s second term (future value) we would use the repurchase rate of each customer. The repurchase rates are shown in Appendix 4.3.

4.6 WACC, Cost of Money, and Administration Cost Estimations

As explained in section 3.5, the discount rate would preferably be replaced by cost of equity which is determined by the financial calculations. At KB, the WACC is estimated as 30% by the formula mentioned in section 3.5.

The Average Cost of Money and Administration Cost are also estimated at KB and the amounts are r c = 16% of loan amount and C a = 2% of loan amount.

Administration cost of 2% for our unit of measurement which is a B2B lending relationship means for every 100 Rials granted as a loan, 2 Rials is spent on employees, expert evaluations of collateral, maintenance, etc. So in our CLV model we will subtract 2% of each loan amount as its administrative cost.

4.7 CLV Calculations

In chapter 3 we discussed the following possible CLV model to be used for borrowers of a commercial bank. This model was stated as

114 Ni Ni −ti CLV i =ƒ[](rL(ti ) −rc(ti ) ) −Ca(ti ) Lti 1( +WACC ) + ti&=0)))))))) ')))))))) ( Term 1

Ni +Mi ([]1( − PD )( r −r ) −C L (Rate )( 1+ g)Ni +Mi −ti ) ƒ i L(ti ) c(ti ) a(ti ) ti repuri 1( +WACC )ti −(Ni +Mi ) ti&=Ni)+1))))))))))) ')))))))))))) ( Term 2 in which, Term1 was the current value of the borrower and Term2 was the future value of them consisted of retention value and up-selling potential. For the current value, the lifetime is considered from the date of first granted loan to that firm till now (2007). For the future value, the lifetime is considered from now till the end of estimated lifetime of the sector that the firm belongs to. The calculations for future value are done in MATLAB and the program is attached as Appendix 4.5. The results are shown in Table 4.7.

As we can see, some future values have negative amounts. This is true for customers with high PDs who carry both qualitative and quantitative scores of less than 55% or bear low qualitative score (< 55%) and high quantitative score (>55%).

Table 4. 7. Current value and future value of customers Customer Customer Current Val Future Val Current Val Future Val ID ID 1 1.0901E+10 1.659E+10 51 5120334783 -4.46E+10 2 2145848263 8.063E+09 52 4149600000 -1.394E+10 3 3.0803E+10 4.419E+09 53 6857809920 6.922E+09 4 5747049448 664717680 54 1.8295E+10 6.358E+10 5 1.0183E+10 1.497E+10 55 6071574671 1.07E+10 6 2.0349E+10 2.489E+10 56 5914556279 1.567E+10 7 4886509686 8.071E+09 57 2.4461E+10 9.724E+10 8 2.784E+10 3.306E+10 58 5.2499E+10 8.644E+09 9 1.0338E+10 4.787E+09 59 1.7232E+10 2.674E+10 10 5497034400 1.234E+10 60 5.8505E+10 1.004E+11

115 Customer Customer Current Val Future Val Current Val Future Val ID ID 11 3120102480 3.291E+09 61 523024700 1.047E+10 12 2899427037 346698000 62 6.0655E+10 3.161E+11 13 2.3268E+10 1.01E+10 63 9568000000 2.423E+10 14 1.3752E+10 -2.146E+09 64 1.677E+10 6.922E+10 15 2.5977E+10 4.608E+10 65 1.04E+10 1.174E+11 16 3926054868 5.61E+09 66 716126341 1.122E+10 17 2429354743 4.862E+09 67 1.3635E+10 515876900 18 1.1574E+10 1.597E+10 68 8243046710 -2.984E+10 19 3.2138E+10 5.049E+10 69 2.0052E+10 -3.705E+10 20 2806669216 3.366E+09 70 4.2212E+10 -7.504E+10 21 1.4391E+10 1.795E+10 71 9449235484 -6.65E+09 22 1.999E+10 1.975E+10 72 7024452240 -1.367E+10 23 3.3695E+10 2.693E+10 73 3335673046 -9.416E+09 24 2.9007E+10 1.077E+10 74 1365000000 -6.754E+09 25 2.2903E+10 2.036E+10 75 1.7875E+10 1.172E+10 26 7127400842 3.74E+09 76 1.5135E+10 -1.503E+10 27 3.2215E+10 9.724E+10 77 9328800000 -2.251E+10 28 5.8835E+10 1.638E+11 78 2666757600 -2.701E+09 29 3.4245E+10 5.545E+10 79 2.2832E+10 -5.982E+10 30 3073270997 2.468E+09 80 1712289800 -7.316E+09 31 2498143353 1.307E+09 81 7891623947 -1.73E+10 32 8207157726 9.556E+09 82 5109251526 -8.844E+09 33 8664255434 6.732E+09 83 1.5964E+10 -5.189E+10 34 4535190400 4.189E+09 84 1.1779E+10 -3.473E+10 35 6005429689 1.578E+09 85 9889148438 -2.303E+10 36 2.2439E+10 2.057E+10 86 1.9848E+10 -1.73E+11 37 2.3963E+10 3.291E+10 87 3.5723E+10 -1.367E+11 38 3.1338E+10 -3.903E+10 88 3.5723E+10 -2.262E+11 39 4118460667 1.622E+09 89 1.9311E+10 -7.68E+10 40 6297587881 -4.631E+10 90 9210257333 -4.824E+10 41 828199999 -3.773E+09 91 1.1174E+10 -6.271E+10 42 2.3886E+10 2.394E+10 92 4843541950 -2.09E+10 43 1.196E+10 4.712E+10 93 7960170400 -6.312E+10 44 4.879E+10 5.117E+10 94 2.0748E+10 -1.153E+11 45 2.5057E+10 8.677E+10 95 9328800000 -2.894E+10

116 Customer Customer Current Val Future Val Current Val Future Val ID ID 46 3.0461E+10 4.384E+10 96 1.9829E+10 -2.262E+11 47 1.5874E+10 4.264E+10 97 1.5439E+10 -1.763E+11 48 5.2061E+10 4.249E+10 98 2392000000 -7.504E+09 49 2.5385E+10 -6.895E+09 99 1924000000 -2.98E+10 50 2.4536E+10 -2.573E+10 100 4.0858E+10 -2.814E+11 101 5200000000 -3.484E+10

These segments’ PDs are 0.99 and 0.72 respectively, so when we substitute them in term2 of Equation 3.1, the total subtraction of the amount from cost of loan yields a negative amount. This means the higher the risk of customer, the higher the PD and the lower the customer value. Total value of each customer is shown in Table 4.8.

Table 4.8. Total CLVs of customers Customer Total CLV Customer Total CLV Customer Total CLV ID ID ID 1  9!    "9! 2 9!  9!  "9! 3 9!  9!   4    "9!  " 9! 5 9!      "9! 6  9!  " 9!  "9! 7 9!  " 9!  9! 8 9!  9!  

9 9!   9!  "9!

10 9! 9!  "  

11     9!  "9!

12      9!  " 9!

117 Customer Total CLV Customer Total CLV Customer Total CLV ID ID ID

13 9!  9!  " 9!

14 9!    9!  "9!

15  9!   9!  "9!

