Volume 9 Number 2 Jun 2016 Published by

Airiti Press Inc. Taipei office: 18F, No. 80, Sec. 1, Chenggong Rd., Yonghe Dist., New Taipei City 23452, Taiwan, R.O.C.

International Journal of Intelligent Technologies and Applied Statistics

Vo l. 9, No. 2 Jun. 2016 ISSN 1998-5010

Copyright © 2016 by Airiti Press Inc. All rights reserved.

To subscribe write to: International Journal of Intelligent Technologies and Applied Statistics (IJITAS), 18F, No. 80, Sec. 1, Chenggong Rd., Yonghe Dist., New Taipei City 23452, Taiwan, R.O.C. Phone: (886) 2-2926-6006, Fax: (886) 2-2923-5151 E-mail: [email protected]

Printed in Taiwan. International Journal of Intelligent Technologies and Applied Statistics

Vol. 9, No. 2 Jun. 2016

CONTENTS

The Impact of Noise Traders on Macrodynamics 91 Panupong Sukkerd

Do Copulas Improve an Eff iciency of Seemingly Unrelated Regression Model? 105 Pathairat Pastpipatkul, Paravee Maneejuk and Songsak Sriboonchitta

Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 123 Montri Singhavara, Aree Wiboonpongse, Yaovarate Chaovanapoonphol and Thaworn Onpraphai

A Cluster Analysis of Bank Lending Behavior by Using Self-Organizing Map: The Case of Japan 145 Satoru Kageyama

Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 153 Giam Quang Do and Chaiwat Nimanussornkul

Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 169 Chanita Panmanee and Aree Wiboonpongse

International Journal of Intelligent Technologies and Appld ie Statt is ics

Vol.9, No.2 (2016) pp.91-104, DOI: 10.6148/IJITAS.2016.0902.01 © Airiti Press

The Impact of Noise Traders on Macrodynamics

Panupong Sukkerd*

Faculty of Economics, Kasetsart University Sri Racha Campus, Chon Buri,

ABSTRACT This study constructs a closed economy dynamic stochastic general equilibrium (DSGE) model with noise traders to study the processes of change which occur throughout stock market and macroeconomy. The study finds that the noise-trader shocks on expected share return directly affect the stock market through their excess demand for shares and indirectly affect the macroeconomy through the changes in behaviors of households and good producers. The positive noise-trader shocks increases share price which implies that share return decreases. Then, the informed traders prefer selling their shares because of over-price share. The noise traders get loss after share price declines. However, they still buy shares at lower price level since they believe an average buying strategy is the best way. For changes in behaviors, the noise traders work more but consume less since they prefer holding more shares while the informed traders behave opposite because they get higher income. Last, an increase in aggregate employment requires the higher investment in order to product more goods.

K eywords: Macrodynamics; Noise traders; DSGE; Stock market

1. Rationale

In the stock market, there are some investors trading shares on imperfect information as if they traded on prefect one. Those investors are called as noise traders and their financial activity is called as a noise trading. That is, their trading causes the excess price volatility and they can gain the benefit from this inefficient market. However, all of them also have to bear risk in order to take a chance to gain the excess return. Some noise traders are definitely able to beat the market and get extra return while others become losers. Furthermore, the volatility also has the harmful effect on the informed traders, investors rationally trade shares, because the change in share price directly affects to their return. Those investors may not invest in the firms if share price is so volatile. However, there are possibilities that the behavior of noise traders may affect not

* Corresponding author: [email protected] 92 Sukkerd only the stock market, but also real markets such as labor market and good market. This ambiguity motivates us to construct the dynamic macroeconomic model with stock market to answer whether noise trading affects the real sector as well as how interaction between variables in stock market and macroeconomy. The impulse response of dynamic stochastic general equilibrium (DSGE) model is a powerful instrument to reveal the change in economic behavior so we create our stylized model to generate the macrodynamic results. The results help us to understand the different trading behaviors between informed and noise traders. Furthermore, those results can reveal the difference in their decision of consumption and work which affect to good producers as well. This study describes three sectors in model as households, financial intermediary and good producers. Then, we will simulate the model by differing the noise-trader shocks on the dynamic macroeconomy. Last, we find the conclusion on the impact of noise-trader shocks on macrodynamics.

2. Noise trader

The conventional wisdom mentions that the stock price always projects the intrinsic value of stock because it reflects all information of the stock every times, thus all rational investors cannot receive any excess returns from trading the stock [8]. The previous principle is known as the random walk theory. This theory is challenged the alternatives. One of alternatives is the theory of noise trading which argues some investors are possible to make decisions under imperfect information and then those investors usually trade securities irrationally. Furthermore, they sometimes beat the market by receiving the excess returns from market inefficiency. However, their short-term speculative activities also cause a negative externality by magnifying the market volatility. The aspect of noise traders causing volatility is initially proposed by Black [2]. This approach argues that there are noise traders surviving in financial market [2, 7, 9, 18]. Those investors irrationally trade shares because they cannot access perfect information [2, 14, 17]. Their trading strategy causes price distortion and noise-trader risk. Moreover, this risk impedes the price adjustment because it limits the ability of arbitrage. This inefficiency allows noise traders to seek the extra gain by bearing their risk [3, 4, 6, 14]. Noise traders are able to get more return than other types of investors [6], but this statement is inconsistent [15]. However, an exact advantage of noise traders is that they help increase the market liquidity [2, 3].

3. Model

There are three sectors in this model; households (financial intermediary owners and noise traders), financial intermediary (informed traders) and good producers. The Impact of Noise Traders on Macrodynamics 93

3.1 Households

First, we assume the amount of households is a unity and the households are heterogeneous. There are 1 – SN financial intermediary owners and SN noise traders in our economy. The variables relating to financial intermediary owners and noise traders are indexed by the subscript I and N, respectively.

3.1.1 Financial intermediary owners

The financial intermediary owners supply their labor N I ,t in good producers for receiving the wage W t as compensation. They spend their wealth for consumption CI ,t as well as investing in financial intermediaries F t. Then, financial intermediaries pay F the fund return Rt and an exogenous fraction ω of their profits Πt back to the owners.

Last l y, T I,t states the transfer payments. The owners’ period utility function is given by,

U I ,t = ln C I ,t + xln(1 – N I ,t ) (1)

Their budget constraint is formed as,

F CIt,+ F t = WN t It, + R t−− 11 F t +()1 −ω Π+tT It, (2)

where x is the labor weight in utility. They try to maximize their lifetime utility (1) subject to their budget constraint (2). Then, we get the optimal conditions for consumption, labor supply and amount of fund as follows,

1 Λ=It, (3) CIt,

CIt, Wt = χ (4) 1− NIt,

ΛIt,1+ F β ERtt=1 (5) ΛIt,

where β is the discount factor, Et denotes the rational expectation operator and the Lagrange multiplier on owners’ budget constraint is presented by ΛI ,t. 94 Sukkerd

The shadow price is determined by marginal utility of consumption presented by the Equation (3). An Equation (4) is an intra-temporal condition which represents how financial intermediary owners choose between work and consumption in the same period. We call an Equation (5) that an inter-temporal condition or Euler equation which ensures the owners will smooth their life-time consumption.

3.1.2 Noise traders

The noise traders work in good producers for getting wages in order to consume goods and invest themselves in shares S N ,t for saving their wealth. Notice that they trade the shares themselves instead of indirectly trading through financial intermediaries. Their trading strategies rely on imperfect information so they sometimes trade shares irrationally. We apply how noise traders hold shares from the Gertler and Karadi [11]. They gain utility from the amount of consumption C N ,t and disutility from the hours of work N N ,t so the utility function is formed by,

U N ,t = ln C N ,t + xln(1 – N N ,t ) (6)

Their budget constraint is defined as,

SS CNt,,,+ P t S Nt + TF Nt = W t N Nt, ++() P t DIV t S Nt,1 + T Nt , (7)

The noise traders’ transaction fee function is approximated by the multiples of the rate of transaction fee adjusted by a scaled factor and the amount of transaction,

2 SS (8) TFNt,,=Η−φ ( PSt Nt PS t )

S where Pt stands for the share price, DIV t represents dividends from holding shares, φ is the rate of transaction fee, H is a scaled factor for adjusting an transaction fee base with the frequency of trading per quarter and S stands for the average amount of shares which is normalized to be one without loss of generality. They maximize their lifetime utility (6) subject to their budget constraint (7). Then, we get the optimal conditions for consumption, labor supply and share as follows, The Impact of Noise Traders on Macrodynamics 95

1 Λ=Nt, (9) CNt,

CNt, Wt = χ (10) 1− NNt,

 SS1 Λ Nt,1+ S PSt Nt,= PS t +−β ENt,1() Rt+ R t (11 ) 2φΗΛNt,

where ΛN ,t is the Lagrange multiplier on noise traders’ budget constraint, E stands for the noise traders’ subjective expectation operator, S denotes the N ,t Rt definition of gross share return and Rt is the definition of inverted stochastic discount factor. We interpret Equations (9) and (10) to be the same way of Equations (3) and (4), respectively. The Equation (11) represents an optimal share holding condition. This condition reveals that noise traders prefer holding shares more when they expect the excess share return will rise.

3.2 Financial intermediaries

The financial intermediaries are supposed to be the owners’ representative traders to directly trade shares in the stock market. Their trading strategies rely on perfect information so they rationally trade the shares. We can treat financial intermediaries as the informed traders. Their profits are the difference between gains from holding shares and costs from paying fund return. This sector is modified from the Lendvai et al. [13] financial intermediary traders as well as Gerali et al. [10] banks. In this model, the financial intermediaries are represented as the informed traders who trade shares SI ,t on perfect information so they exactly forecast the share return. Our financial intermediaries receive funds from financial intermediary owners at a fund rate and then they use the funds and their net worth to invest in the shares. The financial intermediaries get revenues from receiving share return, but they have to pay the transaction fee TFI ,t as well as the fund return. Therefore, the informed traders’ period profit function is given by,

SF Π=t()P t + DIV t S It,1− − R t −− 1 F t 1 − TF It , (12) 96 Sukkerd

They accumulate their net worth by the follow law of motion,

NW t = (1 – δI )NW t–1 + ωΠt–1 (13)

Their balance sheet identity is given by,

S Pt S It, = NW t + F t (14)

where NW t is the net worth and δI denotes the cost for managing the financial intermediary’s capital (or net worth) position. Additionally, the functional form of informed traders’ transaction fee is similar to the noise traders’ one. They maximize their lifetime profits (12) subject to balance sheet constraint (14). Here, the following problem is not differentiated with respect to the net worth because the retained earnings are given by an exogenous fraction of the profits. Then, we get the optimal conditions for fund and share,

ΛIt,1+ F µβt= ER It, t (15) ΛIt,

 SS 1 ΛIt,1+ SF PSt It,= PS t +−β EIt,1() Rt+ R t (16) 2φΗΛIt,

where EI,t stands for the informed traders’ subjective expectation operator. The Equation (15) states the shadow price of fund equal to the stochastic discount fund return. We interpret Equation (16) to be similar to Equation (11).

3.3 Traders’ expected stock return

In our model, we have two types of traders who survive in the stock market so we must describe how they are different. Informed traders are very smart to expect a share return because they have a rational expectation. Noise traders have noisy expectation so they cannot exactly forecast a share return. We assume the source of mis-forecast is the noise-trader shocks νt which satisfy the white noise processes without persistence. Therefore, we can define the relationship between two types of traders’ subjective expectations on share return and rational expectation as follow equations, The Impact of Noise Traders on Macrodynamics 97

ΛΛIt,1++SSIt,1 ββEI,1t REttt++= R 1 (17) ΛΛIt,,It

ΛΛ Nt,1++SSNt,1 νt ββEN,1tt R++= Ett Re 1 (18) ΛΛNt,,Nt

Then, we can use Equations (17) and (18) to rewrite the optimal share holding conditions (11) and (16).

3.4 Good producers

We assume that good producers are established by using the initial funds from issuing the shares. They apply the labors and capitals as inputs to produce goods and contribute them to be the consumption of financial intermediary owners and noise traders as well as the retained earnings for re-investment in them. Then, the profits of good producers are paid as dividends to the traders.

The good producers are representative and competitive. They produce output Y t by using technology At, capital K t, and worker’s labor N t as inputs so their dividends (profits) are the difference between the value of output and the cost of inputs. The period dividend function is given by,

DIV t = Y t – W tN t – I t (19)

Their Cobb-Douglas production function and law of motion of capital stock are defined by the follow equations,

1−αα (20) Yt= AK tt N t

K t+1 = (1 – δ)K t + I t (21)

where I t denotes investment, δ states the depreciation rate of capital stock and α denotes the labor share in production as well as 0 < α < 1. They try to maximize their discount future dividends (19) subject to the production function (20) and capital accumulation (21). Then, we get the optimal conditions for investment and capital stock as well as labor demand, 98 Sukkerd

Λtt++11Y QEtt = 1β  ()11 − αδ +−() Qt+1(22) ΛttK +1

Λtt++11Y (23) QEtt=β()11 − αδ +−()Qt+1 ΛttK +1

Yt (24) α = Wt Nt

An Equation (22) tells us that the shadow price of investment in competitive firm is unity. The optimal capital stock condition is denoted by an Equation (23) which shows the present price of a unit of capital equals to the discount future gains consist of marginal product of capital and the value of capital after depreciation. An Equation (24) is the optimal labor demand condition. This condition interprets firms will hire the amount of labor until the marginal product of labor equivalents to the wage rate.

3.5 Aggregations

In this model, there are five types of aggregations consist of aggregate labor, aggregate share, aggregate transaction fee, aggregate consumption and aggregate welfare. Their definitions follow equations,

N t = sNN N ,t + (1 – sN)N I ,t (25)

St = sNSN ,t + (1 – sN)SI ,t (26)

TFt = sNTFN ,t + (1 – sN)TFI ,t (27)

Ct = sNC N ,t + (1 – sN)C I ,t (28)

V t = sNV N ,t + (1 – sN)V I ,t (29)

3.6 Closing the model

In the equilibrium of closed economy, the output is divided between consumption and investment. Therefore, the resource constraint of this economy is, The Impact of Noise Traders on Macrodynamics 99

Y t = Ct + I t (30)

The market clearing in share requires the gross amount of shares has to be normalized to unity because it gets rid off the effect of changing in supply of shares. Thus, the total share is,

St = 1 (31)

Additionally, we have two definitions of transfer payment for each type of households. The transfer payment is the multiple of each type of proportion of households and previous-period transaction fee. Therefore, both equations are similarly written as follows,

T I ,t = (1 – sN)TFt–1 (32)

T N ,t = sNTFt–1 (33)

3.7 Exogenous processes

The noise or non-fundamental shocks is only one exogenous shocks in our model. The exogenous law of motion is given by,

νt = εν,t (34)

where εν,t is the normally distributed white noise processes.

3.8 Parameterization

Most of parameters are picked up from Tanboon [16] while others follow Gerali et al. [10], Khanthavit [12], Alp and Elekdag [1], Chen et al. [5] and Lendvai et al. [13] as well as our calculation. Additionally, we categorize these parameters into three groups. The first group of parameters is about the households. The discount rate β is set to 0.9926 which implies the long-term real risk-free interest rate of annualized 3 percent. Labor weight in utility x is set to be 1 which means the disutility of labor is no scaling.

Fraction of noise traders sN is assumed to be 0.50 in the benchmark case which tells us that the half of population is noise traders. Average amount of securities in each-typed traders’ portfolio S is normalized to be one. The rate of transaction fee is set to be 0.0025 because the fee is charged about 25 basis points of trading value. We assume 100 Sukkerd all traders trade the shares two times a month which implies the scaled factor H is 6, following Lendvai et al. [13]. Furthermore, the standard deviation of noise- trader shocks σν is set to be 0.0407, follows the standard deviation of share return on Thailand Stock Exchange in Khanthavit [12]. The second group of parameters relates to the financial intermediaries. We assume the fraction of retained earnings of profits ω to be 0.09. The cost for managing the financial intermediary’s capital position δI is set to 0.1049, following Gerali et al. [10]. The last group involves the firms. Steady-state total factor productivity A is normalized to be one. We choose the value of labor share in production α to be 0.60 which value is between 0.50 in Chen et al. [5] and 0.70 in Tanboon [16]. Capital depreciation rate δ is set to 0.0105 which equivalent to the average depreciation rate of annualized 4.2 percent, following Tanboon [16]. Then, this model is simulated by using Dynare software. Furthermore, we choose the second-order perturbation methods and 10,000 periods of simulations with the random draws of the noise-trader shocks in order to report the results.