16    "9!  "9!

17     " 9!  " 9!

18  9!  "9!  "9!

19 9!  9!  "9!

20   9!  "9!

21  9!  9!  " 9!

22 9!  9!  "9!

23 9!  9!  " 9!

24 9!   9!  "9!

25 9!  9!  "9!

 9!  9!  " 9!

  9!  9!  "9!

 9!  9!  " 9!

 9!   9!  "9!

    9!  "9!

    9!  "9!

 9!  9!  " 9!

 9!   9!  " 9!

34    "9!

118 As could be seen in Table 4.8, from the first 66 customers who have originally been introduced as reliable customers, six customers have got negative CLVs (9%) while in the rest of 35 customers that we named risky customers, four customers have positive CLVs (11.5%).

The CLVs of risky and reliable customers were tested for normality and they didn’t come out to be normal sets. So we conducted Wilcoxon rank sum test in MATLAB as follows:

[p,h,stats]=ranksum(CLVX,CLVY)

p = 1.9208e-013

h = 1

stats = zval: -7.3542 ranksum: 754

The very small amount of p means we reject the null hypotheses of having equal medians. So the medians of the two populations are significantly different.

4.8 Fine-tuning the Model

As was shown in Figure 1.1, after we develop the model we should look for its possible flaws. In order to do so, we ran the sensitivity test on major elements of the model to see which one has the highest sensitivity. The results are shown below in Figure 4.2 through Figure 4.6.

119 The Sensitivity Percentage of Value Regarding No. remaining years to customer lifetime 1

0.9

0.8

0.7

Value percentage Change 0.6

0.5

0.4

0.3

0.2

0.1

0 0 5 10 15 20 25 30 No Remaining years to customer life

Figure 4.2. Sensitivity percentage of Future Value regarding remaining years of lifetime

The above figure shows the sensitivity of Future Value for different amounts of remaining lifetime for the customer. As it is shown, when the lifetime is very short the estimated value is highly sensitive but as it increases the amount will become very low. For example if the estimated remaining lifetime is only 5 years, one-year change in that parameter will change the amount of value for about 20% while when remaining lifetime is 15 years, a one-year change will change FV by less than 3%. We can conclude that the amount of estimated value is more accurate for customers whose lifetime is more than 10 years. (The other factors are assumed to be fixed: PD = 0.02, Repurchase = 75%, Growth = 30%).

On the other hand Figure 4.3 shows the sensitivity of Future Value for different amounts of repurchase rate. As it is shown when the repurchase rate is above 10%, FV is less sensitive and changes about 3%. For example if the estimated repurchase rate is about 20%, one percentage change in that parameter will change FV for about 2.5%.

120

The Sensitivity Percentage of Value Regarding Repurchase level 100

90

80

70

60

50

40 Value percentage change

30

20

10

0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Repurchase Level

Figure 4.3. Sensitivity percentage of Future Value regarding repurchase rate

But for repurchase rates below 10%, the sensitivity is very high. Other factors are assumed to be fixed: PD = 0.02, Remaining lifetime = 12, Growth = 30%.

Figure 4.4 shows how much the percentage change of Future Value would be if the growth rate changes 1%. For example if the estimated growth rate is 30%, the value will change 4.35% with one percentage change in the growth rate. As it is shown the highest sensitivity level is at 65% growth rate, assuming other factors to be fixed. (PD = 0.02, Remaining Year = 12, Repurchase = 75%).

121 The Percentage Sensitivity of Value regarding different growth rates 5

4.5

4

3.5 Value percentage change Value percentage

3

2.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Growth Rate change

Figure 4.4. Sensitivity percentage of Future Value regarding different growth rates

Figure 4.5 shows sensitivity of Future Value regarding different amounts of probability of default. As it is seen the most sensitive parameter here is probability of default and it is most sensitive at PD=0.3 level, which is more than 200%. In other words for example if estimated PD is 0.2 and it changes by 0.005, then the percentage change in FV would be about 4.5%. As it is shown the estimated amount of value with probability of default from 0.25 to 0.35 will be very sensitive and could be inaccurate assuming other factors are fixed (Repurchase =0.75, Remaining lifetime = 12, Growth = 30%). In other words this means our FV is very close to zero in this range of PD and it causes the amount of (FV 1-FV 2) /FV 1 go to infinity in this range.

122 The Sensitivity of Value Regarding different amounts of probability 2.5

2

1.5

1 Value percentage change percentage Value

0.5

0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probability of Default

Figure 4.5. Sensitivity of Future Value regarding different PDs

As could be seen in Figure 4.6, all lines intersect around point PD= 0.3 and the value of FV is zero around this point. So the numerator of the second term in equation 3.2 which we called the Future Value, is very close to zero around PD=0.3.

Amount of Value for different grwoth rates and PDs 3000000 40% growth 30% grwoth 2000000 20% grwoth 10% grwoth 50% growth 1000000 60% growth

0

-1000000

-2000000 Life Life time Value

-3000000

-4000000

-5000000

-6000000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Probablity of Default

Figure 4.6. Future Value for different growth rates and different PDs

123 4.9 Summary of the Chapter

This chapter basically went through the case study and model testing. We first calculated the qualitative risk scores of customers regarding our Delphi results from chapter 3. Then the quantitative score of same customers were extracted from information available at KB. The qualitative and quantitative scores were then used for PD calculation of customers. Customers were segmented based on their qualitative and quantitative risk scores being below or over 55%. So we had four segments based on customers’ risks and for each segment we counted the number of customers who have had past dues of more than three months and divided that by total of number of customers in that segment. As a result we got four PDs, each assigned to one segment and based on these PDs we calculated the future value of customers. The total value of each customer was then derived from Equation 3.2 and financial data available at KB.

The future values of customers came out to be negative for customers with PD= 99% and PD= 72%. This resulted in having 88.5% total estimated CLVs negative for risky customers and 91% positive CLVs for reliable customers. The Wilcoxon rank sum test verified that the medians of the distributions that these two medians come from are different.

At the end we tested the sensitivity of different parameters including PD, repurchase rate, growth rate, and remaining lifetime. The results showed that our mathematical developed model is most sensitive regarding parameter PD. This was because for PD = 1- (C a / (r L - rC)) the numerator of FV is zero and so the future value becomes zero. So our developed model is most accurate for PD # 1- (C a / (r L - rC)).

124

Chapter 5

Conclusion and Future Research

5. Conclusion and Future Research

During the past four chapters we went through the motivations for this research along with details of research steps undertaken to get the research to this point of conclusion. In this chapter, we will review the whole steps briefly and will mention limitations of this research regarding each step. The applications of the results for financial services industry are also discussed. Further research that could be conducted for verification and/or completion of this research is also discussed at the end of this chapter.

125 5.1 Over view of the Research

Customer value could be measured for customers of all industries, including service providers, retailers, and producers. In each of these industries, concept of relationship marketing is a force toward making long-term relationships with their customers and to do so, they have to know the value of their relationships so that they can better decide which relationships to keep and invest in. So any specific industry decision maker should pay a great deal of concern to the necessity of this measurement for his/her organization.

CLV according to Berger and Nasr, (1998) helps firms quantify customer relationships, and also customer lifetime value looks at what a retained customer is worth to the organization now, based on the predicted future transactions and costs. It is therefore far more useful than historic customer analysis as a basis for developing marketing strategies to maximize shareholder value (Ryals 2002). The applications of CLV which were reviewed in chapter 2 are summarized in Table 5.1.

Table 5.1. CLV Applications

Application Author

Segmentation/Product (Kim et al., 2006),(Hwang et al., 2004),(Liu and Shih, recommendation/ 2005),(Kumar et al., 2004),(Wu et al., 2005) Retention

Marketing campaign (Evans, 2002),(Ryals, 2002),(Hwang et al., 2004)

(Ryals and Knox, 2006),(Liu et al., 2007)/service effort allocation model,(Berger and Bechwati, 2001)/budget Resource allocation allocation,(Reinartz et al., 2005)/balancing Acq and Ret resources,(Berger and Bechwati, 2001)/allocation of promotion budget to maximize customer equity,

126 Application Author

(Venkatesan and Kumar, 2004)/ A CLV framework for resource allocation strategy

Merger and acquisition (Kumar et al., 2004), (Haenlein et al., 2007)

(Bayon and Gustsch, 2002),(Rust et al., Customer equity 2000),(Gummesson, 2004),(Wang and Hong, 2006),(Berger and Bechwati, 2001) (K.Stahl et al., 2003),(Bayon and Gustsch, 2002),(Ryals Value creation and Knox, 2006)

There exist many mathematical models for CLV measurement in literature. The basic one has been developed by Berger and Nasr (1998) and till 2007 that Crowder developed an abstract model for CLV estimation many researchers tried to focus on modeling specific aspects of CLV. The assumptions and specifications of some of these models which were reviewed in chapter 2 are brought in Table 5.2.

Table 5.2. CLV model characteristics, assumptions, and industry usage

Author Model specifications/ assumptions Industry

(Blattberg and Emphasis on the optimal level of acquisition Deighton, 1996) N.A. and retention spending, aggregate data

Sales take place once a year, Constant N.A. (Berger and Nasr, retention spending and retention rate, Numerical 1998) Constant contribution margin, aggregate examples data provided Markov chain modeling (MCM) N.A

Switching matrix would provide the ability Numerical (Pfeifer and to model both retention and migration examples Carraway, 2000) behavior provided Sample customers represent the customer (Rust et al., 2000) base of the firm, Purchase in a unit time N.A. occurs in intervals inversely to the average,

127

Author Model specifications/ assumptions Industry

aggregate data number of purchases Constant margin for each segment varies (Blattberg et at al., across time, retention rate and acquisition 2001) probability for each segment vary across N.A. time, aggregate data (Gupta and Constant Retention rate, Constant R.Lehmann, 2003) contribution margin/ constant growth rate N.A.

for the contribution margin, aggregate data

Measures three dimensions of CLV: current Contractual (Hwang et al., 2004) value, potential value, and loyalty. setting

Defection probability is measured for churn Tested on a (Kim et al., 2006) evaluation wireless

RM: data mining communication

Models CLV in four dimensions (Potential value, current value, loyalty, Network Contractual potential). This model actually brings into setting (Wu et al., 2005) account the effect of (WOM). Tested on RM: Time series, Probit modeling, health care Probabilistic Neural Networks (PNN) industry

Propose two hybrid methods that exploit the merits of the WRFM-based method and the Tested on a (Liu and Shih, 2005) preference-based CF method to improve the hardware retail quality of recommendations. store RM: Data mining

Individual customer level Estimates i) purchase frequency (using a generalized gamma distribution to model inter-purchase time), (ii) contribution margin (using panel-data regression (Kumar et al., 2006) Non-contractual methodologies), and (iii) marketing cost setting, tested (calculating the total marketing cost on a retail store separately for each customer based on the various marketing channels expected to be used to interact with that customer). Disaggregate data

(Haenlein et al., Using CART and MCM, probabilistic Tested on data

128

Author Model specifications/ assumptions Industry

2007) nature of MCM allows accounting for the of a retail bank, inherent stochasticity in customer non-contractual relationships, MCM is used for both setting retention and migration states.

Pareto/NBD modeling of CLV, Time when the customer becomes inactive is unknown

to the analyst, Customer can make any Non-contractual number of purchases and can become (Fader et al., 2007) setting, tested inactive any time. on Tuscan Limitations: the input data requirement for Lifestyle data each model should be data string of transactions for less than two years

N.A (Crowder et al., Abstract model Tested on some 2007) financial data

With all the applications that CLV measurement could have for organizations, still traditional profitability measurement is used in Iran. It is argued by Stahl (2003) that accounting-based profitability measures do not adequately reflect the value of a firm the main reasons being that (1) accounting methods differ widely, (2) risks are not adequately taken into account, (3) investment requirements are ignored, (4) dividend policy is not reflected and (5) the time value of money is ignored. These flaws have been the motivation for conducting this research and we basically focused on considering risk in our CLV measurement because the financial risk score of customers that are currently used in Iran in banking are unreliable due to unreliability of financial statements and limited available financial data.

5.1.1 Research Design

An overview of the whole research steps is shown in Figure 5.1. In order to gain a risk-adjusted CLV model, we first had to model the risk. Since the target customers of this research were business loan applicants, by risk of relationship we meant the probability of default of customers. Due to unreliability of financial statements we couldn’t rely solely on financial statements, so we modeled the risk based on both

129

What risk model best solves the inaccuracy of current scoring models

Delphi method steps Questionnaire 1: Initial collection of Selecting experts factors

Brainstorming for important factors Questionnaire 2: Validation of initial list of factors

Narrowing down factors Questionnaire 3: Choosing most important

Assigning weight to factors Questionnaire 4: Grading the extracted factors

Grading business customers according to Delphi output

Qualitative credit score modeling of Quantitative credit score modeling of business customers business customers by logit

Credit Risk Matrix

PD estimation of customers

Embedding PDs in CLV calculation for future revenue estimations

Risk adjusted CLV model

Figure 5.1. Flowchart of the research

qualitative and quantitative scores. The qualitative risk factors were extracted through Delphi method by 23 panelists nationwide of whom 16 were credit experts and 7

130 were risk experts. Then 101 firms were chosen from KB’s customers to be graded on the basis of these extracted attributes and the qualitative risk scores of these firms were determined.

These customers were those who have had at least 2 years of relationship with KB and still continue to be customers. The quantitative risk scores of these same customers were calculated with logistic regression. The PD of each customer was assigned based on 4 different score classes that we classified customers according to their qualitative and quantitative scores.

On the other hand a CLV approximation model for loan applicants of a bank was developed using “1-PD” for future revenue generation of loan applicants. This mathematical model reflected the loan amount, the assigned interest rate for the year it was taken, the repurchase rate of the customer, an approximate growth rate of 29%, a WACC of 30%, cost of money of 16% of the loan amount, and administration cost of 2% of loan amount.