4. Simulations and results

We simulate our model by introducing the noise-trader shocks. The size of shocks is set a one standard deviation in quarter. All of impulse responses are displayed in percentage change. The timing of impulse response is measured in quarterly frequency. Figure 1 shows the impulse response to noise-trader shocks which is presented by the solid line and describe the impact as following. The positive noise-trader shocks cause sharply increase in the noise traders’ demand for shares. This excess demand pushes share price up. An increased price implies that share return is gone down. The informed traders perceive the reduction in share return so they are willing to sell their shares. However, the share price cannot go up overtime. The noise traders perceive that they buy the share at the high price after declining in the price of share. Nevertheless, they still buy shares at lower price for getting the lowest average price because they thought that it is the bes t wa y. In the initial stage, noise traders supply more labor but less consume in order to save more money for buying shares. Their utility begins with the negative percentage. Then, they believe that they are wealthy because there are many shares in their portfolios and the share price is getting higher. Therefore, they desire to supply labor less but consume more. These processes cause their utility to rise to be positive. However, their utility is slummed down because they turn to work more but less consume after they knew that they misunderstood about their wealthy. So, their utility is declined later. The Impact of Noise Traders on Macrodynamics 101

Figure 1. Impulse response to a one standard deviation of noise-trader shocks. 102 Sukkerd

The financial intermediary owners initially work less but consume more because they receive the higher return from the financial intermediaries. So, their utility is positive in the beginning period. Then, their labor sharply increases in short time because good producers’ labor demand pushes the wage up. Thus, their utility is gone down to be negative. Lastly, their consumption, labor and utility convert to zero. For the good producers, the aggregate labor is positive in the beginning. This requires positive investment in order to optimize the production. The output is positive as well. Next, the aggregate labor and investment are sharply declined and become the negative so the output turns to be negative. Then, all of those variables adjust themselves to zero percentage change in the long-run.

5. Conclusions and limitations

5.1 Conclusions

This paper constructs the DSGE model to study the impact of noise-trader shocks on Macrodynamics. We employ the most of parameters from Tanboon [16], Gerali et al. [10], Khanthavit [12], Alp and Elekdag [1], Chen et al. [5] and Lendvai et al. [13] while some unknown parameters are ourselves calculated. The second-order perturbation method is chosen to simulate the model’s results. The model helps us to understand how the processes of change which occur throughout stock market and macroeconomy when they face the noise-trader shocks. The study demonstrates that the noise-trader shocks on expected share return directly affect the stock market through their excess demand for shares and indirectly affect the macroeconomy through their changes in households’ behavior. For direct effect, the optimistic noise traders want hold more shares for speculation. This higher demand increases share price which implies that share return decreases. Now, the informed traders know that shares are over price so they prefer selling their shares. This supply declines the share price slowly which make the noise traders get loss. However, they still prefer buying cheaper shares because they believe the average buying is best practice at that time. For indirect effect, the positive noise-trader shocks affect the behavior of households as follows. Initially, the noise traders work more but consume less since they prefer more saving to hold more shares while the informed traders behave opposite because they have higher income from financial intermediaries. However, the aggregate labor increases which requires more investment in order to product output more. The policy implications suggest that social planners can understand the impact of noise traders on stock market and real sectors. They can forecast the result of noise-trader shocks and find the instrumentality or mechanism to improve the economic efficiency and social welfare. For investors, the investing in stock The Impact of Noise Traders on Macrodynamics 103 market through financial intermediaries such as mutual funds may be better than households directly invest in the stock market because the mutual funds has more experts and perfect information.

5.2 Limitations

The model of this study is based on the real business cycle (RBC)’s simplifying assumptions. We may relax the assumptions in the future study. The investors are treated to be only two types of traders, informed and noise traders. We may add the other types of traders such as the liquidity traders. Furthermore, the traders are treated to invest only in the stock market. Actually, the traders can allocate their wealth to several markets such as bond market, money market and real estate.

References

[1] H. Alp and S. Elekdag (2012). Shock Thera py: W hat Role f or Thai M onetary P olic y? International Monetary Fund, Washington, DC.

[2] F. Black (1986). Noise, The Journal of Finance, 41, 528-543.

[3] R. Bloomfield, M. O’Hara and G. Saar (2009). How noise trading affects markets: An experimental analysis, Review of Financial Studies, 22, 2275-2302.

[4] G. W. Brown (1999). Volatility, sentiment, and noise traders, Financial Analysts J ournal, 55(2), 82-90.

[5] V. Chen, B. Cheng, G. Levanon, A. Ozyildirim and B. van Ark (2012). Projecting global growth, Economics Program working paper, The Conference Board, New York.

[6] J. B. De Long, A. Shleifer, L. H. Summers and R. J. Waldmann (1990). Noise trader risk in financial markets, Journal of Political Economy, 98, 703-738.

[7] J. B. De Long, A. Shleifer, L. H. Summers and R. J. Waldmann (1991). The survival of noise traders in financial markets, The Journal of Business, 64, 1-19.

[8] E. F. Fama (1965). Random walks in stock market prices, Financial Analysts J ournal, 21, 55-59.

[9] T. Foucault, D. Sraer and D. J. Thesmar (2011). Individual investors and volatility, The Journal of Finance, 66, 1369-1406.

[10] A. Gerali, S. Neri, L. Sessa and F. M. Signoretti (2010). Credit and banking in a DSGE model of the Euro area, Journal of M oney Credit and Banking, 42(s1), 107-141. 104 Sukkerd

[11] M. Gertler and P. Karadi (2013). QE 1 vs. 2 vs. 3 ...: A framework for analyzing large-scale asset purchases as a monetary policy tool, International Journal of Central Banking, 9(S1), 5-52.

[12] A. Khanthavit (2010). Regime-varying behavior of stock returns in Thailand’s stock market and its implications toward asset allocation and risk management strategies, Manuscript, Faculty of Commerce and Accountancy working paper, Thammasat University, Bangkok, Thailand.

[13] J. Lendvai, R. Raciborski and L. Vogel (2013). Macroeconomic effects of an equity transaction tax in a general-equilibrium model, J ournal of Economic D ynamics and Control, 37, 466-482.

[14] A. Shleifer and L. H. Summers (1990). The noise trader approach to finance, The Journal of Economic Perspectives, 4(2), 19-33.

[15] R. W. Sias, L. T. Starks and S. M. Tiniç (2001). Is noise trader risk priced? J ournal of Financial Research, 24, 311-329.

[16] S. Tanboon (2008). The Bank of Thailand structural model for policy analysis, Economic Research Department working paper, Bank of Thailand, Bangkok, Thailand.

[17] B. Trueman (1988). A theory of noise trading in securities markets, T he J ournal of Finance, 43, 83-95.

[18] B.-C. Witte (2013). Fundamental traders’ “tragedy of the commons”: Information costs and other determinants for the survival of experts and noise traders in f inancial markets, Economic M odelling, 32, 377-385. International Journal of Intelligent Technologies and Appld ie Statt is ics

Vol.9, No.2 (2016) pp.105-122, DOI: 10.6148/IJITAS.2016.0902.02 © Airiti Press

Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model?

Pathairat Pastpipatkul, Paravee Maneejuk* and Songsak Sriboonchitta

Faculty of Economics, Chiang M ai University, Chiang M ai, Thailand

ABSTRACT The conventional SUR model has a strong assumption of normally distributed residuals which might not be realistic. What this paper suggests is to take an advantage from Copulas approach in order to relax this normality assumption. Therefore, we introduce the Copula- based SUR model as an alternative to the conventional SUR model. The performance and accuracy of the proposed model are evaluated through the simulation study. We then apply the Copula-based SUR model to the real data set of Thai rubber and compare the estimation result with the result obtained from the conventional SUR model. The results of this paper could support our suggestion that the Copulas can be used appropriately to relax the assumption of normality of residuals in the conventional SUR model.

K eywords: Seemingly unrelated regression; Copula; Model comparison; Thai rubber

1. Introduction

What we are interested here is the seemingly unrelated regression (SUR) model, which was first introduced by Arnold Zellner in 1962. The model is used to construct a system of different linear equations where all equations are estimated simultaneously [9]. By its properties, the SUR model could gain more efficiency in estimate the model, since it can combine information from the different equations by letting errors terms of each equations be related. But apart from its strength, the SUR model also has a strong assumption of normally distributed error terms. Roughly speaking, it is one of the important assumptions for the regression analysis, in which the residual of each equation in the system is normally distributed. Let’s think about the case that we have non-normal residuals, should we continue using the conventional SUR model? This question gives rise to a possibility that the SUR model might be improved

* Corresponding author: [email protected] 106 Pastpipatkul, Maneejuk and Sriboonchitta by eliminating this assumption. What we suggest in this paper is to consider the Copula approach. It is useful due to its ability to joint the different marginal distributions of residuals. In particular it makes the model more realistic and far from the assumption of normality. Therefore in this paper, we are going to relax the normality assumption of error terms in the SUR model by applying Copula approach to the model. Each equation in the SUR model will be allowed to have different marginal distributions of residuals, and not need to be normally distributed. For instance, the margin of one equation can be considered to have Student t distribution while another equation whose margin is normally distributed. Then Copula function will link them as a joint distribution [3, 5]. Thus, this paper would like to introduce the Copula-based SUR model as an alternative to the conventional SUR model. In order to define how precise of our suggested model, the model will be evaluated its performance and accuracy through a simulation study, before applied to the real data set. Then the evaluation results will be discussed in the second to the last section of this paper. For an overview of this paper, we begin with our intention to suggest the Copula-based SUR model in the introduction. Next, Section 2 is about a discussion on methodology followed by Section 3 which is about the estimation of the Copula-based SUR model. We then try to evaluate an accuracy of our proposed model, so we take into account a simulation study as shown in Section 4. The next to the last section is about application. The data set related to Thai rubber is used as an application since it is one of the most important agricultural exportations of Thailand. And also the nature of Thai rubber market, it is determined by both demand and supply of the rubber which can be constructed by a system equations which are a character of the SUR model. Then, we will talk about estimated result obtained from the Copula- based SUR model and compare the result to the conventional SUR model. And the last section is a conclusion.

2. Methodology

The method mainly used in this paper is about the seemingly unrelated regression (SUR) model which will be described in Section 2.1. Then, we will devote the following part to the Copulas families applied in this paper. After that we will introduce our proposal “the Copula-based SUR model” in the estimation. As we mention in the introduction, we will make a comparison between our proposed model and the conventional SUR model in the application. For more accuracy we decide to estimate the conventional SUR model using all well-known possible estimation method where the Bayesian is one of our concerns. It is believed that there is some complicated procedure for the Bayesian. Hence, we will refresh our knowledge of the Bayesian inference on SUR model by recalling its concept in the last part of this section before devote the rest of this paper to the Copula-based SUR model. Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 107

2.1 Seemingly Unrelated Regression (SUR) model

The SUR model was first introduced by Zellner [9] as a generalized system of linear regression. Considering the structure of SUR model, it consists of several regression equations, in which they are allowed to have their own dependent variables. Here we have M regression equations where each has N independent variables; the system of M equations can be shown as follows.

′ yxtt,1= ,1βε 1 + t ,1 … ′ yxtM,,= tMβε M + tM ,

The system above can be written as a vectorial form as follows:

Y t = X t β + εt (1)

Let Y t be a vector of dependent variables, yt,i, i = 1, …, M. A matrix of independent variables (regressors) is denoted as X t, where xt,i j, i = 1, …, M, j = 1, …,

N. β denotes a matrix of an unknown parameters (regression coefficients), and εt 2 is a vector of the error terms, where εt = [εt,1, εt,2, …, εt,M]′ and εt,i ~ N(0, σi ), i = 1, …, M. The important assumption of the SUR model which let it gain the efficiency of estimation is that the error terms are assumed to correlate across equations [1, 2].

Thus, it can be estimated jointly, E[εiaεib|X ] = 0; a ≠ b whereas E[εiaε jb|X ] = σij [6 ]. T h e model allows non-zero covariance between the error terms of different equations in the system. Considering a variance-covariance matrix for M equations, it can be written as follows.

σσ σσ 11 II 1M 11  1M  '  Γ = ∑⊗I =  , where Σ=()εε  tt  σσ σσ M 1 IIMM M 1 MM and I is an identity matrix. Hence from Equation (1), the error terms of the SUR model is assumed to be εt ~ N(0, Γ), and this system equation can be estimated by

-1 -1 -1 βsure = (X′Γ X ) X′Γ Y 108 Pastpipatkul, Maneejuk and Sriboonchitta

2.2 Copulas

A linkage between the marginal distributions was first introduced by Sklar in 1959 [8] called Sklar’s theorem. Then, it was well described in Nelson [5] as the dependence in Copula. Following Nelson [5], this paper applies the most used copulas of two important classes of Copula, namely Elliptical Copulas and Archimedean Copula to model the dependence structure of the SUR model. Here, the Elliptical class consists of the symmetric Gaussian and t-Copula. The Archimedean Copula consists of the asymmetric Fank, Clayton, Rotated Clayton (90 and 270 degrees), Gumbel, Rotated Gumbel (90 and 270 degrees), and Joe.

Elliptical Copulas

(1) Gaussian Copula Considering the case of n-dimensional, the Gaussian Copula can be defined by [7].

Σ −− n -11- Cu(1 ,..., un )=ΦΦ n (11 (u ),..., Φnn ( u )) (2)

Σd where Φd is n-dimensional standard normal cumulative distribution and Σn is a variance-covariance matrix. The density of the Gaussian Copula is given by

Φ-−1 11()u −−-111 - −- 1 −- 1 = Σ Φ Φ ⋅Σ − ⋅ cu(1 ,..., un ) ( detn )( 11 (u ) nn ( u )) ( n I )  (3) 2  −-1 Φ ()u 1 n

(2) T-Copula T-Copula is one of the copulas in the Elliptical Copulas. It has a second parameter [7] and degree of freedom which is denoted by ν. In the case of n-dimensional we can def ined the t-Copula by

−−-11- tuv()1 tuvn ( ) c( u ,..., u ) =  f() x dx (4) 1 n ∫∫−∞- −∞- tv1 () where fx() is n-dimensional t-density function with degree of freedom ν tv1 () −-1 and tv is the quantile function of a standard univariate tν distribution. In the estimation, the density of t-Copula is def ined by Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 109

f( tu−−-11 ( ),.... tu- ( )) t = vP,1 v v n (5) cuvP,1( ,..., un ) n −-1 Πi=1 ft vv( ( u i ))

where f ν,P is the joint density of a td(ν,0,P) distributed random vector and P is the correlation matrix implied by the dispersion matrix Σ.

Archimedean Co pula

In 1986, Genest and MacKay have provided the general form of the Archimedean Copula which can be defined by

−-1 n Φ( Φ (u11 ) + ... +Φ ( un )) , ifΣii= Φ ( u ) ≤Φ (0) Cu(1 ,..., un ) =   0,otherwise where Φ is a strictly decreasing function. Let’s consider the density function of the Copulas in this Archimedean class consisting of Clayton, Gumbel, Frank, Joe, and the rotated cases, i.e. Rotated Clayton and Rotated Gumbel (90 and 270 degrees).

(1) Clayton Copula The Clayton Copula is usually referred to in the bivariate case [7]. It is a Copula of the multivariate Pareto and Burr distributions. Hofert et al. [3] have proposed the density function of Clayton Copulas as follows.

nn−1 C =++θ −+-(1θθ ) −-nn( +1/ ) cuuθθ(11 ,...,n )∏∏ ( k 1)( u jn ) (1 tuu ( ,..., )) (6) kj=01=

θ controls the degree of dependence, 0 < θ < ∞. If θ → ∞ the Clayton Copulas will converge to the monotonicity Copula with positive dependence. But if θ = 0 it will corresponds to independence.

(2) Gumbel Copula The density function of Gumbel Copula can be defined by

n −- θ −1 ∏ ( logu j ) Gn j=1 G1/θ cuuθ(1 ,...,n )= θ Cuuθ (1 ,..., n ) n Ptuu nn,1αθ( (( ,..., ) ) (7)

tuθ ((1 ,..., unj )∏ u j=1 110 Pastpipatkul, Maneejuk and Sriboonchitta

αα GG nk−−nk−−−−−− n n j j nnnn!!!! kk kjkjkjkj    αα     nn j j G nG k ααααααααGG()(1)()(1)=−Σ=−Σ=−Σ=−Σ- nknk n n α α α α jsn jsn (,)(,) (,)(,) jSk jSk j j=Σ=Σ=Σ=Σkk (( − −- − −1) 1)nn j j where Pxn,1α ()= Σk= αα nk ( ) x, and nknknknk()(1)()(1)nn=n=n k= k= k k snsn (,)(,) (,)(,) jSk jSk j j jj=j=1j=1=11        (( 1) 1) kkkk!!!! jnjnjn jn        in which kn∈{}1,..., , and s, S are the stirling numbers of the first and second kind.

(3) Frank Copula The density function of Franke Copula can be defined by

θ exp(−Σ-θ n u ) F = dF−1 jj=1 (8) cuuθθ(1 ,...,n ) ( )Lihuu−−-(nn 1) ( ( 1 ,..., )) F 1−− exp(-θ ) huθ (1 ,..., un )

n Fn=−−−-θ 1 −−-θ where huθ (1 ,.., unj ) (1 e )∏{} 1 exp(u ) j =1

(4) Joe Copula The density function of Joe Copulas is defined by

n n n θ exp(−Σθ u ) −ΣFθ θdF−11 jj=1 F θ dF−1 exp(cuu(jj= ,...,1u ) )= ((1− un ) Lihuu ( ( ,..., )) n cuu( ,..., )= ( )Lihuu ( (θ ,..., )) θθ1 n∏exp(−Σθθ j u ) −−(nn 1) 1 exp(JF −Σθ u ) θθ1 nF−−(nn 1) 1 dF−1 F F 1−−= exp(θjj )=1 dF−1 huhu( ,..., ,... u ujj=1 ) (9) = J d−=1 j 1 JJα θθ 1 nn 1−− exp(cuuθθ(θ1 ) ,...,n ) ( )cuuLihuuhuθ −−((nn 1)cuu ,..., ,... (θθ( u (1 ,..., ) 1 ,...,= θn ) )) ( )(1Lihuu−−−(nnhuP 1) ( )) ( 1 ,..., ( )) ) −−θ θ11,nnn JJF θαF 1 exp( ) 1huθ−−( exp(1 ,..., ,...θ u u )n ) 1− huhuθ (1 ,...,,... u un ) −Σθ n θθ11nn F θ dF−1 exp(jj=1u ) nn cuuθθ(1 ,...,n )= ( )Lihuu−−(nn 1) ( ( 1 ,..., )) 1−− exp(θ ) huFJJ( ,..., u )= −− θ J nJ−1 k where huθθ (111 , ... unjnjn )∏ (1 (1u ) ) and Pxn,0α ()= Σk= a nk (α ) x jj=11

(5) Rotations of Copulas Here we also consider the rotations of Clayton Copula and Gumbel Copula by 90 and 270 degrees. We are interested in the rotated Copulas since we hope to obtain tail dependent copulas for all possible corners [7]. To be concrete, we consider the case of bivariate copulas where the respective density function can be defined by

c90(u1, u2) = c(1 – u1, u2) and c270(u1, u2) = c(u1, 1 – u2) for the any rotated copulas (90 and 270 degrees), respectively. In addition the corresponding h-functions are given by

h270(u1, u2) = h(u1, 1 – u2), and h90(u1, u2) = 1 – h(1 – u1, u2). Therefore, the derivatives

of h-functions with respect to u2 can be obtained as ∂2h90(u1, u2) = -∂2h(1 – u1, u2) and

∂2h270(u1, u2) = -∂2h(u1, 1 – u2), and similar for the other derivations [4].

2.3 Bayesian inference on SUR model

Even if the Bayesian approach is not our mainly concerned in this paper, we still devote this section to the Bayesian inference on a conventional SUR model since it is going to be compared to our proposed model in the application part (see Section 5). According the Bayes’ theorem, the posterior estimation can be formed by Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 111

Posterior probability ∝ likelihood × Prior probability pr(θ, Σ|Y, X )α L(Y|θ, Σ, X )pr(θ, Σ)

In this study, we apply the Gibb sampling to sampler the parameters in the posterior probability, which are θ and Σ. The initial starting values for these parameters are computed from the least square method. Then the obtained parameters, θ0 and Σ0 become the starting values for sampling the parameters. To estimate the parameters, the procedures are the following: First, drawing θ j from pr(θ|Σ j–1, Y, X ). Next, drawing Σ j from pr(Σ|j β j–1, Y, X ). These procedures would be completed by a Gibb burn-in 5,000 times in order to eliminate the non-stationary distribution of the Markov chain. Thereafter, these new parameters,  and Σ , are θ1 1 used as the starting values, then repeat another iteration 50,000 times to obtain a final of random draws.

3. Estimation of the Copula-based SUR model

In this paper, we consider the SUR model which consists of two equations (M = 2) in order to represent demand and supply equations of Thai rubber. Thus, the bivariate Copula with continuous marginal distribution is conducted in the estimation. Before estimation, we begin with checking the stationary of the data series using the Augmented Dickey-Fuller test. Next, the estimation procedures of Copula-based SUR model involve four steps. First, we estimate the conventional SUR model using a maximum likelihood technique to obtain the initial values. Second, we construct the SUR Copula likelihood using the chain rule, here we have

∂∂22∂∂22 = Fu(1Fu , u( 2 )1 , u 2 ) CFu (11 CFu ( ( ),11 ( Fu11 ( ), Fu ))11 ( )) (10) ∂∂∂∂ ∂∂∂∂ uu12uu12 uu1uu212 = f112( uf )112( f u ( u ) 2 f ) cFu ( u( 2 11) (cFu ( ), 11 ( F 2 (), uF 2 )) 2 ( u 2 ))

From Equation (10), u1 and u2 are the marginal distributions which can be Guassian or Student’s t distribution. f 1(u1) and f 2(u2) are normal functions of demand and supply equations, respectively. Density function of Copulas is denoted as c(F1(u1), F 2(u2)). In this study, we are interested in the following Copula families, i.e. Gaussian, T, Clayton, Gumbel, Frank, Joe, and the rotated Copulas in order to construct the joint distribution function of a bivariate random variable with the univariate marginal distribution. Then, we transform Equation (10) using a logarithm, and we get

T lnL=Σ+id=1 (ln lX (θθ 1 ) ln lX (2 s ) ++ ln fu11 ( ) ln fu 2 ( 2 ) + ln cFuFu (11 ( ), 2 ( 2 )) (11 ) 112 Pastpipatkul, Maneejuk and Sriboonchitta

where lnl(θ2|X d) and lnl(θ2|X s) are the logarithm of the likelihood function of demand and supply equations, respectively. The logarithm of the likelihood function can be def ined by

TT1 ln LL =−−Σ= - ln(2π ) ln( ) (YX −ββ )′ ( YX − ) (12) 2 22( Γ )

Moreover, the last term of (11), ln c(F 1(u1), F 2(u2)) denotes the bivariate Copula density which is assumed for Gaussian, T, Clayton, Gumbel, Frank, Joe, and the rotated Copulas (See Section 2.3). Here, we employ the maximum likelihood estimation (MLE) to maximize the SUR Copula likelihood function (12) in order to obtain the final estimation results for the Copula-based SUR model.

4. Simulation study

In this section, we carry out a simulation study to evaluate performance and accuracy of the Copula-based SUR model. In this simulation study, the Monte Carlo simulation is conducted to simulate the dependence parameter for the families of Copulas that we concerned, i.e. Gaussian, Student-t, Clayton, Frank, Joe, Gumbel, Clayton 90 degree, and Gumbel 90 degree. We set the true parameters value for these copula functions as shown in Table 1 ~ Table 3. To model the error term

(U t) of two equations, we assume these errors to follow a Normal and Student-t distributions. The quantile function is used to converse the simulated uniform u and ν to be (U t) where U t ~ N(0,1) for normal margins and U t ~ t( 0,1,V ) for Student-t margins. Finally, we construct the Copula-based SUR model using the specified parameter from the Table 1 ~ Table 3 and obtained the simulated dependent and independent variables for our model. In this section, we considered three possible cases of simulation studies shown as follow:

Case 1: Copula-based SUR with Normal Margins. Case 2: Copula-based SUR with Student-t Margins. Case 3: Copula-based SUR with Normal-Student-t Margins.

Table 1 ~ Table 3 show the result of the Monte Carlo simulation investigating the maximum likelihood estimation of the Copula-based SUR model. We found that the mean parameters of the three cases are very close to their true values with low standard derivations -they are reasonable and acceptable. Table 1 ~ Table 3 also demonstrate the accuracy of estimation. Overall, the Monte Carlo simulation suggests that our proposed model, the Copula-based SUR model, is quite accurate. Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 113

Table 1. Simulation results of Copula-based SUR with Normal Margins.

True Estimate True Estimate Copula SD Copula SD value value value value Intercept1 1 1.2181 0.1004 Intercept1 1 1.0673 0.0905 Beta1 2 2.1619 0.0963 Beta1 2 2.1215 0.0905 Sigma1 1 1.0226 0.1709 Sigma1 1 0.9308 0.0633 Gaussian Intercept2 4 4.3305 0.0701 Student-t Intercept2 4 4.1451 0.1794 Beta2 -2 -2.2342 0.2443 Beta2 -2 -1.9238 0.1533 Sigma2 2 2.2633 0.2014 Sigma2 2 1.8747 0.1256 Copula 0.5 0.4884 0.1566 Copula 0.5 0.4315 0.0978 Intercept1 1 1.2891 0.0967 Intercept1 1 0.9428 0.0995 Beta1 2 2.0998 0.0671 Beta1 2 2.0006 0.0356 Sigma1 1 0.959 0.0631 Sigma1 1 1.0127 0.0651 Joe Intercept2 4 4.1532 0.2139 Clayton Intercept2 4 4.0414 0.207 Beta2 -2 -2.3777 0.1361 Beta2 -2 -1.9407 0.099 Sigma2 2 2.154 0.1349 Sigma2 2 2.1013 0.1341 Copula 0.5 2.4221 0.3156 Copula 3 3.2836 0.5391 Intercept1 1 1.2474 0.11 Intercept1 1 0.9387 0.915 Beta1 2 1.921 0.0431 Beta1 2 1.9663 0.0825 Sigma1 1 1.1032 0.0745 Sigma1 1 0.9172 0.0647 Gumbel Intercept2 4 4.5516 0.2148 Frank Intercept2 4 4.1009 0.1602 Beta2 -2 -1.8701 0.0656 Beta2 -2 -2.0832 0.1519 Sigma2 2 2.146 0.1456 Sigma2 2 1.6138 0.113 9 Copula 3 3.8821 0.4223 Copula 2 1.4452 0.6098 Intercept1 1 1.1337 0.1004 Intercept1 1 1.0561 0.0998 Beta1 2 2.0297 0.0566 Beta1 2 2.0459 0.0566 Sigma1 1 1.0135 0.069 Sigma1 1 1.0084 0.0647 Gumbel Clayton Intercept2 4 3.84 0.1843 Intercept2 4 4.0414 0.1973 90 90 Beta2 -2 -1.925 0.1135 Beta2 -2 -1.7782 0.1345 Sigma2 2 1.8471 0.1265 Sigma2 2 1.99 0.1298 Copula -2 -2.169 0.2249 Copula -2 -2.1097 0.398 114 Pastpipatkul, Maneejuk and Sriboonchitta

Table 2. Simulation results of Copula-based SUR with Student-t Margins.

True Estimate True Estimate Copula SD Copula SD value value value value Intercept1 1 1.2032 0.1434 Intercept1 1 0.9772 0.1479 Beta1 2 1.9757 0.1211 Beta1 2 2.0695 0.0984 Sigma1 1 1.5074 0.1065 Sigma1 1 1.4072 0.0997 V1 5 4.5874 0.2148 Student-t V1 5 6.398 8.3677 Gaussian Intercept2 4 4.1935 0.1117 Intercept2 4 3.7814 0.1325 Beta2 -2 -2.0661 0.1008 Beta2 -2 -1.9203 0.0827 Sigma2 1 1.0545 0.0747 Sigma2 1 1.1842 0.0846 V2 10 7.8564 1.5781 V2 10 12.6238 2.8624 Copula 0.5 0.6029 0.0377 Copula 0.5 0.5093 0.0776 Intercept1 1 1.1267 0.113 6 Intercept1 1 0.6669 0.1201 Beta1 2 1.9589 0.0829 Beta1 2 1.9728 0.0585 Sigma1 1 1.2123 0.0868 Sigma1 1 1.244 0.0908 V1 5 5.1424 1.3337 Clayton V1 5 4.5071 1.4512 Joe Intercept2 4 4.1287 0.1129 Intercept2 4 3.7195 0.115 8 Beta2 -2 -1.9817 0.0803 Beta2 -2 -1.9095 0.0617 Sigma2 1 1.1685 0.0837 Sigma2 1 1.1314 0.0836 V2 10 8.7854 1.8007 V2 10 8.933 2.5451 Copula 2 2.0491 0.6999 Copula 3 3.1394 1.0 411 Intercept1 1 1.2538 0.1231 Intercept1 1 1.0706 0.1151 Beta1 2 1.8066 0.1378 Beta1 2 1.9552 0.1204 Sigma1 1 1.2 211 0.0863 Sigma1 1 1.1903 0.9415 V1 5 4.8881 1.1253 Frank V1 5 3.9571 1.5212 Gumbel Intercept2 4 3.81904 0.117 6 Intercept2 4 4.1437 1.0031 Beta2 -2 -2.0273 0.106 Beta2 -2 -1.8837 0.0541 Sigma2 1 1.154 0.0816 Sigma2 1 1.091 0.0814 V2 10 8.9134 2.5142 V2 10 9.5554 3.2751 Copula 3 2.8164 0.4521 Copula 2 1.958 0.6421 Intercept1 1 0.8797 0.091 Intercept1 1 1.1301 0.1215 Beta1 2 2.0868 0.0669 Beta1 2 2.2871 0.1226 Sigma1 1 1.0432 0.0737 Sigma1 1 1.1682 0.0826 Clayton V1 5 5.0142 1.6654 V1 5 4.4435 0.2575 Gumbel 90 Intercept2 4 4.1835 0.0963 Intercept2 4 4.0536 0.1038 90 Beta2 -2 -1.9931 0.0801 Beta2 -2 -1.9568 0.0992 Sigma2 1 1.0087 0.0715 Sigma2 1 1 0.077 V2 10 8.5941 1.0321 V2 10 9.8841 1.0354 Copula -2 -3.7936 0.7461 Copula -2 -2.1097 0.4768 Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 115

Table 3. Simulation results of Copula-based SUR with Normal-Student-t Margins.

True Estimate True Estimate Copula SD Copula SD value value value value Intercept1 1 1.1516 0.116 5 Intercept1 1 0.9981 0.1153 Beta1 2 1.8954 0.0869 Beta1 2 1.8524 0.1047 Sigma1 1 1.1985 0.0847 Sigma1 1 1.3758 0.1245 V1 5 4.9985 1.0264 Student-t V1 5 3.6681 1.0001 Gaussian Intercept2 4 4.1937 0.1813 Intercept2 4 3.9256 0.2351 Beta2 -2 -2.1281 0.1441 Beta2 -2 -2.3317 0.1451 Sigma2 1 1.9497 1.1254 Sigma2 1 0.9585 0.9858 Copula 0.5 0.6342 0.1543 Copula 0.5 0.5014 0.114 3 Intercept1 1 1.1026 0.110 4 Intercept1 1 0.8207 0.1337 Beta1 2 1.9312 0.1023 Beta1 2 1.9732 0.0714 Sigma1 1 1.1 0.0837 Sigma1 1 1.4276 0.105 V1 5 4.2794 0.2455 Clayton V1 5 4.7218 1.4084 Joe Intercept2 4 4.0957 0.1975 Intercept2 4 3.7593 0.2371 Beta2 -2 -1.8596 0.4286 Beta2 -2 -1.8373 0.1449 Sigma2 1 1.5016 0.152 Sigma2 1 2.2468 0.4752 Copula 2 1.7487 0.9865 Copula 3 3.1074 0.4724 Intercept1 1 0.9354 0.1088 Intercept1 1 0.8025 0.1324 Beta1 2 2.0551 0.0668 Beta1 2 2.0331 0.1307 Sigma1 1 1.3692 0.0985 Sigma1 1 1.3265 0.0938 V1 5 4.9951 0.1224 Frank V1 5 5.0003 0.0115 Gumbel Intercept2 4 3.9994 0.1857 Intercept2 4 3.8572 0.1752 Beta2 -2 -2.0321 0.0985 Beta2 -2 -2.1315 0.1737 Sigma2 1 1.2218 0.112 2 Sigma2 1 1.7464 0.1229 Copula 3 3.3422 0.5481 Copula 2 1.6871 0.3776 Intercept1 1 1.2726 0.1322 Intercept1 1 0.9496 0.113 2 Beta1 2 2.5524 0.1315 Beta1 2 2.044 0.0718 Sigma1 1 1.4031 0.0766 Sigma1 1 1.1053 0.0785 Clayton Gumbel V1 5 5.1141 0.9584 V1 5 5.0462 1.5902 90 90 Intercept2 4 4.0015 0.0154 Intercept2 4 4.0585 0.2409 Beta2 -2 -1.5515 0.2089 Beta2 -2 -1.9982 0.1419 Sigma2 1 1.5498 0.0614 Sigma2 1 0.6236 0.4154 Copula -2 -2.1001 0.4879 Copula -2 -2.5617 0.5184 116 Pastpipatkul, Maneejuk and Sriboonchitta

5. Application: Thai rubber market

In this section, we apply our proposed model to the real data set of Thai rubber in order to evaluate its efficiency through the real data. We choose the rubber since it is one of the most important agricultural exportations of Thailand, which creates a large economic value for Thai economy. In practice, the rubber market consists of two main equations, i.e., demand and supply of the rubber. They are used to determine the equilibrium price and quantity for the market. Here, we can apply the SUR model to model these equations simultaneously since it is a system equations, which is a character of the SUR model. Therefore, it is reasonable to use it as an application for the SUR model. We then try to investigate the efficiency of our proposed model by comparing the Copula-based SUR model to the conventional SUR model using this data set. In this application, the conventional SUR model is estimated by (i) Bayesian estimator (ii) Maximum Likelihood estimator (MLE), and (iii) Least Squared estimator (LS), while our proposed model is estimated by MLE (See Section 3). We choose these three estimation methods as they are classical and wildely used to estimate SUR model. However we consider that all of them still be used under the strong assumption -normally distributed error terms- of the conventional SUR model. Unlike our suggested model, we do relax this strong assumption. The MLE and LS are often used to estimate not only the SUR model but also many other models, so we may be familiar with them. Unlike the Bayesian estimation, it is far from those estimation techniques. Roughly speaking, In the Bayesian estimation, we get in the probability density functions (or prior) and get out the posterior density functions, rather than a single point as in the MLE. More discussion about the Bayesian inference on SUR model is proposed in Section 2.3.