5.2 Research Outcomes

The risk scoring method that was discussed in this research was used for PD estimation of customers which is summarized in the matrix of Figure 5.2. Qualitative credit score creditscore Qualitative

1 2

PD = 7.1% PD = 2.2%

0.55

3 4

PD = 100% PD = 72.4% 0 0.55 Quantitative credit score

Figure 5. 2. The risk score matrix

131 Quadrant 2 of this risk matrix shows customers with qualitative and quantitative scores more than 0.55 which means they could be the most profitable customers to be cross-sold or up-sold.

Customers in quadrant 3 on the other hand have the highest probability of default estimated as 100% for our sample, but due to possible errors in our estimation and the limited sample size, assigned PD to this quadrant is mentioned as 99%. Customers of this quadrant carry the highest risk so should be monitored with more care.

These PDs were then embedded in term 2 of our developed CLV estimation model indicated as:

Ni Ni −ti CLV i =ƒ[](rL(ti ) −rc(ti ) ) −Ca(ti ) Lti 1( +WACC ) + ti&=0)))))))) ')))))))) ( Term 1

Ni +Mi ([]1( − PD )( r −r ) −C L (Rate )( 1+ g)Ni +Mi −ti ) ƒ i L(ti ) c(ti ) a(ti ) ti repuri 1( +WACC )ti −(Ni +Mi ) ti&=Ni)+1))))))))))) ')))))))))))) ( Term 2

for PD # 1- (C a / (r L - rC))

Each element of this formula was either calculated from financial records of customers (like the year and amount of loan granted), or the information available at KB documents such as WACC estimation. The results of this CLV estimation model revealed 88.5% negative amounts for customers of quadrant 3 (PD=99%) and quadrant 4 (PD=72.4%) vs. 9% negative CLVs for other two quadrants of lower PDs. This indicated that customers with lower PDs (which means less risky customers) would bring more value for the bank during their lifetime, although all 101 customers have had positive current values and are willing to continue relationship with the bank. So this risk-adjusted CLV model actually helped management of KB to see the future of a retained customer and prevent probable losses. For instance by having customers sorted based on their

132 current values from the lowest to the highest like Table 5.3, the management could see that the two lowest current value holders could make high values during their lifetime while customer # 92,93,94, and 95 bring negative total values in long term due to their risk although being among top 10% current value holders.

Table 5. 3. CLVs sorted based on lowest current value to the highest /    /  ë  C  ë  Ç  /[ë  9!   9!  9!  9!  9! "  "  9!  9! " "9!  9! "  " 9!  9! "9! "9!  9!   9!  9! "  "9!   9!  9!  9!  9!  9! "  "  9!  9!  9!  9!    9!   9!     9!  9!   9!   9! "   "9!  9!  9!  9!    9!  9! "9! " 9!   9!  9!   9! " 9! "9!  9!   9!  9! "  " 9!   9! "  9! "  9!  9! "  9! " 9!  9!   9!  9!    9!  9!  9!  9!   9!  9!   9!  9! " 9! " 9!  9!   9!  9! "9! " 9!   9!    9!  9! "9! "  9!

133 /    /  ë  C  ë  Ç  /[ë  9! "9! "9!  9!   9!   9! " 9! " 9!  9!  9!  9! "  9! "9!  9! "9! "9!  9! " 9! "9!   9! "   9!  9!   9!  9! " 9! " 9!  9!    9!  9!  9!   9!  9! 9!  9!    9!  9! "9! "9!  9!   9!  9! " 9! "9!  9!    9!   9!    9!  9! "  9!   9!   9!  9! "9! 9!   9! "9! "9!  9!  9!  9! "9! "9!  9!  9!  9!   9!  9!     9!   9!   9!  9! "9! " 9!  9! "9! "9!  9! "9! "9!  9!   9!  9! " 9! "9!  9!     9!  9! "9! " 9!   9!  9!  9! "9! " 9!   9!    9!  9!   9!  9!  9!   9!   9!

134 /    /  ë  C  ë  Ç  /[ë   9!   9!   9! "9! "9!  9!  9!   9! "   9!  9!  9!  9!  9!   9!  9!  9!      9!  9!    9!  9! "9! "9!  9!    9!  9!    9!  9!  9!   9!     9!  9! "9! " 9!  9! "9! "9!   9! " 9! "  9!  9! " 9! " 9!  9!   9!  9!     9!  9!    9!  9!  9! 9!  9!  9! 9!  9! 9! 9!

We could have customers in two dimensions of current value and future value for easier decision making of the management. This classification is shown in Figure 5.3.

Customer Values

4E+11 3E+11 2E+11 1E+11 0 0.0E+00 1.0E+10 2.0E+10 3.0E+10 4.0E+10 5.0E+10 6.0E+10 7.0E+10 -1E+11 Future Value Future -2E+11 -3E+11 -4E+11 Current Value

Figure 5.3. Customer values based on two value dimensions

135 As can be seen in Figure 5.3, majority of our sample customers have low current and future values while a few have more than 3.0E+10 current value and positive future values, who are possibly the highest revenue generating customers of the bank.

5.4 Research Applications for Financial Service Providers

The risk adjusted CLV model that we developed would have various applications for banking industry. We bring up some of them in this section.

1. Segmentation/product recommendation: segmentation of customers was once made when we developed the risk matrix (four PDs for four quadrants) for our customers. We could also segment customers based on their CLVs in ascending order and then have positive CLVs as one segment and negative CLVs as another segment like Figure 5.4.

Highest CLV

1.35E+12 62% Profitable segment positive CLVs

38% negative Lowest CLV Risky segment CLVs -2.4E+11

Figure 5.4. Segmentation based on CLVs

For each segment then the management of the bank would decide on improvement of relationship with the customer. For instance the management could check to see how much the bank spends yearly on a customer with high risk who has got a negative CLV and would then decide on cutting those expenses. On the other hand the low risk high value customers who are considered the highest revenue generating assets would be offered different kinds of financial services from loans to insurance while the other segment which seemed less profitable in long term would be improved for their risk

136 according to the obtained risk matrix and management of KB has decided not to expose its services much to this segments’ customers. The management would scrutinize to explore if these relationships are worth the level of resources they receive or whether they should be managed at lower service levels. The process has actually helped KB to identify which of their relationships are truly key relationship lending and are worth special services.

2. Loan pricing: the CLV model that we developed could help management of a bank to price loans. This means customers who are shown to create high value in long term would be offered lower interest rates for their loans. 3. Retention: Management of the bank could decide which relationships are worth keeping due to long term value that they create for the bank and would put more effort in retaining them. 4. Shareholder value creation: as exhibited in chapter 1, there is a great loss happening to the bank due to defaults that customers make. If these losses are prevented by the help of tools like our risk adjusted CLV model, then the shareholder value is increased dramatically. 5. Lending advisory: the risk matrix that we developed could be used for lending advisory to the firms. Customers in quadrant 3 have the highest probability of default estimated as 100% for our sample. Customers of this quadrant carry the highest risk so should be monitored with more care. The management of the bank could improve these relationships by leading them through quadrant 1 by having more subjective considerations or by pushing them through quadrant 4 by having more objective considerations. For instance firms in quadrant 3 of our risk matrix could be pushed towards quadrant 1 by improving their qualitative attributes or customers of quadrant 1 could progress to quadrant 2 by having better financial traits. This lending advisory could be a revenue generation method for the bank. 6. Collateral settings: the CLVs of customers could also be used for setting collateral conditions for customers. This means the less risky the customer, the more valuable he is in long term, so less collateral would be set for him.

137 5.5 Research Limitations

In each step of the research some restrictions existed and we bring up some of them in this section.

5.5.1 Limitations of Delphi Method Procedure

a) Expert selection: the main limitation in our Delphi process was having fewer than required risk experts on the panel. Since risk management is a brand new major in Iran, the risk managers and those who could be considered risk experts in banking were very few in number. This limitation caused our Delphi risk panel to have less than required panelist and having had between 10-18 experts on this panel would have increased the reliability of our extracted attributes.

b) Contacting experts: originally we as Delphi researchers were supposed to send questionnaires, from the first one to the fourth one, to our experts. For our research though, we needed to put so much time explaining the relationship lending concept to our experts. Although this concept seemed to have been existing in the banking industry, the experts had never faced it with this term and some weren’t recognizing that they had this lending technique in their banking system. So a long telephone conversation or physical presence of Delphi researcher along with first questionnaire was inevitable. This caused the first questionnaire’s phase to be a very long process.

c) The Delphi method that we conducted for this research had been adapted for initial part of our research purpose which was risk scoring. We extracted 13 attributes on which the panelists had consensus on so we never checked for inter-relation of extracted attributes which would have caused inaccuracy of the results.

5.5.2 Limitations of Scoring

a) Qualitative scoring of the firms was done with one member of the credit committee who had been among our Delphi experts and only 25 firms were graded by the head of the branch that had granted the credit facility. If we had the opportunity to have at

138 least two opinions for grading the firms based on Delphi attributes, the results would have been more accurate. Unfortunately we were not able to do the follow-up for 75 firms due to some restrictions like unavailability of the head of the branch (some were no longer at KB or had been in other cities rather than Tehran) and time constraint to continue with the rest of the research.

b) Quantitative scoring as mentioned earlier in chapter one faced to major limitations. First was very limited stored data that we had which was due to young age of the bank. Second reason was unreliability of the financial statements which caused high error (35% on average) and made the accuracy of the modeling decrease.

5.5.3 Limitations of CLV Modeling

The developed risk adjusted CLV model of our research was designed for credit facility applicants but since the guarantee and LC’s data are not stored in computer (still on paper), it wasn’t possible to have all credit facility applicants considered in the model. Loan applicants were the only group we had good enough digital data about.

The growth rate of customers was considered a constant due to limited amount of years of relationship that customers had with KB. If the relationship durations had been longer, then a specific growth rate could be determined at least for each industry if not for each customer and the results would have become more accurate.

5.6 Future Research Directions

There are different aspects of this research which could be followed by researchers of this area.

The Delphi method that we used was adapted for the purpose of this research hence fore we did not go through the ranking process after consensus. Instead we asked experts to assign weight to each attribute. Further research could be done on our Delphi results for ranking the risk attributes that we had got. The ranking could then be used for other research purposes in financial services’ industry.

139 Also the attributes that were extracted from our Delphi method could have been inter-related and just because we trust the experts’ attitudes, we never check for this inter- relation. A statistical test could be conducted on these extracted attributes to check if any of them are inter-related.

In order to have more accurate results for grading the firms, it would be possible that firms are chosen based on availability of more than one grader for them. So in addition to the member of the credit committee of our case study who knew all chosen firms, at least another member of the committee or the head of the loan granting branch could be asked to grade that firm.

For the risk matrix that we developed, point (0.55, 0.55) was estimated with trial and error. This point could be estimated with optimization methods which would indicate more sensitivity in results.

There is very limited variety of products that we have in Iranian financial services which would limit us to loyalty and current value estimations. Further research could be conducted to broaden our CLV model in another dimension of potential value meaning that one could analyze the cost of different services and then by analyzing past behavior of customers predicting what potential value the customer could bring for the organization.

The collateral value of each customer is also an important issue. This could be estimated as a coefficient being considered in future revenue generation of loan applicants.

There are also many researches done on switching behavior of customers. It would be a complement of this research to have the switching behavior for banking industry with MCM methodology for instance and embed it into our CLV model.

Similar CLV model could be developed for other industries considering the specifications of that industry and risk factors of the customers in that industry.

140 References: ARAVINDAKSHAN , ASHWIN , et al., 'CUSTOMER EQUITY: MAKING MARKETING STRATEGY FINANCIALLY ACCOUNTABLE', JOURNAL OF SYSTEMS SCIENCE AND SYSTEMS ENGINEERING, 13/4 (2004), 405-22. BAAS , TIMO and S CHROOTEN , MECHTHILD , 'Relationship Banking and SMEs: A Theoretical Analysis', Small Business Economics, 27 (2006), 127-37. BAYON , TOMAS and G USTSCH , JENS , 'Customer Equity Marketing: Touching the Intangible', European Management Journal, 20/3 (2002), 213-22. BERGER , PAUL D. and N ASR , NADA L., 'CUSTOMER LIFETIME VALUE: MARKETING MODELS AND APPLICATIONS', Journal of Interactive Marketing, 12 (1998). BERGER , PAUL D. and B ECHWATI , NADA NASR , 'The allocation of promotion budget to maximize customer equity', Omega, 29 (2001). BLATTBERG , R. C. and D EIGHTON , J., 'Manage marketing by the customer equity test', (74, 1996), 136-44. BURN , JANICE and A SH , COLIN , 'A dynamic model of e-business strategies for ERP enabled organisations', Industrial Management & Data Systems, 105 (2005), 1084-95. COURTHEAUX , R.J. , 'Database Marketing: Developing a Profitable Mailing Plan', Catalog Age, (1986). CROWDER , MARTIN , J.HAND , DAVID , and K RZANOWSKI , WOJTEK , 'On optimal intervention for customer lifetime value', European Journal of Operational Research, (2007). DERMINE , J and B ISSADA , Y, Asset and Liability Management (Prentice Hall, 2002). DWYER , F.R., 'Customer lifetime valuation to support marketing decision making', Journal of Direct Marketin, 3 (1989), 8-15. EVANS , GOERGE , 'Measuring and managing customer value', work study, 51 (2002), 134- 39. FADER , PETER S., H ARDIE , BRUCE G.S., and L EE , KA LOK , 'RFM and CLV: Using Iso- Value Curves for Customer Base Analysis', Journal of Marketing Research, 12 (2005), 415-30. FADER , PETER S., G.S.HARDIE , BRUCE , and J ERATH , KINSHUK , 'Estimating CLV using aggregated data: The Tuscan lifestyles case revised', Journal of Interactive Marketing, 21/3 (2007). GUMMESSON , EVERT , 'Return on Relationships (ROR): The value of relationship marketing and CRM in B-2-B contexts', Journal of Business and Industrial Marketing, 19 (2004), 136-48. GUPTA , S., L EHMANN , D., and S TUART , J., 'Valuing Customers', Journal of Marketing Research, 41 (2004), 7-18. GUPTA , SUNIL and R.LEHMANN , DONALD , 'Customers as assets', Journal of Interactive Marketing, 17/1 (2003), 9-24. HAENLEIN , MICHAEL , K APLAN , ANDREAS M., and B EESER , ANEMONE J., 'A Model to Determine Customer Lifetime Value in a Retail Banking Context', European Management Journal, (2007). HWANG , HYUNSEOK , J UNG , TAESOO , and S UH , EUIHO , 'An LTV model and customer segmentation based on customer value: a case study on the wireless