Data

The data used in this application is the data set related to Thai rubber consisting of Thai rubber export prices (Pexport), Oil prices (Poil), Malaysian rubber export prices (Pmalay), Exchange rates (EX), Producer prices (Pproducer), Rainf all (RF), and Rubber prices from Tokyo commodity exchange (TOCOM). The data set is monthly frequency data collected from M1/2007 to M2/2015, covering 100 observations. Then we transform the data to the growth rates before estimation.

T he model specif ication

The SUR model consisting of demand and supply equations of Thai rubber can be specified as follow: Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 117

d export oil malay Q=+++αβ P β P β P + β EX + U tt1 1 2 3 4t 1, t tt s export producer Q=++αδ P δ P ++ δ RF δTOCOM +U t2 1 tt2 3 tt4 t 2, where the error terms of the SUR model are dependent (See Section 2.1). We begin this part with checking stationary of the data sets; the Augmented Dickey-Fuller (ADF) unit root test is used to confirm if the data sets are stationary. We do not show the result in this paper but we found that all variables passed the test at level with probability equals to zero, meaning all of them are stationary.

5.1 Selecting a Copulas

As we suggest adding Copula density function into the likelihood (See Section 3) in order to propose the Copula-based SUR model. Now we have to select an appropriate Copula-based SUR model, and decide which type (or family) of Copulas is best-fit for our data. We consider a set of Copula families as the candidate models, and employ classical ways of selecting the models called the Akaike information criterion (AIC) and the Bayesian information criterion (BIC). The results are exhibited in Table 4.

Table 4. AIC and BIC criteria f or model choice.

Marginal distributions Copulas Type 1 Type 2 Type 3 Type 4 1.6997 58.1878 24.2191 22.3294 Gaussian 1.9880 58.4762 24.5074 22.6177 1.5008 13.2541 20.4090 27.4498 Student’s t 1.7892 13.5424 20.6973 27.7382 1.7098 40.4737 88.6958 39.9138 Clayton 1.9981 40.7621 88.9842 40.2021 1.6614 17.5686 23.4182 86.0829 Rotated Clayton 90 degrees 1.9497 17.8570 23.7065 86.3713 1.6612 25.4381 29.7150 26.1106 Rotated Clayton 270 degrees 1.9496 25.7264 30.0034 26.3989 1.6513 43.3755 17.6921 43.3970 Gumbel 1.9396 43.6638 17.9805 43.6853 1.5793 20.5012 28.6594 40.7777 Rotated Gumbel 90 degrees 1.8676 20.7896 28.3710 41.0661 1.5799 11.1656 19.7847 77.3132 Rotated Gumbel 270 degrees 1.8682 11.4 53 9 20.0730 77.6016 1.7372 23.0008 29.4733 73.5191 Frank 2.0255 23.2891 29.7616 73.8075 1.6359 49.1998 27.4556 40.0768 Joe 1.9243 49.4882 27.7440 40.3651 118 Pastpipatkul, Maneejuk and Sriboonchitta

In this simple case of Copula-based SUR model where the system consists of demand and supply equations, we assume Guassian and Student’s t distributions for the error terms. Then we can define the four possible types of combining the marginal distributions of demand and supply, respectively. The four types can be shown as follows:

Type 1: Guassian / Guassian distributions Type 2: Guassian / Student’s t distributions Type 3: Student’s t / Guassian distributions Type 4: Student’s t / Student’s t distributions

Table 4 shows the values of AIC and BIC for making a comparison among the given Copula families as well as a comparison of the four types of combining the marginal distributions. To choose the best-fit Copulas and marginal distributions for our data, we seek for the minimum AIC and BIC values. We found that the minimum values of AIC and BIC are 1.5008 and 1.7892 (bold and underlined numbers), respectively, meaning that the Student’s t Copula and type 1 are chosen. It implies that the Student’s t Copula is considered to be the one that best-fit for being a linkage between the error terms of demand and supply compared to other given Copulas. In addition, the margins of demand and supply are considered to be normally distributed.

5.2 Estimating parameters of demand and supply equations

In this part, we intrend to evaluate the efficiency of our proposed model and compare with the conventional SUR model. To specify the efficiency, we have to show that the estimation results obtained from our model is accurate -- the estimated parameters should be close to the result obtained from other candidates -- with low standard derivations in the same way as the simulation study. We also try to investigate that our Copula-based SUR model gives more precise estimates than the conventional methods. To compare models, we suggest to consider the mean squared error (MSE), the Sum of Squared Residuals (SSR), and the standard errors (Std. err.) of estimated coefficients, and define the best model as the model which gives the minimum values of MSE and SSR. Table 5 shows the estimation results of Copula-based SUR model1. We found that Thai rubber export prices (Pexport) is significant and the coefficient is negative which follow the theory of demand. Next, Malaysian rubber export prices (Pmalay) and exchange rate are significant and their coefficients are positive implying that

1 The Copula-based SUR model used to compare with other conventional models, here is the Student’s t Copula being a linkage between the error terms, whose margins are considered to be normally distributed (See Table 1). Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 119 the higher the Malaysian export prices and the exchange rate, the higher export demand of Thai rubber. For supply-side, Thai rubber export prices (Pexport) is not significant, but only just so, its coefficient would indicate a positive relationship between price and supply of Thai rubber which also follow the theory of supply. Next, rainfall is statistically significant where the coefficient is negative indicating that the higher level of rainfall, the lower supply of Thai rubber, which is in line with nature of rubber tree.

Table 5. Estimation of Copula-based SUR model for demand and supply in comparison with other estimators.

Copula-based SUR Bayesian SUR SUR (MLE) SUR (LS) Demand Coeff. Std. err. Coeff. Std. err. Coeff. Std. err. Coeff. Std. err. Intercept 0.0186 0.0275 0.0089 0.0179 0.0088 0.0150 0.0088 0.0154 Pexport -1.382 ** 0.6460 -1.899 * 0.9762 -1.896 ** 0.8177 -1.897 ** 0.8395 Poil 0.1830 0.1341 -0.0262 0.2048 -0.0150 0.1706 -0.0150 0.1750 Pmalay 1.215 ** 0.5790 1.8419 * 1.0052 1.828 ** 0.8380 1.829 ** 0.8602 Exchange rate 1.828 ** 0.7308 1.4745 1.0287 1.4979 * 0.8622 1.4970 * 0.8848 Copula-based SUR Bayesian SUR SUR (MLE) SUR (LS) Supply Coeff. Std. err. Coeff. Std. err. Coeff. Std. err. Coeff. Std. err. Intercept 0.0363* 0.0210 0.0289 0.0219 0.0291 0.0196 0.0291 0.0201 Pexport 0.4036 0.5048 0.5819 0.4446 0.5637 0.3956 0.5635 0.4060 Pproducer 0.0131 0.0093 0.0193 0.0223 0.0186 0.0199 0.0187 0.0204 Rainfall -0.004** 0.0014 -0.0030 0.0031 -0.0031 0.0027 -0.0031 0.0028 TOCOM 0.2406 0.2004 -0.4223 0.2946 -0.4069 0.2610 -0.4068 0.2679 Sum of Squared Residuals (SSR) Copula-based SUR Bayesian SUR SUR (MLE) SUR (LS)

SSRdemand 2.1889 2.1997 2.2006 2.2006

SSRsupply 3.6105 3.6156 3.6168 3.6168 SSR (system) 5.7994 5.8153 5.8174 5.8174 Mean Squared Error (MSE) Copula-based SUR Bayesian SUR SUR (MLE) SUR (LS)

MSEdemand 0.02459 0.02472 0.02473 0.02473

MSEsupply 0.04057 0.04062 0.04064 0.04064 Note: “*” and “**” denote rejections of the null hypothesis at the 10% and 5% significance levels, respectively. The underlined numbers denote the lowest Std. err., SSR, and MSE values. 120 Pastpipatkul, Maneejuk and Sriboonchitta

M odel comparison

We compare the Copula-based SUR model -- it is estimated by MLE -- with other conventional SUR models estimated by Bayesian estimators, MLE and LS. The result (Table 5) shows that our model performs slightly better than the others, especially compared to Bayesian and least squared estimators. First, let’s focus on the number of significant parameters. The Copula-based SUR model beats the others in terms of the number of significant parameters is greatest (5 parameters) implying that our model has high potential to analysis the data. Second, we consider the standard error of coefficients (Std. err.) which inform how accurate the estimations are, and we found that our model is much more precise than the conventional SUR estimated by Bayesian estimators and LS because the standard errors of our model are smaller for all parameters. But unlike the SUR estimated by MLE, some estimated parameters are more accurate since they have the smaller standard errors than the Copula-based SUR model, i.e. intercept terms (demand and supply) and Thai rubber export prices (supply). Finally, we consider a perspective of model comparison, and employ the mean squared error (MSE) and the sum of squared residuals (SSR) as the evaluation measure. The results show that Copula-based SUR is the most accurate model for this data set since it has the lowest values of MSE and SSR compared to the others.

5.3 Testing f or residual normality

As we mention as earlier, the conventional SUR model has the strong assumption of residuals which have to be normally distributed. But if we found that the residuals are non-normally distributed after obtaining the estimated results, how can we feel confident because of the contradictory results. Here we try to check the normality of residuals of the conventional SUR model estimated by MLE, LS, and Bayesian estimators. Figure 1 illustrates the distribution of residuals for judging the normality of the distributions. We have the histograms of residuals of the conventional SUR model estimated by MLE (top left), Bayesian estimator (top right), LS (bottom left), and the Copula-based SUR model (bottom right). We also superimpose a normal density function on the histograms (red line) in order to compare with the histograms. We consider the histogram of residual since it can be used to check whether the deviation is normally distributed. The error terms are considered to be normally distributed if there exists a symmetric bell-shape histogram which distributed around zero. Figure 1 shows that the histogram of residuals of demand seem to correspond well to the assumption except the conventional SUR estimated by MLE (top left) which is quite not normally distributed. There exists many peaks asymmetrically located in each side of zero. It suggests that the conventional SUR model estimated by MLE may be violated. But for the supply-side, all models seem not to correspond fairly well to the assumption of normality. The histograms indicate that they exhibit an asymmetric bell-shape histogram and also a flat asymmetric Do Copulas Improve an Efficiency of Seemingly Unrelated Regression Model? 121 tail extending over the interval approximately [-0.5, 0, 0.8]. It also suggests that the conventional SUR models estimated by these given estimators are violated. However this is not going to be a problem for Copula-based SUR model since the normality assumption does not restrict the ability of our Copula-based SUR model.

Figure 1. Histogram of the residuals.

6. Conclusion

Since we do not satisfy the assumption of normally distributed residuals appeared in the conventional SUR model, the Copula density function is applied into the likelihood to relax this normality assumption. Hence, we come up with the Copula-based SUR model and consider it as an alternative to the conventional SUR model. We begin by evaluating performance and accuracy of our proposed model through the simulation study. We found that the mean parameters of the proposed model are very close to their true values with reasonable standard derivations. Then, we consider the real data set of Thai rubber to be applied. We estimate the Copula-based SUR model using MLE, and then compare the estimation results to 122 Pastpipatkul, Maneejuk and Sriboonchitta the conventional SUR model estimated by MLE, LS, and Bayesian estimators. We found that our model perform very well, and gives more precise estimates than the conventional model. Based on the data set of Thai rubber, the results show that Copula-based SUR model is the most accurate model for this data set since it has the lowest values of MSE and SSR compared to the others. Moreover the parameters estimated by Copula-based SUR model mostly have the smaller standard errors than the others, and the number of significant parameters is greatest. According to the overall results of this paper, it is shown that the Copula-based SUR model could perform very well in both simulation study and application to the real data set. Therefore, we could conclude that it may be better to consider the Copulas to be used to relax the strong assumption of normally distributed error terms appeared in the conventional SUR model. Especially when we face to the non-normal residuals, we better choose the Copula-based SUR model, rather than the conventional one.

References

[1] B. H. Baltagi (1988). The efficiency of OLS in a seemingly unrelated regressions model, Econometric T heory, 4, 536-537.

[2] M. Bilodeau and P. Duchesne (2000). Robust estimation of the SUR model, The Canadian Journal of Statistics, 28, 277-288.

[3] M. Hofert, M. Mächler and A. J. McNeil (2012). Likelihood inference for Archimedean copula in high dimensions under known margins, J ournal of Multivariate Analysis, 110, 133-150.

[4] G. Masarotto and C. Varin (2012). Gaussian copula marginal regression, Electronic Journal of Statistics, 6, 1517-1549.

[5] R. B. Nelsen (2010). An Introduction to Copulas (2nd ed.), Springer, New York.

[6] P. Pastpipatkul, P. Maneejuk and S. Sriboonchitta (2015). Welfare measurement on Thai rice market: A Markov switching Bayesian seemingly unrelated regression. In V.-N. Huynh, M. Inuiguchi and T. Denoeux (Eds.), Integrated Uncertainty in K nowledge M odelling and Decision M aking (pp. 464-477), Springer, Nha Trang, Vietnam.

[7] U. Schepsmeier and J. Stöber (2014). Derivatives and Fisher information of bivariate copulas, Statistical Pa pers, 55, 525-542.

[8] M. Sklar (1959). Fonctions de répartition à n dimensions et leurs marges, Publications de l’Institut de Statistique de L’Université de Paris, 8, 229-231.

[9] A. Zellner (1962). An efficient method of estimating seemingly unrelated regressions and tests for aggregation bias, Journal of the American Statistical Association, 57, 348-368. International Journal of Intelligent Technologies and Appld ie Statt is ics

Vol.9, No.2 (2016) pp.123-143, DOI: 10.6148/IJITAS.2016.0902.03 © Airiti Press

Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level

Montri Singhavara1*, Aree Wiboonpongse2, Yaovarate Chaovanapoonphol2 and Thaworn Onpraphai2

1Faculty of Economics, Maejo University, Chiang Mai, Thailand 2Faculty of Agricultural, Chiang M ai University, Chiang M ai, Thailand

ABSTRACT This research aims to study decision approach in using resources together with trade- offs between income from growing crops for food and for biofuel. The study used data from survey in 2013 from 870 agriculturists in Tambol Nikhom Thung Phothalay, Muang District, Kamphangphet Province. To achieve in finding net profit together with energy security in a community level, Interactive Fuzzy Linear Programming with Fuzzy Parameters have been used. Result of the study reveals that highest net profit and satisfaction level come from growing rice, followed by cassava, sugar cane, and maize. In situation of lack of resources (water supply and labor) and slump in agricultural product price, the community would have more income from crops that produce ethanol (tapioca chips and sugar cane) to substitute income from food-producing crops, together with higher energy security in community level.