141 telecommunication industry', Expert systems with Applications, 26 (2004), 181- 88. JACKSON , BARBARA , 'Build customer relationships that last', Harvard Business Review, 64 (1985), 120-228. JACKSON, D.R, 'Achieving Strategic Competitive Advantage Through the Application of the LongTerm Customer Value Concept', Journal of Database Marketing, 4 (1996), 174-86. JAIN , DIPAK and S INGH , SIDDHARTHA S., 'Customer lifetime value research in marketing: A review and future directions', Journal of Interactive Marketing 16 (2002). K.STAHL , HEINZ , M ATZLER , KURT , and H.HINTERHUBER , HANS , 'Linking customer lifetime value with shareholder value', Industrial Marketing Management, 32 (2003), 267-79. KALWANI , M.U and N ARAYANDAS , 'Long-term manufacturersupplier relationships: do they pay off for supplier firms? ' Journal of Marketing, 59 (1995), 1-16. KIM , B. and K IM , S., 'Measuring up-selling potential of life insurance customers: application of a stochastic frontier model', Journal of Interactive Marketing, 13 (1999), 2-9. KIM , J., 'e-CRM for e-Business, Gurm', (2000). KIM , SU-YEON , J UNG , TAE -SOO , and S UH , EUI -HO, 'Customer segmentation and strategy development based on customer lifetime value: A case study', Expert systems with Applications, 31 (2006), 101-07. KOTLER and F OX , Strategic Marketing for Educational Institutions (Prentice-Hall, 1995). KUMAR , V. and S HAH , DENISH , 'Building and sustaining profitable customer loyalty for the 21st century', Journal of Retailing, 80 (2004), 317-30. KUMAR , V., R AMANI , GIRISH , and BOHLING, TIMOTHY, 'Customer lifetime value approaches and best practice applications', Journal of Interactive Marketing, 18 (2004). KUMAR , V., S HAH , DENISH , and V ENKATESAN , RAJKUMAR , 'Managing retailer profitability-one customer at a time', Journal of Retailing, 82 (2006), 277-94. KUMAR , V. and G OERGE , MORRIS , 'Measuring and maximizing customer equity: a critical analysis', J. of the Academy of Marketing Science, (2007). KUTNER , MICHAEL H., et al., 'logistic regression, poisson regression, and generalized linear models', Applied Linear Statistical Models (fifth edn., Mc Graw Hill, 2005), 1420. LINDGREEN , ADAM and C RAWFORD , LAN , 'Implementing, monitoring and measuring a programme of relationship marketing', Marketing Inteligence and Planning, 17/5 (1999), 231-39. LIU , BEN SHAW -CHING , P ETRUZZI , NICHOLA C., and S UDHARSHAN , D., 'A service effort allocation model for assesing customer lifetime value in service marketing', Journal of Service Marketing, 21/1 (2007), 24-35. LIU , DUEN -REN and S HIH , YA-YUEH , 'Hybrid approaches to product recommendation based on customer lifetime value and purchase preferences', The Journal of Systems and Software, 77 (2005), 181-91. ---, 'Integrating AHP and data mining for product recommendation based on customer lifetime value', Information and Management, 42 (2005), 387-400.

142 MASON , C.H., 'Tuscan Lifestyles: Assesing Customer Lifetime Value ', Journal of Interactive Marketing, 17 (2003), 54-60. OKOLI , C. and P AWLOWSKI , S.D., 'The Delphi Method as a research tool: an example, design considerations and application', Information and Management, 42 (2004), 15-29. PFEIFER , PHILLIP E. and C ARRAWAY , ROBERT L., 'Modeling customer relationships as markov chains', Journal of Interactive Marketing, 14/2 (2000). REINARTZ , WERNER , T HOMAS , JACQUELYN S., and K UMAR , V., 'Balancing Acquisition and Retention Resources to Maximize Customer Profitability', Journal of Marketing, 69 (2005), 63-79. REINARTZ , WERNER J. and K UMAR , V., 'On the Profitability of Long-Life Customers in a Noncontractual Setting: An Empirical Investigation and Implications for Marketing', Journal of Marketing, 64 (2000), 17-35. ROSSET , SAHARON, et al., 'Customer Lifetime Value Models for Decision Support', Data Mining and Knowledge Discovery, 7 (2003), 321–39. RUST , R.T, L EMON , KATHERINE N, and Z EITHAML , 'Return on marketing: Using customer equity to focus marketing strategy', Journal of Marketing, 68 (2004), 109-27. RUST , ROLAND T., Z EITHMAL , VALERIE A., and L EMON , KATHERINE N., Driving customer Equity: How Customer Lifetime Value is Reshaping Corporate Strategy (2000). RYALS , LYNETTE , 'Are your customers worth more than money?' journal of Retailing and Consumer Services, 9 (2002), 241-51. RYALS , LYNETTE and K NOX , SIMON , 'Measuring and managing customer relationship risk in business markets', Industrial Marketing Management, (2006). SABZEVARI , HASSAN , S OLEYMANI , MEHDI , and N OORBAKHSH , EAMAN , 'A comparison between statistical and Data Mining methods for credit scoring in case of limited available data', Credit Scoring and Credit Control (Edinburgh/ Scottland, 2007). SAUNDERS , M, L EWIS , P, and T HORNHILL , A, Research methods for business students (2'nd edn.: Prentice Hall, 2000). SCHMIDT , R.C, 'Managing Delphi surveys using nonparametric statistical techniques', Decision Sciences, 28/3 (1997), 763-74. SCHMITTLEIN , D.C and M ORRISON , D.G, 'Who are they and what will they do next?' Management Science, 33 (1987), 1-24. STAHL , HEINZ K., M ATZLER , KURT , and H.HINTERHUBER , HANS , 'Linking customer lifetime value with shareholder value', Industrial Marketing Management, 32 (2003), 267-79. VENKATESAN , RAJ KUMAR and K UMAR , V., 'A customer lifetime Value framework for customer selection and resource allocation strategy', Journal of Marketing, 68 (2004), 106-25. VERHOEF , P. C. and D ONKERS , B., 'Predicting customer potential value an application in the insurance industry', (32: Elsevier, 2001), 189-99. WANG , HSIAO -FAN and H ONG , WEI -KUO , 'Managing customer profitability in a competitive market by continuous data mining', Industrial Marketing Management, 35 (2006), 715-23. WU, LIHUA , L IU , LU, and L I, JING , 'Evaluating Customer Lifetime Value for Customer Recommendation', IEEE, (2005).

143 Appendix 3. 1. Questionnaire 1 of Delphi process

Dear respondent, As you know, the factors which are influential in continuation of lending relationship between a lender and a business borrower in financial services are divided into two main categories of qualitative and quantitative factors. The purpose of this questionnaire is to extract the qualitative factors that would indicate the risk of this relationship in Iran. Please note that by risk factors of lending relationship we mean factors that increase the probability of default of a relationship borrower. According to what is said please identify as many factors as possible, which in your opinion and regarding to your experience, are influential in riskieness of lending relationship with a business relationship borrower. Please give a brief description of the reason for your choice if possible.

We sincerely appreciate your time in advance

144 Appendix 3. 2. Questionnaire 2 of Delphi process Dear respondent, As you know, the factors which are influential in continuation of lending relationship between a lender and a business borrower in financial services are divided into two main categories of qualitative and quantitative factors. The attached list shows 26 of these factors which are identified by experts of financial services in Iran including you. The purpose of this questionnaire is to have a second opinion on these factors and to make sure the list satisfies the required information on the original question of the research which was “identification of the factors influential in the probability of default of a business relationship borrower.” Please verify the attributes on the list below as either related or non-related to 1) the purpose of the research and/ or 2) the identified category. At the end please also verify your own list of attributes and make necessary changes if applicable. We sincerely appreciate your time in advance

Summarized If a non- Category Risk Factors description by the risk factor, should experts verify the change reason to… Environmental Factors Different economic Economic sector of the firm sectors like agriculture, construction, … have different risks More competitors bring Competence capability more risk due to loss of job The less under governmental

145 Degree of deregulation regulations, the less risky since the governmental regulations are not stable Growth rate of the firm's The higher the domestic industry growth, the less the risk Impact of inflation on firm's The higher the impact, activity the higher the risk The higher the impact, Independency from imports the higher the risk The higher the market Market share of the firm share, the lower the risk

The better the Performance of the firm in the performance, the lower banking system the risk The better the Relationship of the firm and relationship the lower its clients the risk If the number and variety Number of suppliers of suppliers is high, the risk is lower If the number and variety Number of buyers of buyers is high, the risk is lower The more the WOM, the Extent of word of mouth the less the risk of losing client can bring customer Internal Factors

Management quality of the Experience of the

146 firm management in the field, related education The higher the sales Sales fluctuations of the firm fluctuations, the higher the risk If high, means Ratio of variable cost/ fixed technology usage and/or cost internal management is weak, so the risk is high Liquidation capability Type/amount of collateral and and type of promissory reliability of cosigners notes, stocks, residential property, account receivable The higher the Activity permits of the firm compatibility of permits with activity, the lower the risk The higher the volume of Production capacity of the production, the lower the Firm risk

Five options identified Reliability of the firm's from low risk to high financial statements risk: Audited financial statements, un-audited financial statements, tax statements, balance sheet only, no financial statement If the usage is beneficial

147 Credit facilities’ usage and matches the firm’s purpose activities, the risk is lower Lender- borrower Relationship The longer you know the Longevity of the relationship customer, the lower the information asymmetry, the lower the risk The more reliable the Reliability of the referee referee, the lower the risk

Number of bought services by Could be both positive or the firm negative in terms of risk Concentration of Extent of checking account checking account activity within the Bank activities in the bank means more reliability and control Concentration of credit Extent of credit activities activities in the bank within the Bank means more reliability and control If the growth is high and Growth rate of the firm's matches the firm’s credit activity with the bank activity growth, the risk is lower

148 Appendix 3.3. Questionnaire 3 of Delphi process Dear respondent, As you know, the factors which are influential in continuation of lending relationship between a lender and a business borrower in financial services are divided into two main categories of qualitative and quantitative factors. The purpose of this questionnaire is to extract significant qualitative factors that would indicate the risk of this relationship in Iran. Please note that by risk of lending relationship we mean the risk of default of a client. According to what is said, please identify at least ten factors from the list below that in your opinion and regarding to your experience are influential in probability of default of a business customer. Please give a brief description of the reason for your choice if possible.

We sincerely appreciate your time in advance

Risk Factors Your reason for this selection Environmental Factors Economic Sector of the Firm Competence Capability Degree of Deregulation Growth Rate of the Firm's Industry Impact of Inflation on Firm's Activity Independency from Imports Market Share of the Firm Number of Suppliers Number of Buyers Internal Factors Management Quality of the Firm Sales Fluctuations of the Firm Ratio of Variable Cost/ Fixed Cost

149 Type of Collateral and Reliability of Cosigners Activity Permits of the Firm Production Capacity of the Firm Reliability of the Firm's Financial Statements Credit facilities’ usage purpose Lender- borrower Relationship Longevity of the Bank-Firm Relationship Number of Products Bought by the Firm Firm's Checking Account Activity within the Bank Firm's Credit Activities within the Bank Growth Rate of the Firm's Credit Activity with the Bank

150 Appendix 3.4. Experts’ Votes of the two rounds

Total Total RE CE RE CE

Round1 Round2 Round1 Round1 Round2 Round2 Economic Sector 11 12 9 2 9 3 Competence capability 12 14 8 4 10 4 Degree of deregulation 12 13 9 3 9 4 Industry's Growth Rate 12 15 8 4 10 5 Inflation 5 7 4 1 4 3 Dependency from 12 13 6 6 8 5 Imports Market Share 14 12 8 6 9 3 Number of Suppliers 13 13 9 4 10 3 Number of Buyers 15 13 14 1 9 4 Management Quality 22 21 16 6 15 6 Sales Fluctuations 7 5 5 2 2 3 Variable Cost/ Fixed 1 2 0 1 1 1 Type of Collateral 16 19 14 2 15 4 Activity Permits 8 6 7 1 5 1 Production Capacity 12 13 11 1 8 5 Financial Statements 11 17 9 2 11 6 Longevity of Reltnsp 15 14 10 5 8 6 Number of Bought 7 4 3 4 2 2 Products Account Activities 19 15 12 7 10 5 Credit Activities 8 13 5 3 9 4 Growth Rate of Credit 12 11 7 5 6 5 Identified Usage 13 14 10 3 7 7

151 Appendix 3.5. Questionnaire 4 of Delphi process

Dear respondent, As you may know, the factors which are influential in continuation of lending relationship between a lender and a business borrower in financial services are divided into two main categories of qualitative and quantitative factors. The purpose of corresponding previous questionnaires was to extract the qualitative factors that would indicate the risk of this relationship in Iran. Please note that by risk of lending relationship we mean the probability of default of a client. According to what is said, please grade the importance of each of the extracted factors on the scale shown in front of each factor. Please note that the purpose of this questionnaire is to indicate the importance of each factor regardless of its positive or negative effect on relationship risk. Please give a brief description of the reason for your choice if possible.