K eywords: Trade-offs; Land used; Biofuel

1. Introduction

Thailand has a policy of energy security or sustainability similar to many other countries. Ministry of Energy has set an important strategy, that is, civil participation in developing infrastructure in energy area and support the investment of this in community level. At present, important agricultural products to produce ethanol fuel is sugar cane (molasses) and cassava. Generally, cost of producing ethanol from cassava is higher from molasses. So, to support or sustain producing ethanol from cassava, the government has requested fuel traders to buy ethanol from cassava and molasses at the same price in the ratio of 62:38. However,

* Corresponding author: [email protected] 124 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai in the situation that benzene price is going downward, this results in decreasing demand for ethanol, resulting in using ethanol not meeting the target set by Ministry of Energy. Analysis of trade-offs is important and is in widespread use to estimate results and make decision for agricultural production. Trade-offs analysis means the exchange emerged from compromise process and is becoming more important in choosing limited resources along with stakeholders who may have contradictory targets. Understanding dynamic of trade-offs in agricultural system with production and nature change based concept is considered achievement in the purpose of sustainability and food security in the future [17]. Important approach has been developed to analyze trade-offs and this consists of 4 methods: (1) Participatory Research. This is usually used for qualitative data, and useful for setting scenarios and other indicators from stakeholders relating to decision making [4, 6], but this is not appropriate for calculating trade-offs. (2) Empirical analyses or Experimental method. This is beneficial for calculating behavior according to data collected from different situations. Graph for trade-offs can be calculated to show value of various indicators. Strength point of this is to show results of several choices of the system but the weak point is that results are limited by set of only data collected that might not be suitable for forecasting results of other system that has different set of data [32]. (3) Simulation model. This is used to survey alternative from simulation data. The simulation is used for estimating trade-off curves. This is used for studying dynamic of trade-offs which can be different along different time [22, 34]. And (4) Optimization models. This is to employ mathematical programming to estimate trade-off curve that is adapted when changes occurred by decision makers towards the parameters [10, 12]. Optimization is used to estimate potential of synergy and alleviation of trade-offs between objectives or targets but this approach has limitation when socio-cultural factors take important role in the system [17]. From the 4 factors, we can see different strong and weak points. So, appropriate solution is to bring them to use together for better explain what happen in reality. For example, in multi-criteria analysis, method which comes from participatory research and optimization approach, stakeholders set weight according to criteria along with setting importance for each objective in order to study changes in trade-offs [21]. However, parameter value used in agricultural planning model is usually fuzzy or imprecise. This can cause problem in decision making [1]. Setting parameter value according to knowledge and experience of experts is appropriate for agricultural planning but traditional linear programming cannot solve fuzziness of all fuzzy linear programming. Fuzzy set is brought to use with multi-objective linear programming for agricultural land use planning [8, 23]. And in a situation where there must be prioritizing fuzzy objectives, Sinha et al. [29] used pre-emotive priority FGP to plan crop growing by considering trade-offs emerging from membership function of targets. Nevertheless, multi-objective linear programming might face problems in finding appropriate solution due to contradicting of some objectives. Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 125

So, interactive multi-objective linear programming [30] was created to solve such problems. Choo and Atkin [5], and Fichefet [7] have developed STEM (STEP method) which consists of 2 main principles: (1) finding appropriate solution according to Pareto principle, using the concept of minimax. (2) using Decision Maker to help compare and contrast in improving satisfaction level to become higher. This study, therefore, focuses on finding decision approach in using resources i.e. land, water, labor and budget by using trade-offs analysis to achieve 2 main objectives. Firstly, producing crops for biofuel for energy security. Secondly, making income from human and animal manufacturing industry according to Figure 1 which covers conceptual framework of ethanol producing, human and animal food manufacturing from growing cassava, sugar cane, rice and maize.

Figure 1. Inputs use and trade-off between income from Food/Feed and Ethanol manufacturing.

2. Concept of interactive Fuzzy linear programming with Fuzzy parameters

Sakawa and Yano [25, 27, 28] have developed the method of interactive model which results in satisficing for decision makers from optimum solution set according to by changing reference point of membership level and/or level. Setting the function of membership of multi-objective linear programming is shown below in Figure 2. When we change objective function into membership function, this make us become interested in the concept of solution approach of M – α – Pareto optimal in addition from α – Pareto o ptimal by calculating set of solutions according to the concept of minimax [26, 33] and can develop equation for finding solution as below. 126 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

Figure 2. Membership function.

minimizeminimize max max {}{} u uˆˆii−− u u ii ii () () c c x x ikik==1,...,1,..., subjecttosubjectto x x∈∈ X X()() ab ab , , (1)

(,,)(,,)abcabc∈∈ ( ( ABC ABC , , , , ) )αα

By specifying u = (u1, …, uk) That is, membership reference level is based on function of membership ui(cix). Finding appropriate solution for problem (1) can be done by using problem (2) as the following [27].

minimize v L −-1 subjectto cxiα ≤ diR ( uˆi − v ), i ∈∪ I13 I R R −-1 (2) cxiα ≥ diL ( uˆi − v ), i ∈∪ I23 IL LR x∈ Xa (αα , b ).

As function of membership characterized by up or down level and might be non-linear function to solve problem (2) smoothly, we have to use characteristics of strictly monotone function of membership function (with fuzzy goals) diL(∙) and R −-1 L −-1 diR(∙) as cxiα ≥− diL () uˆ i v and cxiα ≤− diR () uˆ i v, by considering ratio of trade- offs between multi-objective membership function, using simplex multipliers from Equation (3) with value ν* comes from solving problem (2) as shown in Lagrangian f unction. Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 127

m L L−−-1* -1*R LR L= cx1α +∑πiR{ cx iα − d iR() uˆˆ i − v} + ∑∑ πλiL{d iL () u i −− v cxiα} + j( a jαα x− b j ) iI∈∪13 IRLiI∈23 ∪ I j=1 (3)

In the following equations, πiL, πiR and λ j are simplex multipliers Haimes and Chankong [9] had shown the ratio of trade-offs between membership function.

∂ucx()LL d′ () cx −11αα = 1R 1 π ∈∪ ≠ LLiR , iI13 IR , i1 (4) ∂ucxi() iαα diR′ () cx i

∂ucx()LL d′ () cx −11αα = 1R 1 π ∈∪ RRiL , iI23 IL (5) ∂ucxi() iαα diL′ () cx i

LL Criteria in considering ratio of Trade-offs is that if the ratio−∂ - ucx11(αα ) ∂ ucxii ( )1 = , that means there can be no more increasing satisfaction for any objective function among these functions although satisfaction is lowered from either of each function. If the LL ratio 00 <−∂< - ucx11 (αα ) ∂ ucxii ( )1 < , this means there can be increase in satisfaction from L LL function ucxii()α and if −∂- ucx11(αα ) ∂ ucxii ( )1 > , this means the increase satisfaction from L function ucx11()α can be better done.

3. Multi-period plant growing land allocation model

Since the conceptual framework in ethanol and food production from agricultural product from Figure 1 is related to time frame in growing, taking care, harvesting and selling, so multi-objective and multi-period models are developed to show dynamic of agricultural process as discrete time periods [13] along 3 years as the following.

F (1) Equation of profit from crop selling to food factory ( Zi ).

3 32 3 32 3 FF F Zi = ∑ Pit Qit − ∑∑ Cijt Area ijtv − ∑ Clit HiLa it − ∑∑ CWiwt HiW iwt −∀ ∑ i t Bor t i t=1 t=11v = t=1 tw=1 =1 t=1 (6) 128 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

E (2) Equation of profit from selling crop product to ethanol factory and fuel traders ( Zi ).

33 33 3232 33 3232 EEE E EEEEE Z EE= P it Q +−() PEit CE  it QE −E C Area − Cl HiLa − CW HiW i Zi =∑ Pitit Qit ∑ +−() PEit CEit QEit it ∑∑ −ijt Cijt Area ijtv ijtv ∑ −it Clit HiLa it it ∑∑ − CWiwtiwt HiW iwt iwt = ∑= ∑ =∑∑ = = ∑= ∑∑ = t1t=1 t1t=1 t11tv=11v = t1t=1 tw1tw=11 =1 3 3 −∀i Bor i ∑−∀ttitt Bor i (7) = ∑ t 1 t=1

(3) Equation of constraints. (3.1) Transfer of water supply in dry and rainy seasons.

 TrWtw(= 1) −≤ HiWtw(= 2) WatQtw(= 1) ∃ t TrW −≤ TrW HiW ∃ t tw(= 1) (t −= 1)( w 1) tw(= 2) (8) 423   ∑∑∑CWUij Areaijtv+− TrW t( w= 1) TrW (t −= 1)( w 1) ≤ WatQtw(= 3) + HiWt( w= 4) ∃ t ivj=111 = =

(3.2) Limitation of underground water usage during dry and rainy seasons.

T  ∑ HiWtw(= 2) ≤∃ WQRw=2 t t=1 (9) T  ∑ HiWtw(= 4) ≤∃ WQRw=4 t t=6

(3.3) Limitation of land area appropriate for growing plants in the studied area (soil series 7 and 33/38).

4 ∑ Areaijtv ≤ Alandv ∀∀ t v (10) i=1

(3.4) Limitation of selling amount of crop production, as amount of product from crop i E F that agriculturists sell to ethanol factory ( Qit ) and food factory (Qit ) has less value than product gained from using the area ν.

QQEF+ ()it it ( 11) ≤ Areajtv ∀ v ∀∃ i t Y i Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 129

Volume of ethanol gained from crop i has less value than product of crop i that is sold to the factory and processed into ethanol.

E CtE i ⋅ Q E ()it (12) QEit ≤∃ t Y i

(3.5) Limitation of household and hired labor for crop growing, taking care and harvesting. Use of labor for activity for crop i has less value than household labor plus hired labor.

2 ∑ LQAreaij ijtv− HiL it ≤ ALt ∀∀∃ j i t (13) v=1

Amount of highest hired labor for crop i at time t.

 HiLit ≤ AHiLt ∀∀ i, t (14)

(3.6) Use of land for harvesting and taking care has less quantity than use of land

for growing. Areaiv( t+= 1)( j 2) ≤ Areaitv( j= 1) ∀ t ∀∀ i v

Areaiv( t+= 1)( j 3)2) ≤ Areaitvitv( j j= 1) ∀∀∀∀ t ∀∀ i v v

Areaiv( t+= 1)( j 2) ≤ Areaitv( j= 1) ∀ t ∀∀ i v Areaiv( t+= 1)( j 3)2) ≤ Areaitvitv( j j= 1) 2) ∀∀∀∀ ∀∀∀ t t ∀∀ i i v v v

Areaiv( t+= 1)( j 3) ≤ Areaitv( j= 1) ∀∀∀ t i v (15) Areaiv( t+= 1)( j 3) ≤ Areaitv( j= 1) 2) ∀∀∀ ∀∀∀ t t i i v v

Areaiv( t+= 1)( j 3) ≤ Areaitv( j= 2) ∀∀∀ t i v Areaiv( t+= 1)( j 3) ≤ Areaitv( j= 2) ∀∀∀ t i v

(3.7) Money transfer for both income, expenditure, loan and budget (t = 1 ~ 36).

4 4 4 42 4 FEFE   E E ∑Pit Qit + ∑ Pit Qit + ∑ PEit QEit − ∑∑ Cijt ⋅ Area ijtv −⋅− ∑ Clit HiLa it ======i1 i1 i1 iv11 i1 (16) 42 ∑∑CWiwt⋅ HiW iwt +⋅ Bc t BoW t + TranFt − TranFt−1 ≤ Budgett ∀ t iw=11 =

(3.8) Last month money transfer has no less value than loan plus agricultural budget.

36 36 ∑∑−−(1itt ) * Bor + TranF t=36 ≥ Budgett (17) tt=1 130 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

Sets: i is the types of crops i ∈ I, I = 1 cassava, I = 2 maize, I = 3 sugar cane, I = 4 rice. F is the selling of product to food factory. E is the selling of product to ethanol factory. j is the crop growing activity j ∈ J, J = 1 is the growing, J = 2 is the taking care, J = 3 is the harvesting. t is the time that agricultural activity happens t ∈ T, T = month 1 ~ 36. ν is the soil series that are suitable for growing crops v ∈ V, V = 1 is soil series 7, V = 2 is soil series 33/38. w is the water supply in crop growing area w ∈ W, W = 1 is irrigation during dry season, W = 2 is underground water during dry season, W = 3 is irrigation during rainy season, W = 4 is underground water during rainy season.

Decision variables: F Qit is amount of crop i sold to food factory at time t (unit: kg ). E Qit is amount of crop i sold to ethanol factory at time t (unit: kg ).

Areai jtv is land use of crop i according to activity j at time t of soil series v (unit: rai).

HiLit is the hired labor of crop i at time t (unit: man-hour).

H iW tw is amount of underground water use during dry and rainy seasons at time t (unit: cubic meter).

T rW (t–1)w is irrigation transferred from time t (unit: cubic meter).

Bort is amount of loan at time t (unit: baht). E QEit is amount of ethanol from crop i which was produced from ethanol factory for selling at time t (unit: liter).

HiLt is amount of hired labor at time t (unit: man-hour).

TranF t is amount of money transfer at time t (unit: baht).

4. Membership f unction

(1) Membership function for objective function. As agriculturists and community need possible highest profit, objective function is a problem of maximization but net profit might be unreliable or difficult to estimate. So, we specify membership function as Fuzzy max membership function (Figure 2b solid line) as exponential f u n c t i o n [ 24 ]. (2) Membership function for resource. This is set as Fuzzy min membership function (Figure 2a dash line) because it is according to limitation problem of resources, by giving importance to lowest value of resources that can be accepted; they are

water supply (W atQtw, WQRw) and hired labor (AHiLt). (3) Membership function of technical coefficient of constraint equation. This consists Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 131

of average product of crop per rai (Y i), coefficient of water use for crop (CWU ij)

and amount of ethanol from land usage in crop growing (CtEi). This membership function employs Fuzzy min membership function (Figure 2a solid line) and has characteristics of exponential function, showing importance of risk-free [3].

Coefficient of water use for crop (CWU ij) and ethanol volume from land use for

crop growing (CtEi) employ membership function type of symmetric triangular possibilistic distribution (Figure 2c dash line) because uncertainty of information can be possible both in upward or downward ways from either of any value resulted from the experiment or scientific testing [18, 31]. (4) Member function for coefficient value of objective equation. Since uncertainty can occur from both income and cost, parameters gathered are in information interval consisting membership function of selling prices of crop products and FE E ethanol (,PPit  it and PE it ) . Since agriculturists usually face uncertainty of selling price, pessimistic world view is considered important situation [3]. Therefore, we set membership function as Fuzzy min membership function (Figure 2a E dash line). As for membership function of cost of ethanol production ()CE it , most agriculturists face rising cost of produce. To find appropriate solution, considering highest possibility, we therefore use Fuzzy max membership function (Figure 2b dash line).

5. Data f or studying

5.1 Technology and raw materials for ethanol producing in community level

At present, many educational institutions in Thailand realize the importance of biofuel production in a community level. So, refinery machines have been developed for factories (size of community). The main criteria are that using low budget and can be used with agricultural products that are of many types and easily available in a community. To make it accord to the above criteria, least processed products have been chosen to reduce cost of storage and transportation as well as considering amount of other raw materials to be used in other processed industry, i.e. tapioca chips and sugarcane juice (instead of molasses)

5.2 Constant value of resource amount from constraint equation

From secondary information from relevant government offices, i.e. department of land development, social development and agricultural office of Muang District, we can have parameter value of physical geography, economy and society as the following. 132 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

(1) Areas that are suitable for crop growing, Nikhom Thung Pho Thalay includes 16 villages. Data from land development department shows the highest proportion of 2 soil series: series 7 (suitable for rice and maize growing) and 33/38 (suitable for maize and sugar cane growing). So, limitation of land area has been set according to these 2 soil series which total 12,555 and 14,770 rai accordingly [14].

(2) Budget of agriculturists (Budgett). There are overall 1,148 households doing agriculture [16]. Income occurred from agricultural activities totaled 7,736,467.67

baht/month, making average capital of agriculturists (Budgett) equal to 2,320,940.30 baht/month (30% of overall agricultural income).

(3) Amount of water resource from irrigation (W atQ tw) and underground water

(WQRw). The sampling area is in the area that that get Tho-Thong Daeng and Wang Bua irrigation projects. Water supplying is divided into 2 periods i.e. rainy season (June ~ September) (total amount 175 ~ 180 million cubic meter) and dry season (December ~ March) (total amount 135 ~ 150 million cubic meter) [11]. And data from interview with officials of Bureau of Groundwater Resources Region 7 identified that 10 from 16 villages have more than average groundwater resource (100 ~ 200 gallon/minute) or 200,000 ~ 250,000 cubic meter/rai in dry season (December ~ March) and 350,000 ~ 400,000 cubic meter/rai in rainy season (June ~ September) [2].

(4) Amount of household labor (ALt) and hired labor available in the area (AHiLt). Average members of family doing agricultural work are around 3 persons. Hired labors in the area totaled 989 ~ 3,500 persons [16]. By specifying 1 labor working 8 hours per day (8 man-hour) and working days total 25 days per month, man-hour of household labor can be calculated as 600 and hired labor as 197,800 ~ 700,000 man-hour/month.

5.3 Technical coef ficient of constraint equation

By surveying field data by random sampling of agriculturists growing cassava, maize, sugar cane and rice which total 263, 265, 283 and 59 agriculturists accordingly from villages that has soil series 7 and 33/38 more than 40% of agricultural land area, we found the following details.

(1) Average yield per rai of crop (Y i) and ethanol amount from land used for growing

crop (CtEi). As each agriculturist has different production efficiency, including deviation in information given by agriculturists, Table 1 is shown as information interval and since ethanol ratio from cassava and sugar cane are averagely about 333 and 70 liter/1,000 kg. Accordingly, calculation of ethanol can be done as shown in Table 1.

(2) Ratio of labor usage in crop growing activity (LQij). From field survey, we know about labor usage for all crop growing activities (planning, taking care and

harvesting) and also from coefficient water use of crops (CWU ij) along with Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 133

uncertainty of data of water usage for each crop as shown in Table 2. In part of

interest rate of loan (it), agriculturists have a cost of loan for crop growing in a loan rate as a good-grade retail (MRR equal to 7.0% per year).

Table 1. Average yields of each crop categorized according to soil series.

Yield (kg/rai) (Y i) Ethanol amount (liter/rai) (CtEi) Crop Soil series 7 Soil series 33/38 Soil series 7 Soil series 33/38 Lowest Highest Lowest Highest Lowest Highest Lowest Highest Cassava 4,700.9 4,858.2 4,362.7 4,985.2 1,565.4 1,617.8 1,452.8 1,660.1 Maize 634.2 674.0 618.4 668.4 - - - - Sugar cane 12,250.6 13,333.3 11,53 9.9 2,099.0 857.5 933.3 807.8 846.9 Rice 802.5 940.0 850.3 900.0 - - - -

Table 2. Cost of producing, man-hour and water usage of crops.

Cost (baht/rai) (Ci jt) labor (man-hour/rai) (LQij) Water used by Crop Taking Taking crops (cubic m/ Planting Harvesting Planting Harvesting care care rai) (CWU ij) Cassava 704.6 2,606.4 2,773.7 19.7 36.2 22.8 1,900 ~ 2,000 Maize 908.6 4,187.9 1,797.0 17.8 45.0 44.2 400 ~ 700 Sugar cane 1,090.7 5,585.9 3,013.8 36.2 69.3 41.6 1,200 ~ 2,000 Rice 2,479.2 2,290.2 1,783.8 13.0 48.0 40.0 1,022 ~ 1,137

5.4 Coef ficient value of objective equation

 F (1) Selling price of agricultural product to food factory ( Pit ) and community ethanol  E factory ( Pit ). From surveying sampling agriculturists in planting year of 2011- 12 and 2012-13 together with secondary information from Office of Commercial Affairs Kamphaengphet Province [20], we can know about selling price rate as E in Table 3 while ethanol price ()PE it is usually lower than reference price set by government office because fuel traders have higher bargaining power as shown in interval in price in Table 3.

(2) Cost of crop production (Ci jt). By using information gathered from agriculturists growing cassava, maize, sugar cane and rice, we can get relevant average of

costs shown in Table 2 while cost of hired labor (Clit) equals to 300 baht/day or 37.5 baht/man-hour. And as the studied area is in irrigation area, this makes less

cost for water usage, so cost of water used for crops (CW iwt) is set as 1 baht/cubic mater. 134 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

Table 3. Selling price of agricultural products to food and community ethanol factories.

Price interval or crops (baht/kg) and ethanol (baht/liter) Crops 2 011-12 2012-13 2013-14 Lowest Highest Lowest Highest Lowest Highest Cassava (May ~ Apr.) 2.00 2.20 2.27 2.30 - - Maize (May ~ Aug.) 6.55 7.45 6.23 6.52 6.85 7.50 Maize (Sep. ~ Dec.) 8.50 9.93 6.23 6.50 - - Sugar cane (Nov. ~ Oct.) 1.02 1.07 1.00 1.05 - - Rice (May ~ Aug.) 8.20 9.75 8.20 9.99 7.80 8.21 Rice (Nov. ~ Feb.) 8.00 8.50 8.30 10.25 - - Ethanol (May ~ Apr.) 21.50 22.16 22.0 23.60 - - Ethanol (Nov. ~ Oct.) 21.50 24.0 25.50 27.16 - -

E (3) Cost of producing ethanol from crops ()CE it . From renewable energy department, ministry of energy, cost of producing ethanol from tapioca chips and fresh sugar cane is 7.1 and 6.12 baht/liter. From present technology, rate of using raw materials for ethanol producing (from tapioca chips has average cost of 3 kg/ liter and as for (fresh) sugar cane, 14.29 kg/liter. So, interval of cost in producing E ethanol ()CE it is shown in Table 4.

Table 4. Cost of producing ethanol according to technology and uncertainty of raw material prices.

Raw material (baht/kg) Raw material (baht/liter) Interval cost of ethanol Crops at level of alpha (0.2 ~ 1) at level of alpha (0.2 ~ 1) producing (baht/liter) 2 011-12 2012-13 2 011-12 2012-13 2 011-12 2012-13 Tapioca chips 2.34 ~ 2.50 2.58 ~ 2.60 7.03 ~ 7.51 7.74 ~ 7.81 14.61 ~ 16.1 14.91 ~ 16.4 (Fresh) 1.02 ~ 1.07 1.00 ~ 1.05 14.68 ~ 15.34 14.39 ~ 15.0 21.46 ~ 23 21.12 ~ 22.6 Sugar cane

6. Results and discussion

To get solution according to M – α – Pareto, it is done by Equation (2) together with multi-objective and multi-period model, that is, Equations (6) ~ (17). This has process and following results. Step 1: calculation maximum and minimum values of each objective equation by specifying level α = 0.2 and α = 1. Result of calculation Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 135 will inform scope of uncertainty of net profit all along 3 years. Considering fair competition of each objective equation, target value of net profit has been set as 200 ~ 2.1 million baht. Step 2: setting form of (membership function) according to membership function (4). Step 3: setting α (0 ≤ α ≤ 1) together with at several levels. As value of α which is near 0 indicate optimistic world view (but chance of this solution in practical reality is low) while value of α which is near 1 has a characteristics of pessimistic worldview (solution is highly possible) [3]. Results from calculation revealed that value of membership of each objective equation and ratio of trade-offs between level of membership from objective equation are as following.

6.1 Membership level, net profit and ratio of trade-offs

When specifying α to be near 0 (optimistic), membership value and satisfaction level becomes higher (ui is near 1) together with increase in net profit of all objective equations, particularly net profit of rice and cassava (ethanol) (Figure 3g and 3i) which has net profit of 185.7 and 115.4 million baht (comparable to 50 and 26.2% of sum of net profit of other crops). And when value of α becomes higher (pessimistic becoming higher), this results in value of membership and net profit of all objective equation become lower (Figure 3a, 3c, 3e, 3g, 3i and 3k). Membership level of maize becomes lower a lot because it has high cost of produce and choice of selling channel is only for animal food factory (cassava and sugar can have more selling channels).

As for specifying level of reference membership uiˆi , =1,....,..., kk at several levels, this nearly has no effect to membership level difference and to ratio of trade-offs between objective equation of net profit of any crop, except for objective equation from maize growing in which net profit can increase along with the increase of uˆi and has highest value of 41.41 million baht (uˆi = 1). So we have chosen uˆi = 1 to analyze together with value of α at different levels and by using Equations (3) ~ (5), we can get ratio of trade-offs as shown in Figure 3b, 3d, 3f, 3h, 3j and 3l where ratio of trade-offs of membership level from all objective equations when compared to maize (UMf) have value higher than 1 (Figure 3b, 3f, 3h, 3j and 3l graph curve). This results in increase in net profit from rice, cassava and sugar cane that can be done better than maize. Rice can be grown in substitute of maize very well in the interval of α = 0.4 ~ 0.6, that is, ratio of trade-off is 155.7 and 138.9 (Figure 3h URf04 and URf06). But when production resources decrease (labor and water), also plus negative economic factor (lower selling price but higher cost of produce) (pessimistic), this results in lower advantage of growing rice over maize, while advantage of cassava and sugar cane over rice become higher (Figure 3b, 3f, 3j and 3l). Growing and selling tapioca chips to ethanol factory (UCe) can be shown from ratio of 88.7 (Figure 3j at UCe1). 136 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

Figure 3. Membership level and trade-offs between different objectives. Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 137

Selling cassava and sugar cane to ethanol factory causes satisfactory and net profit in a higher level than selling them to food factory. This can be observed from ratio of trade-off value over 1 both in optimistic and pessimistic view (Figure 3j at Ucf and Figure 3l at Usf). Moreover, selling cassava for ethanol producing can substitute growing and selling sugar cane for food and ethanol well when specifying α < 0.6 (rather optimistic view) but when price, budget and resource situation become negative α ≥ 0.6, selling sugar cane to ethanol and food factories causes more profit when compared to selling cassava (ethanol), as shown in ratio of trade-offs value 0.9 and 0.7 accordingly (Figure 3j at UCe1). Moreover, net profit from selling cassava and sugar cane to ethanol factory has made satisfaction or net profit level become higher than growing and selling rice either in optimistic or pessimistic views. This can be checked from trade-off value point below 1 (Figure 3h at UCe and Use).

6.2 Resource allocation

Supposing decision maker has neither too much pessimistic nor optimistic in aspects of selling price, budget, technical coefficient and resource amount, when considering chance in increasing production of ethanol from using fresh sugar cane together with tapioca chips, it should be specified that α = 0.6 and uˆi = 1 and this emerges the following allocation approaches. (1) Land and water usage. From the model, use of most overall land (90% of all available land) occurred in November ~ February 2014 (growing cassava, sugar cane and rice). Cassava use overall highest land area (14,069 rai), followed by maize (8,922 rai) but it is only one time use during 3 year-round (September ~ December 2012) while rice growing use 4 rounds of land in production which equals to using land area of 8,422 rai (third largest land usage and happened in round 4 of growing season (in-season May ~ August 2014)). Sugar cane uses similar land area to rice (8,352 rai) and begins growing in near the end of year 2 until end of year 3. Results of this study are different from real usage by agriculturists (2012-13) [15] that used the area for growing rice most (14,792 rai) followed by sugar cane, maize and cassava (2,300,665 and 500 rai accordingly). In part of water use, it is in similar direction with the use of land for crop growing, i.e. having volume more than 4.6 million cubic meters/ month in November 2013 ~ February 2014. Cassava used water supply the most (Cas3w and Cas7w) because it used land area more than other crops. But when compared to the use of water per 1 rai, it is found that rice used water the most (Rice7w and Rice3w) (Figure 4). 138 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

30,000 1,500,000

Rai

25,000

20,000 oubic mafer 1,000,000

15,000

10,000 500,000

5,000

0 0

Jan-12 Mar-12 May-12 Jul-12 Sep-12 Nov-12 Jan-13 Mar-13 May-13 Jul-13 Sep-13 Nov-13 Jan-14 Mar-14 May-14 Jul-14 Sep-14 Nov-14

Cas7 Cas3 Mai7 Mai3 Sug7 Sug3 Rice7 Rice3

Cas7w Cas3w Mai7w Mai3w Sug7w Sug3w Rice7w Rice3w

Figure 4. Land and water used for cultivation of cassava, maize, sugar cane and rice.

(2) Quantity, value of employment, economic returns and energy security. From the study, it is found that there is extra hired labor besides household labor for growing every kind of crops, particularly in harvesting. From all labor in the area, hired labor totals 1,993 person, worth as 6 million baht per month. Most of labor activities related to growing and harvesting in April ~ May and August ~ November. From approach in resource allocation according to model of community agriculturists, there will be net returns along 3 years about 393.98 million baht, more than average income of 3 years of agriculturists in this Tambol which is worth 281.1 million baht [16]. In part of energy security, amount of ethanol produced from cassava and sugar cane has at least 250,572 liters/month (α = 1) and maximum as 674,723 liter/month (α = 0.2) while average demand for gasohol and benzene per household in northern region is 692 baht/month [19]. When calculating along household in Tambol Nikhom Thung Pho Thalay [16], demand for ethanol would equal to 47,934 liters/month (supposing from using ethanol 100%). So, ethanol amount produced can be more than enough to sell to other areas or other industries that really need it.

7. Summary and discussion

From aim of ministry of energy that wants to support the production and use of biofuel (Alternative Energy Development Plan: AEDP 2012-2021), this resulted in the creation of “agricultural community project in the cluster of food and energy crop” in Nikhom Thung Pho Thalay, Mueang District, Kamphaengphet Province. But controversy in competition between biofuel and food and feed industry is important and should be taken into consideration carefully. From the study with Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 139

Interactive Fuzzy Linear Programming with Fuzzy Parameters together with trade- offs analysis, main conclusion emerges as follow. From value of trade-offs, growing rice has more advantages than growing cassava, sugar cane and maize for selling to food and feed factories within optimistic view framework. That is, level of resources (water and labor) is plenty and economic factors are in favor (high selling price and low budget for production). On the other hand, when insufficiency of resources and unfavorable economic factors happen, growing crops (cassava and sugar cane) for ethanol can best substitute growing crops for human and animal food, including rice, cassava and sugar cane as food. Community can gain more income from plants for ethanol to substitute decreasing food and feed crops, simultaneously gaining community level energy security. However, the community still earn profit from food crops in high ratio (more than 60%) (Figure 5), especially as rice is still being produced in the community. The result is like this because the model for the study aims to target profit based on satisficing and compromising principles. Moreover, using (fresh) sugar cane instead of molasses can support more using of tapioca chips for producing ethanol because of aspect in amount of ethanol producing process. The good point is that this increases the choice in selling products for agriculturists and reduce worries about competition for molasses for other industries.

1.20 100% profit SugE

90% profit CasE 1.00 80% profit RiceF 70% 0.80 profit SugF 60% profit MaiF 0.60 50% profit CasF 40% 0.40 satisfy CasF SATISFY LEVEL SATISFY 30%

20% satisfy MaiF 0.20 10% satisfy SugF

- 0% satisfy RiceF 0.2 0.4 0.6 0.8 1 ALPHA satisfy CasE

satisfy SugE

Figure 5. Satisfaction levels and percentage of net profit from objective equations.

As for approach in resource allocation for the period of 3 years, considering resource and economic situations which are neither too positive nor too negative

(α = 0.6, uˆi = 1), there emerges an approach for both area and time allocation when compared to use of growing area for 4 crops in growing year 2012-13 [14]. That is, there should be increasing area for cassava growing for 13,569 rai (beginning during middle of year 2) , increasing maize growing area for 8,267 rai (only one season growing), increasing sugar cane area for 6,052 rai (beginning near the end of year 2), 140 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai and reducing rice growing area for 6,370 rai (growing only in-season rice). Moreover, water amount needed for agriculture tended to be higher in the growing year 2013/14 which directly correlated to the use of area for cassava and rice growing. As for labor needed, this increased a lot during 2nd and 4th quarters for growing and harvesting activities. As for ethanol produced, if the amount was more than need in the community, it can be sold to make income and stored as reserved energy for energy security without affecting food production. Suggestion for policy related issue is presented in the following points. (1) Agricultural production efficiency should be supported to increase yields per rai and plant breeding development should be supported to produce more amount of flour and sugar. (2) Technology in producing ethanol in a community level should be supported to be more efficient with important principles of low budget and flexibility for using raw materials found in local area. (3) Storage warehouse for ethanol in a community level should be promoted together with low-interest rate loan for starting community ethanol enterprise, including finding applying ethanol for other industries to increase more channels in selling in the future. (4) 2nd generation technology for biofuel should be urgently developed to make use of bio- mass from plants used in producing biofuel and this can create income channel particularly bio-mass from maize. (5) agricultural area zoning management by government should be carefully decided by considering agriculturists’ decision and considering relationship between economic factors (price, budget, and marketing), nature’s uncertainty factor and time factor (growing seasons).

References

[1] A. Biswas and B. B. Pal (2005). Application of fuzzy goal programming technique to land use planning in agricultural system, Omega, 33, 391-398.

[2] Bureau of Groundwater Resources Region 7 (2012). Groundwater System of Tambol Nikhom Thung Pho Thalay, Muang District, Kamphang phet. Available at http://bgr7.dgr.go.th

[3] C. Carlsson and P. Korhonen (1986). A parametric approach to fuzzy linear programming, Fuzz y Sets and S ystems, 20, 17-30.

[4] M. Chaudbury, J. Vervoort, P. Kristjanson, P. Ericksen and A. Ainslie (2013). Participatory scenarios as a tool to link science and policy on food security under climate change in East Africa, Regional Environmental Change, 13, 389- 398.

[5] E. U. Choo and D. R. Atkins (1980). An interactive algorithm for multicriteria programming, Computer and Operations Research, 7, 81-87. Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 141

[6] R. DeFries and C. Rosenzweig (2010). Toward a whole-landscape approach for sustainable land use in the tropics, Proceedings of the National Academy of Science of the United States of America, 107, 19627-19632.

[7] J. Fichefet (1976). GPSTEM: An interactive multiobjective optimization method. In A. Prekopa (Ed.), Progress in Operations Research (Vol. 1, pp. 317-332). North- Holland, Amsterdam.

[8] A. P. Gupta, R. Harboe and M. T. Tabucanon (2000). Fuzzy multiple-criteria decision making for crop area planning in Narmada river basin, Agricultural S ystems, 63, 1-18.

[9] Y. Y. Haimes and V. Chankong (1979). Kuhn-Tucker multipliers as trade-offs in multiobjective decision-making analysis, Automatica, 15, 59-72.

[10] P. B. R. Hazell and R. D. Norton (1986). M athematical Programming f or Economic Analysis in Agriculture, Macmillan, New York.