We sincerely appreciate your time in advance

Factor Very Poor Little Average Little Strong Very poor poor strong strong

Environmental factors

1.Number of competitors

2.Impact \ degree of government regulations on firm’s activities

3.Impact of imports on client’s business

4.Client’s market share

5.Domestic growth of firm’s industry

6.Buyers’ bargain power

152 Factor Very Poor Little Average Little Strong Very poor poor strong strong

7.Suppliers’ bargain power

Internal factors

8.Management quality of the firm

9.Type of collateral and credibility of cosigners

10.Firm’s production/sales capacity

11.Reliability of firm’s financial statements

Lender-borrower Relationship

12..Longevity of relationship with lender

13.Firm’s checking account activity within the bank

153 Appendix 3.6 . Risk Attribute Weights Assigned by Delphi Experts

A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12 A13 E1 6 6 5 6 5 5 4 7 4 2 7 6 6 E2 3 6 6 6 4 6 2 6 4 6 7 6 7 E3 6 5 5 6 4 7 4 5 3 3 5 5 6 E4 4 6 4 6 5 5 5 5 5 4 6 4 5 E5 5 5 5 6 6 6 4 6 4 5 7 7 5 E6 4 5 6 6 5 6 5 7 7 4 6 5 6 E7 5 3 7 6 6 6 5 7 5 5 7 6 7 E8 5 6 6 5 7 6 5 7 4 6 7 7 7 E9 4 3 5 5 4 4 2 7 7 5 4 6 6 E10 4 6 5 6 7 7 7 7 7 7 7 6 7 E11 6 7 7 6 6 6 7 6 5 5 4 6 4 E12 2 3 4 4 4 4 3 5 5 3 5 6 6 E13 3 6 6 6 4 6 3 7 4 7 5 6 7 E14 3 6 6 6 3 6 3 6 4 6 7 7 7 E15 2 2 3 5 5 6 4 4 3 5 3 6 5 E16 5 5 6 7 5 7 7 7 6 6 7 5 6 E17 5 6 4 6 6 6 5 7 7 6 7 5 5 E18 4 6 5 6 6 4 3 7 6 5 7 7 6 E19 6 7 6 6 5 6 5 7 7 5 7 4 5 E20 4 2 3 6 6 6 6 7 5 6 6 7 7 E21 4 6 5 6 5 6 6 7 5 4 7 6 7 E22 5 5 4 5 5 7 5 6 6 5 6 6 7 E23 5 6 6 5 5 6 6 7 5 5 7 5 6 4.35 5.13 5.17 5.74 5.13 5.83 4.61 6.39 5.13 5 6.13 5.83 6.09

154 Appendix 4.1. Qualitative scores of the firms Customer ID Qual Score Performance Customer ID Qual Score Performance 1   51     2   52    3   53   4   54    5     55    6   56    7    57    8   58     9   59    10   60   11   61    12   62    13   63    14   64    15    65   16    66    17    67   18   68   19   69    20    70   21   71   22    72    23   73   24   74    25   75   26   76   27   77   28   78     29    79    30    80   31    81    32   82   33   83   34    84   35   85    36   86   37    87   38    88   39    89   40     90     41   91    42   92   43   93    44    94   

155 Customer ID Qual Score Performance Customer ID Qual Score Performance 45   95    46   96   47    97   48    98   49   99    50   100    

156 Appendix 4.2. Quantitative scores of the firms Customer Quant Score Performance Customer Quant Score Performance ID ID 1   51    2    52   3    53   4   54    5   55   6   56    7   57   8   58   9   59    10   60 9"  11   61   12   62   13   63   14   64    15   65    16   66   17    67   18    68   19   69   20     70   21    71    22   72   23   73    24   74   25   75    26   76   27   77   28   78   29   79   30   80   31   81   32   82   33   83     34   84    35    85   36   86    37   87   38   88    39   89    40   90   41   91    42   92   43   93  

157 Customer Quant Score Performance Customer Quant Score Performance ID ID 44   94   45   95    46   96    47     97   48   98   49    99   50    100   101  

158

Appendix 4.3. Geometric average of four consecutive year loan demand     D  () 9!  9! 9!  9! "  9! 9!  9! 9! "   9! 9! 9! 9! "  9!  9! 9! 9! "  9! 9! 9! 9! " 9! 9! 9! 9! " 9! 9! 9! 9! " 9! 9! 9!  9! "  9! 9! 9! 9! " 9! 9! 9! 9! " 9! 9! 9! 9! " 9!  9! 9! 9! "  9! 9! 9! 9! "  9! 9! 9! 9! " 9! 9! 9! 9! "   9! 9! 9! 9! " 9! 9!  9! 9! "   9! 9!  9! 9! "   9! 9! 9! 9!  9! 9! 9! 9!   9! 9! 9! 9!   9!  9! 9! 9!   9! 9!  9! 9!  9! 9! 9! 9!  9! 9! 9!  9!    9! 9!  9!  9!  9! 9!  9! 9!  9! 9! 9! 9!   9!  9! 9! 9!  9! 9! 9! 9!  9! 9! 9! 9!   9! 9! 9! 9!     9! 9!  9! 9!    9! 9! 9! 9!    9! 9! 9!  9!   9! 9! 9! 9!   9! 9! 9! 9!  9!  9! 9! 9!  9! 9! 9!  9!  

159     D  () 9! 9! 9! 9!   9! 9! 9! 9!  9! 9! 9!  9!  9! 9! 9! 9!  9!  9! 9! 9!  9! 9! 9!  9!  9! 9! 9! 9!   9! 9! 9! 9!   9!  9!  9! 9!  9! 9! 9! 9!  9! 9! 9! 9!   9! 9! 9! 9!  9! 9! 9! 9!   9! 9! 9! 9!  

160 Appendix 4.4. Repurchase rate of customers Customer Real PD Customer ID Real PD Customer ID Real PD ID 1 0.44  1.00  0.80 2 0.70  1.00  0.80 3 0.70  1.00  0.80 4 0.70  0.60  0.80

5 0.60  1.00  0.80 6 0.60  0.50  1.00 7 1.00  0.80  1.00 8 0.80  1.00  1.00 9 1.00  0.30  1.00 10 1.00 0.75  0.75 11 1.00  0.80  1.00 12 0.30  0.80  1.00 13 1.00  0.80  0.75 14 0.30  1.00  0.75 15 0.30  0.50  1.00 16 0.60  0.75  0.80 17 0.80  1.00  1.00 18 0.70  1.00  1.00 19 0.80  0.50  1.00 20 1.00  0.50  1.00 21 1.00  0.50  1.00 22 0.60  1.00  0.50 23 0.60  0.75  0.75 24 0.50  1.00  1.00 25 1.00  1.00  0.70

 0.80  1.00  1.00

 1.00  0.70  1.00

161 Customer Real PD Customer ID Real PD Customer ID Real PD ID  0.80  0.80  1.00

 0.10  0.20  0.70

 0.75  1.00  1.00

 0.60  0.50  1.00

 0.60  0.75  0.50

 0.70  1.00  0.50

34 0.40  1.00

162 Appendix 4.5. Future value function for CLV calculation function [FutureValue] = ValueCal(LoanAmount,PD,RmYear,Growth,Cofund,Admin,Repurchase,WACC ,Rate)

Value1=((1-PD)*Rate-Cofund-Admin).*LoanAmount.*Repurchase); for j= 1:101 i=1:RmYear(j); g(1:RmYear(j))=(1+Growth).^(i-1); AllValue(1:RmYear(j))= (Value1(j).*g)./((1+WACC).^(i-1)); FutureValue(j) = sum(AllValue); i=0; g=0; AllValue=0; end

163