[11] Irrigation Of f ice 4 (2012). Irrigation System of Tambol Nikhom Thung Pho T hala y, Am phoe M uang District, K am phang phet. Available at http://ridceo. rid. go.th/kampang/kampang

[12] S. Janssen and M. K. Van Ittersum (2007). Assessing farm innovations and responses to policies: A review of bio-economic farm models, Agricultural S ystems, 94, 622-636.

[13] H. M. Kaiser and K. D. Messer (2011). M athematical Programming f or Agricultural, Environmental, and Resource Economics, John Wiley & Sons, Hoboken, NJ.

[14] Kamphangphet Land Development Station (2013). Land U se of T ambol N ikhom Thung Pho Thalay, Amphoe Muang District, Kamphang phet. Available at http://r09.ldd.go.th/Station/kpp

[15] Kamphangphet Provincial Agriculture Extension Of f ice (2012). Inf ormation on Agriculture in Kamphang phet Province. Available at http:/ /www. kamphaengphet.doae.go.th

[16] Kamphangphet Provincial Social Development and Human Security Of f ice (2013). Inf ormation about Minimum Quality of Lif e in T haland. Available at http://www.kamphaengphet.m-society.go.th

[17] C. J. Klapwijk, M. T. van Wijk, T. S. Rosenstock, P. J. A. van Asten, P. K. Thornton and K. E. Giller (2014). Analysis of trade-offs in agricultural systems: Current status and way forward, Current O pinion in Environmental Sustainability, 6, 110 -115. 142 Singhavara, Wiboonpongse, Chaovanapoonphol and Onpraphai

[18] Y.-J. Lai and C. L. Hwang (1992). Fuzzy M athematical Programming: M ethods and A pplications, Springer, New York.

[19] National Statistical Of f ice (2013). 2013 Agricultural Census Kamphaeng phet Province. Available at http://www.nso.go.th/

[20] Of f ice of Commercial Af fairs Kamphaengphet Provincial (2011). Inf ormation on M arketing in K amphang phet Province. Available at http://internet.pcoc. demotoday.net/wappPCOC/views/dprice.aspx?pv=62

[21] C. Romero and T. Rehman (2003). Multiple Criteria Analysis f or Agricultural Decisions, Elsevier, Amsterdam.

[22] M. C. Ruf ino, J. Dury, P. Tittonell, M. T. van Wijk, M. Herrero, S. Zingore, P. Mapfumo and K. E. Giller (2011). Competing use of organic resources village- level interactions between farm types and climate variability in a communal area of NE Zimbabwe, Agricultural S ystems, 104, 175-190.

[23] B. Sahoo, A. K. Lohani and R. K. Sahu (2006). Fuzzy multiobjective and linear programming based management models for optimal land-water-crop system planning, W ater Resources M anagement, 20, 931-948.

[24] M. Sakawa (1993). Fuzzy Sets and Interactive Multiob jective O ptimization, Plenum Press, New York.

[25] M. Sakawa and H. Yano (1986). Interactive decision making for multiobjective linear programming problems with fuzzy parameters. In G. Fandel, M. Grauer, A. Kurzhanski and A. P. Wierzbicki (Eds.), Large-Scale M odeling and Interactive Decision Analysis (pp. 88-96), Springer, New York.

[26] M. Sakawa and H. Yano (1987). An interactive satisf icing method f or multiobjective nonlinear programming problems with fuzzy parameters. In J. Kacprzyk and S. A. Orlovskiĭ (Eds.), Optimization Models Using Fuzzy Sets and Possibility T heory (pp. 258-271), D. Reidel, Dordrecht, Holland.

[27] M. Sakawa and H. Yano (1990). An interactive fuzzy satisf icing method for generalized multiobjective linear programming problems with fuzzy parameters, Fuzz y Sets and S ystems, 35, 125-142.

[28] M. Sakawa, H. Yano and I. Nishizaki (2013). Linear and Multiob jective Programming with Fuzzy Stochastic Extensions, Springer, New York.

[29] S. B. Sinha, K. A. Rao and B. K. Mangaraj (1988). Fuzzy goal programming in multi-criteria decision systems: A case study in agricultural planning, Socio- Economic Planning Sciences, 22, 93-101. Trade Offs of Income between Crops for Agricultural Purpose and Energy Purpose in a Community Level 143

[30] R. E. Steuer and E.-U. Choo (1983). An interactive weighted Tchebychef f procedure for multiple objective programming, M athematical Programming, 26, 326-344.

[31] H. Tanaka, H. Ichihashi and K. Asai (1984). A formulation of fuzzy linear programming problems based on comparison of fuzzy numbers, Control and C ybernetics, 13, 185-194.

[32] D. Valbuena, O. Erenstein, S. H.-K. Tui, T. Abdoulaye, L. Claessens, A. J. Duncan, B. Gérard, M. C. Ruf ino, N. Teufel, A. van Rooyen and M. T. van Wijk (2012). Conservation agriculture in mixed crop-livestock systems: Scoping crop residue trade-offs in Sub-Saharan Africa and South Asia, Field Crops Research, 132, 175-184.

[33] A. P. Wierzbicki (1980). The use of reference objectives in multiobjective optimization. In G. Fandel and T. Gal (Eds.), Multiple Criteria Decision Making: T heory and A pplication (pp. 468-486), Springer, Berlin, Germany.

[34] S. Zingore, E. González-Estrada, R. J. Delve, M. Herrero, J. P. Dimes and K. E. Giller (2009). An integrated evaluation of strategies for enhancing productivity and prof itability of resource-constrained smallholder farms in Zimbabwe, Agricultural S ystems, 101, 57-68.

International Journal of Intelligent Technologies and Appld ie Statt is ics

Vol.9, No.2 (2016) pp.145-151, DOI: 10.6148/IJITAS.2016.0902.04 © Airiti Press

A Cluster Analysis of Bank Lending Behavior by Using Self-Organizing Map: The Case of Japan

Satoru Kageyama*

Graduate School of Economics, Osaka University, Osaka, J a pan

ABSTRACT This paper provides a cluster analysis of the bank lending behavior in Japan from 1977 to 2002, in which the bubble economy occurred. The self-organizing map is used for the clustering. This clustering method is a type of artificial neural network and enables us to visualize the similarities among the high-dimensional data on bank lending. A result of clustering indicates that this time period is classified into two periods: before and after 1992. Moreover, we successfully extract the characteristics of bank lending behavior on 8 borrower sectors from another clustering result.

K eywords: Bank lending; Self-organizing map; Bubble economy

1. Introduction

In the 1980s and early 1990s, the significant fluctuation of the asset prices occurred in Japan. This phenomenon is typically referred to as financial bubble, but the mechanism of such a price fluctuation has not been fully elucidated. An empirical study on Japanese real estate market by Ito and Iwaisako [1] showed that the change of bank lending to real estate sector causes the fluctuation of land price in the sense of Granger. This research, however, merely evaluated the correlation between the growth rate of the bank lending and the growth rate of the land price. It is then necessary to understand the bank lending behavior not only toward the real estate sector but also toward other relevant sectors. The artificial neural network is widely used for data mining. The self-organizing map (SOM) is known as a suitable method for cluster analysis and visualization of high-dimensional data. Recent studies have applied the SOM to evaluation of economic phenomena such as bankruptcy, currency crisis, and financial stability [2,

* Corresponding author: [email protected] 146 Kageyama

6, 7]. Despite its wide applicability, no attempts have been made to analyze the bank lending behavior with the SOM clustering. The aim of the present study is to provide a cluster analysis of bank lending behavior with the SOM. First, we classify the amount of bank lending in each year in Japan from 1977 to 2002. Next, we extract the characteristics of the bank lending behavior on 8 borrower sectors.

2. Self -organizing map

This chapter briefly explains the algorithm of the SOM. The SOM is a type of artificial neural network with unsupervised learning method developed by Kohonen [3, 4]. The SOM has a two-layer structure, the input layer and output layer. In the output layer, neurons are arranged on the 2-dimensional grid. Each neuron receives inputs from the input layer and from the other neurons in the output layer. The n input vector is written as an n-dimensional real vector xi ∈ R , where i is the index n of the input. Each neuron has a weight vector mj ∈ R , where j is the index of the n weight vector. At each learning step, the winner neuron mc ∈ R is identified by computing the distances between xi and all the weight vectors mj as follows:

xmic−=min{ xm i − j } (1) j where ||∙|| denotes the Euclidean distance.

Next, the weight vectors are updated in order to reflect the similarity of neighboring neurons as follows:

mj(t + 1) = mj(t) + hc,j(t)[ xi – mj(t)] (2)

where t is the learning step, and hc, j(t) is the neighborhood function. The neighborhood function is given by

2 rrcj− ht()=α ()exp t  − (3) cj, 2σ 2 ()t 

where rc and r j are positions of neurons c and j on the output grid, α(t) is the learning rate, and σ(t) is the width of the neighborhood function. Both α(t) and σ(t) decrease monotonically with time as follows: A Cluster Analysis of Bank Lending Behavior by Using Self-Organizing Map: The Case of Japan 147

α(t) = α(0) (1 – t/T ) (4) σ(t) = σ(0) (1 – t/T ) (5) where T is the maximum number of the learning. The above steps are iterated for all the input data. In our study, the Torus-SOM is used [5]. In the Torus-SOM, the top and bottom sides and the left and right sides of neuron map are designed to be continuously connected.

3. Results and discussions

Figure 1 plots the amount of bank lending to 8 sectors from 1977 to 2002: the “individuals” sector, the “Wholesale and retail trade, eating and drinking places” sector, the “Services (Old basis)” sector, the “Real estate” sector, the “Finance and insurance” sector, the “Construction” sector, the “Transport and communications” sector, and the “Electricity, gas, heat supply and water” sector. The source of data is the statistics published by the Bank of Japan. The data are summarized by Statistics Bureau of Japan (http://www.stat.go.jp/data/chouki/14.htm).

Figure 1. Bank lending to 8 borrower sectors (1977 ~ 2002).

As shown in Figure 1, except for the “Electricity, gas, heat supply and water” sector, the amount of bank lending toward every sector explosively increased from the early 1980s to the mid-1990s. It should be noted that there is a discontinuous 148 Kageyama jump of lending at 1992. After this jump, the rates of increase in bank lending tend to reduce. According to Ito and Iwaisako [1], the land price began to decline in 1991, and thus the year 1992 seems to be a turning point. In the following clustering, the same data as represented in Figure 1 are used. Figure 2 shows the result of the SOM clustering on time series for 26 years. For this clustering, each input vector corresponds to each year. The inputs are 26 vectors of 8-dimension. For example, the vector “1977” has 8 elements: the lending amount for the “individuals” sector in 1977, the lending amount for the “Wholesale and retail trade, eating and drinking places” sector in 1977, the lending amount for the “Services (Old basis)” sector in 1977, the lending amount for the “Real estate” sector in 1977, the lending amount for the “Finance and insurance” sector in 1977, the lending amount for the “Construction” sector in 1977, the lending amount for the “Transport and communications” sector in 1977, and the lending amount for the “Electricity, gas, heat supply and water” sector in 1977. For an example of input data to the SOM, see Appendix. In this example, each row of the data corresponds to each input vector and its label of the input vector. The size of map is 30 × 24 units. The initial size of neighborhood is 30. The number of iterations is 65,000. The rate of learning is 0.01. Every hexagonal unit in the map represents a weight vector which is simulating a neuron. Similar input vectors are adjacent to each other in the map. Units with darker color tell us that its neighboring units have lower degree of similarities.

Figure 2. SOM clustering for time series (1977 ~ 2002).

In Figure 2, the year 1992 is a break point on the map. The time period from 1977 to 1991 and the time period from 1992 to 2002 are divided. In addition, the behavior A Cluster Analysis of Bank Lending Behavior by Using Self-Organizing Map: The Case of Japan 149 in the late 1970s, that in 1980s, and that in 1990s does not resemble each other. These results are in good agreement with Figure 1. We have obtained a clear classification of bank lending behavior on time series. Figure 3 indicates the clustering result on 8 borrower sectors. Here, each input vector corresponds to each borrower sector. The inputs are 8 vectors of 26-dimension, which are obtained by reversing the rows and columns of the data set used in Figure 2. The labels “I,” “W,” “S,” “R,” “F,” “C,” “T,” and “E” denote the vectors of “individuals,” “Wholesale and retail trade, eating and drinking places,” “Services (Old basis),” “Real estate,” “Finance and insurance,” “Construction,” “Transport and communications,” and “Electricity, gas, heat supply and water,” respectively. For example, each element of the vector “R” represents the amount of bank lending in each year toward the real estate sector. The size of map is 30 × 30 units. The initial size of neighborhood is 30. The number of iterations is 50,000. The rate of learning is 0.01.

Figure 3. SOM clustering for 8 borrower sectors.

In Figure 3, there are two groups. One group consists of the “Construction” sector, the “Transport and communications” sector, and the “Electricity, gas, heat supply and water” sector. These sectors mainly involve the infrastructures. The other group consists of the “individuals” sector, the “Wholesale and retail trade, eating 150 Kageyama and drinking places” sector, the “Services (Old basis)” sector, the “Real estate” sector, and the “Finance and insurance” sector. As can be seen in the map, the “Wholesale and retail trade, eating and drinking places” sector resembles both the “individuals” sector and the “Services (Old basis)” sector. These sectors mainly involve the consumption. On the other hand, the “Real estate” sector also resembles the “Finance and insurance” sector. These sectors partly involve the investment. Since the left and right sides are connected in the map, one finds that the “Construction” sector unexpectedly resembles both the “Finance and insurance” sector and the “Transport and communications” sector rather than the “Real estate” sector. We have successfully extracted the characteristics of bank lending behavior on various sectors with the SOM clustering.

4. Conclusions

We have investigated the behavior of bank lending in Japan from 1977 to 2002 by using the SOM clustering. The amount of bank lending toward various sectors explosively increased in this period and there was a discontinuous jump of lending at 1992. The result of clustering also indicates that the year 1992 is a turning point. Moreover, we have successfully extracted the bank lending behavior on 8 borrower sectors. Our results demonstrate that the SOM clustering is a powerful tool for the analysis of behavior of bank lending.

References

[1] T. Ito and T. Iwaisako (1996). Explaining asset bubbles in Japan, M onetary and Economic Studies, 14, 143-193.

[2] K. Kiviluoto (1998). Predicting bankruptcies with the self-organizing map, N eurocomputing, 21, 191-201.

[3] T. Kohonen (1982). Self-organized formation of topologically correct feature maps, Biological C ybernetics, 43, 59-69.

[4] T. Kohonen (2001). Sel f -Or ganizin g M a ps (3rd ed.), Springer, Berlin, Germany.

[5] M. Ohkita, H. Tokutaka, K. Fujimura and E. Gonda (2008). Sel f -Or ganizin g M a ps and the Tool, Springer, Tokyo, Japan. (in Japanese)

[6] P. Sarlin (2013). Exploiting the self-organizing financial stability map, Engineering Applications of Artificial Intelligence, 26, 1532-1539.

[7] P. Sarlin and D. Marghescu (2011). Visual predictions of currency crises using self-organizing maps, Intelligent S ystems in Accounting, Finance and M anagement, 18, 15-38. A Cluster Analysis of Bank Lending Behavior by Using Self-Organizing Map: The Case of Japan 151

Appendix: Example of input data of SOM clustering f or time series

8 #dimlist I C E T W F R S

128476 76664 22540 40667 329237 25048 70209 75048 1977 154904 85145 25450 44584 357020 35783 81699 90578 1978 173835 88072 30985 46789 382024 44396 86203 103707 1979 187217 92165 35464 50810 401809 51312 90798 11415 8 1980 200767 101603 40883 57539 430208 65162 102212 132768 1981 211916 112 014 45635 64555 466354 86838 118 5 2 2 158767 1982 223719 126636 51067 70952 499423 118 4 7 6 139720 194342 1983 236182 140534 54271 76006 534923 153452 164567 231250 1984 257353 155098 55793 84855 561782 179019 2 0117 0 273610 1985 296198 16 3711 55261 91556 572168 225752 266394 325182 1986 372097 170846 55647 100102 592487 287791 312464 396010 1987 442297 18114 9 54275 109339 604602 317395 358133 448610 1988 541436 191766 51083 117110 630914 367492 409857 512592 1989 6117 6 5 199776 51327 128374 655883 377002 424269 578322 1990 649644 215777 53440 139607 642785 361474 447248 602764 1991 772827 286483 58596 158816 812992 467642 517696 715234 1992 768200 297977 59312 161588 804811 476830 544549 728037 1993 762586 306654 60101 16115 4 796620 493075 559121 733020 1994 808917 31128 3 61483 169425 781410 496155 573592 746964 1995 842948 307480 63596 182953 772351 466800 593247 755675 1996 872839 310874 57093 191603 766231 473738 613960 752223 1997 891559 313388 49959 17 3 011 7 5 2911 448479 618169 716021 1998 906191 300166 54871 177215 722979 417526 577052 659299 1999 927167 288362 53453 187743 698682 397471 569555 634393 2000 952409 264778 47623 185640 651218 370509 550532 594342 2001 982638 232407 40954 183676 592209 362502 517869 55113 2 2002

International Journal of Intelligent Technologies and Appld ie Statt is ics

Vol.9, No.2 (2016) pp.153-167, DOI: 10.6148/IJITAS.2016.0902.05 © Airiti Press

Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations

Giam Quang Do1* and Chaiwat Nimanussornkul2

1Faculty of Accounting and Business M anagement, V ietnam National University of Agriculture, Hanoi, V ietnam 2Faculty of Economics, Chiang M ai University, Chiang M ai, Thailand

ABSTRACT The paper examines the correlations among the 5 ASEAN emerging stock markets (Indonesia, Malaysia, Philippines, Thailand and Vietnam), and international gold market. Multivariate GARCH models i.e., CCC-GARCH and DCC-GARCH are employed to evaluate the conditional correlations between these markets. We find that the estimates of conditional correlations in the CCC and DCC models, on average, exhibit low to medium levels. The gold and Vietnam stock markets are especially low correlated with each other and with the remaining stock markets. Moreover, the estimates of the DCC model show that the correlations between the sample market pairs are time varying. Interestingly, in the recent economic downturn, the downward trend appeared in the pair correlations between the gold and stock markets, while the rapid upward trend occurred in the pair correlations between the Vietnam and 4 remaining stock markets.

K eywords: Gold and ASEAN stock markets; Conditional correlations; MGARCH; CCC; DCC

1. Introduction

Studies on the univariate models of conditional volatility have appeared in the literature much more than those on multivariate extensions. A main restriction of the univariate volatility models is that they are modelled and estimated independently from the others. Usually, international stock markets respond to the same shock, at least in part, and the extent to which volatility is transmitted across markets as a change in the volatility of a market tends to cause changes in the volatility of others, so the univariate model may not be correctly specified. In finance, the conditional correlations and covariances between the markets

* Corresponding author: [email protected] 154 Do and Nimanussornkul reflect their possible interdependencies over time. Understanding how conditional correlations between the markets vary overtime is very substantial for constructing portfolio and risk management. Such the motivation, multivariate GARCH models can potentially overcome these deficiencies with their univariate counterparts. In fact, there have been a number of key researches, which examined the volatility transmission mechanism and market interdependencies among major developed stock markets. For instance, Hamao et al. [7], Susmel and Engle [19], and Koutmos and Booth [14] focused on New York, London, and Tokyo; Theodossiou and Lee [20] examined interdependencies between the US, Japan, Canada, and Germany; Koutmos [13] investigated the dynamic interdependence of major European stock markets. Meanwhile, few other works on this issue can be found between emerging and developed stock markets over the world e.g., in Asia-Pacific [16], in Asia [9], in Southeast Asia [18] and in Latin America [4]. In other aspects, there have been a number of studies that focused on cross-border volatility spillover in stock markets [7, 10, 12, 14, 17], in foreign exchange markets [1, 2, 8, 11], and in interest rate markets [21]. Nevertheless, these papers tend to focus on some specific financial markets. Regarding ASEAN stock markets, the studies have been done usually involve with the developed markets, in terms of market integration, correlations and volatility spillovers. The recent economic downturn is happening over the world, leading to a strong decline in global stock markets. In the Association of South East Asian Nations (ASEAN), Vietnam, Thailand, Philippines, Malaysia and Indonesia have emerging stock markets, which are affected negatively from the crisis i.e., sharp declines in its stock markets. Together with high volatility in the stock markets, the world gold market volatility tends to be high during the recent time, due to the worried psychology of investors about a long economic crisis and high inflation. That is the reason why prices of gold have continuously achieved new high records in recent years. However, how correlations among ASEAN emerging stock and gold markets are going on over time, particularly during the recent economic crisis, have not been known. Changes in the conditional correlations between these markets may have significant effects on the portfolio management. To date, no studies have been found that show the conditional correlations among the international gold and ASEAN emerging stock markets. The purpose of this paper is to examine the conditional correlations among ASEAN emerging stock and international gold markets, using the constant conditional correlation (CCC) model and the dynamic conditional correlation (DCC) model that are of the most widely employed multivariate GARCH specifications for studying the dynamic conditional correlations. The remaining of the paper is organized as follows: Section 2 provides data and summary statistics; Section 3 presents model specifications; Section 4 reports the empirical results; and Section 5 provides concluding remarks. Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 155

2. Data and basic statistics

The paper used daily closing indexes of 5 emerging stock markets in ASEAN, namely JKSE, KLSE, PSE, SET and VN-Index (VNI), which respectively represent for the stock markets in Indonesia, Malaysia, the Philippines, Thailand and Vietnam, together with daily prices of the PM London Gold Fix (GOLDFIX), which represents the benchmark for the official gold trading over the world. The sample period for analysis is from July 28, 2000 (the day that Vietnam stock market -- the youngest one opened for trading) to March 31, 2009. Daily data of the 5 stock indexes were downloaded from Reuter and the gold prices were obtained from www.kitco. com. Before doing time series analysis, the data should be checked the statistical adequacy i.e., stationarity or nonstationarity. The two most common methods for unit root test, namely augmented Dickey-Fuller (ADF) and Perron-Phillips (PP) tests were employed. The test results indicate that the null hypothesis of a unit root in the six level series is failed to reject, implying that the 5 stock market indexes and gold prices are nonstationary. However, the null hypothesis of a unit root in the first difference of the 6 level series is strongly rejected, these conclusions are commonly seen in financial time series data (the detail results of the tests are not reported here), so all the 6 daily return series are stationary.

Daily returns of the selected markets, ri,t, are computed as ri,t = 100 × [ln( pi,t) – ln( pi,t–1)], where pi,t and pi,t–1 are the closing indexes of market i on the days t and t – 1, respectively. The plots of returns series of the sample markets are shown in Figure 1, indicating that the mean returns are constant but the variances change over time.

Figure 1. Daily returns of the 6 markets. 156 Do and Nimanussornkul

Table 1 reports a descriptive statistics for each of market return series. The statistics indicates that Vietnam stock market performs the highest mean return (0.099) along with the highest risk (1.713), while Malaysia stock market appears to show the lowest mean return (-0.006) and the lowest risk (0.929). However, gold market performs quite well with relatively high return (0.085), but less risky (1.208) as compared to those of the stock markets. Usually, it is assumed that the higher returns of a financial asset are associated with a higher risk and vice versa. In addition, Table 1 reveals some characteristics in the return series. For instance, all the series show the standard property of return data i.e., heavy-tailed distributions, as expressed by the positive coefficients of the excess kurtosis (> 3). This characteristic is also highlighted by the significant Jarque-Bera normality test. Hence, the distributional properties of the return series seem to be non-normal, meaning that the null hypothesis of normality can be rejected. The theoretical value of the skewness in the case of the normal distribution is zero, however all the return series possess negative skewness, implying that large negative shocks occur more frequently than large positive shocks, except in the Philippines. The heavily negative skewness exists in Malaysia, while the gold market shows slightly negative skewness.

Table 1. Preliminary statistics of stock market returns (from 28 July 2000 to 31 March 2009).

market returns GOLDFIX JKSE KLSE PSE SET VNI Mean 0.0849 0.0409 -0.0064 0.0261 0.0185 0.0991 Maximum 6.8414 11.7 07 3 4.1709 15.2558 10.5770 6.6561 Minimum -7.9719 -10.9539 -9.9785 -8.6981 -16.0633 -7.6557 S td. De v. 1.2084 1.5859 0.9290 1.4 511 1.4927 1.7130 Skewness -0.0891 -0.7508 -1.2384 0.4627 -0.7531 -0.1410 Kurtosis 8.0595 11.0 5 4 5 14.8446 13.8225 15.1686 5.3273 Jarque-Bera test 1,780.2 4,662.7 10,170.8 8,194.8 1,0442.6 381.7 Probability (< 0.001) (< 0.001) (< 0.001) (< 0.001) (< 0.001) (< 0.001) Obs. 1,667 1,667 1,667 1,667 1,667 1,667

3. Model specifications

Since the autoregressive conditional heteroscedasticity (ARCH) model was first invented by Engle [6] and extended then by Bollerslev [3] to be the generalized Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 157

ARCH (GARCH) model, a huge number of works have been done in the empirical studies relating to the univariate GARCH in various areas, especially in financial market volatility. Recently, the univariate GARCH have been developed to the multivariate GARCH (MGARCH) cases that can examine the volatility spillovers as well as the conditional correlations between the markets. To model the conditional correlations between the international gold and ASEAN emerging stock markets, we employed two MGARCH models such as CCC-GARCH and DCC-GARCH. In the MGARCH models, the default equation for the means can be constant, or AR( p), or ARMA( p,q). In general, the conditional mean equations of daily returns of the markets under our consideration in MGARCH models can be written as follows,

rit = E(rit | Ψt–1) + εit , with εit | Ψt–1 ~ N(μit,hit) (1)

εit = hzit it , with zit ~ iid( 0,1) ( 2)

Let i = 1, …, s be the number of the sample markets, t = 1, …, n the number of observations, rit return series of the selected markets, εit = rit – μit the residuals or shocks to market returns, hit the univariate time-varying conditional variances of the selected market returns, Dit = diag() hit a diagonal matrix, Ψt–1 the past information available at time t, the standardized residuals to the zit = εit / hit selected market returns. In constructing the multivariate conditional variance equations, Bollerslev [2] introduced the constant conditional correlation (CCC) model. As defined in Equations

(1) and (2), the conditional covariance matrix, H t, in the CCC model is written as follows

H t = E(εt εt' | Ψt–1) = E(Dtztzt'Dt) = DtE(ztzt')Dt = DtRDt (3)

Let R = E(ztzt') = { ρik} be a symmetric positive definite matrix that ρik =ρki with

ρik = 1, i = k (for i, k = 1, ..., s). Hence, R is the matrix of the constant conditional correlations, {ρik}, between different pairs of the market returns. The CCC model assumes that the univariate conditional variance for the return series, hit, follows a univariate GARCH process [3] as

pq 2 hhit=++ω i∑∑ αε ijitj,,−− β ijitj (4) jj=11= 158 Do and Nimanussornkul

Let αij be the ARCH effects implying the short-run effects of shocks, βij the GARCH effects or the contribution of such shocks to long-run persistence (αij + βij). Bollerslev

[3] indicated that ωi > 0, αij ≥ 0 for j = 1, …, p and βij ≥ 0 for j = 1, …, q are sufficient pqαβ+< conditions for a positive conditional variance hit > 0, and ∑∑j=11 ij j= ij < 1 is the necessary and sufficient condition for the existence of the second moment. The 2 simplest case is GARCH(1,1) model, i.e., ht t−11 ++= βεαϖ ht−11 , but has been most widely used in practice. The assumption of the constant conditional correlations in the CCC models may often be reasonable over shorter time periods, however the DCC model is not advisable as the series are very short and also when time-varying volatility is not an issue [15]. To relax this assumption, Engle [5] proposed the dynamic conditional correlation (DCC) model, which is a generalization of the CCC model. The DCC model can be expressed as follows

εt | Ψt–1 ~ N(0,H t), H t = DtRtDt (5)

Let H be the conditional covariance matrix, ( h ) adiagonal matrix t t = diagD hit it )( of the univariate conditional variance equations, Rt the a conditional correlation matrix. The conditional variance (hit) in the Dt is assumed to follow a univariate GARCH model as given in Equation (4). The difference between DCC and CCC models is that DCC model allows the conditional correlation matrix, Rt, to be time varying i.e., Rt = {ρik,t}. The DCC estimation of conditional variances and correlations is conducted through two stages, so that the estimation of time-varying correlation matrix is easier. For instance, in the first stage univariate volatility parameters in Equation (4) are estimated for each return series, using GARCH model and so the standardized shocks, , are obtained. In the second stage, the zit = εit / hit standardized residuals, zit, obtained from the first stage are used to estimate the parameters of the dynamic conditional correlations, qik,t. In our study, we employ the DCC (1,1) version of Engle [5], so the model can be written as follows

-1/2 -1/2 Rt = {diag(qii,t) } qik,t {diag(qkk,t) } = {ρik,t}, f or i, k = 1, ..., s and t = 1, …, n (6)

(7) qikt,=−−(1θ 1 θρ 2 ) ik + θ 1zzit ,1−− kt ,1 + θ 2 q ikt ,1 −

Engle [5] defined that ρ ik in (7) is the unconditional correlation between zi,t and zk,t that Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 159 has unit variance, obtained from the first stage, and Equation (6) is used to standardize the matrix estimated in Equation (7). In Equation (7) θ1 and θ2 are parameters to be estimated, reflecting the short-term and long-term effects of shocks to the dynamic correlations, respectively. If the estimates of θ1 and θ2 are significantly different from zero, then conditional correlation in the whole is not constant. On the contrary, if the estimates of θ1 and θ2 are not significant, then qik,t in Equation (7) can be interpreted as the CCC model. The DCC model is estimated using the maximum likelihood estimator (MLE). Engle [5] showed the log-likelihood function as

nn 11 −−-1' 1' −= −= ∑∑sL sL ππ ||log)2log( ||log)2log( ++ ++ zRzR zRzR tttt tttt (8) 22tt==11

It is assumed that zt in Equation (8), the standardized residual series obtained from the first stage, = ε / hz ttt , is normally distriduted with zero mean and variance Rt (i.e., zt ~ N(0,Rt)). Interestingly, Rt plays a role as a variance matrix of the standardized residuals and also as a correlation matrix of the residual series, εit = rit – μit.

4. Empirical results

In this section, we first provide the estimates for the univariate conditional variance model, namely the GARCH(1,1) model in the selected markets based on three types of mean specification i.e., the unconditional mean, AR(1) and ARMA(1,1). All the estimates of the parameters are obtained using the Marquardt optimization algorithm in the Eviews 6 econometric software package. Results of the estimation are presented in Table 2. According to the least values of AIC and BIC, together with the statistically significant levels of the estimated parameters in mean equations, the AR(1) process is found in almost all the markets under investigation and also dominates the other two alternatives. The estimates of variance equations in the GARCH(1,1) model show that all the coefficients of the unconditional variance (ω), and the ARCH (α) and the GARCH (β) effects are positive and significant. Therefore, the AR(1)- GARCH(1,1) specification is statistically adequate for both the conditional mean and the conditional variance of those markets. 160 Do and Nimanussornkul

Table 2. The estimates of univariate GARCH(1,1) model.

Market Mean equation Variance equation AIC BIC returns Constant AR(1) MA(1) ω α β 0.0569 0.0093 0.0394 0.9559 3.004 3.017 (0.016) (< 0.001) (< 0.001) (< 0.001) 0.0574 -0.0165 0.0096 0.0394 0.9556 3.005 3.021 GOLDFIX (0.012) (0.551) (< 0.001) (< 0.001) (< 0.001) 0.0708 0.9551 -0.9838 0.0104 0.0396 0.9546 2.996 3.015 (< 0.001) (< 0.001) (< 0.001) (< 0.001) (< 0.001) (< 0.001) 0.1330 0.1446 0.1465 0.8002 3.548 3.561 (< 0.001) (< 0.001) (< 0.001) (< 0.001) 0.1358 0.1291 0.1364 0.1400 0.8086 3.537 3.553 JKSE (< 0.001) (< 0.001) (< 0.001) (< 0.001) (< 0.001) 0.1359 0.0436 0.0869 0.1367 0.1405 0.8080 3.557 3.545 (< 0.001) (0.823) (0.664) (< 0.001) (< 0.001) (< 0.001) 0.0265 0.0126 0.1432 0.8599 2.439 2.453 ( 0.121) (< 0.001) (< 0.001) (< 0.001) 0.0260 0.1506 0.0107 0.1364 0.8668 2.423 2.439 K LSE ( 0.191) (< 0.001) (< 0.001) (< 0.001) (< 0.001) 0.0257 0.3197 -0.1746 0.0106 0.1349 0.8682 2.424 2.443 (0.212) (0.053) (0.318) (< 0.001) (< 0.001) (< 0.001) 0.0671 0.2201 0.1407 0.7633 3.469 3.482 (0.017) (< 0.001) (< 0.001) (< 0.001) 0.0695 0.1067 0.2168 0.1352 0.7683 3.461 3.477 PSE (0.031) (< 0.001) (< 0.001) (< 0.001) (< 0.001) 0.0694 0.1754 -0.0688 0.2177 0.1352 0.7678 3.462 3.481 (0.038) (0.438) (0.769) (< 0.001) (< 0.001) (< 0.001) 0.0706 0.40060 0.1517 0.6609 3.496 3.509 (0.054) (< 0.001) (< 0.001) (< 0.001) 0.0728 0.117 4 0.4236 0.1636 0.6371 3.486 3.502 SET (0.074) (< 0.001) (< 0.001) (< 0.001) (< 0.001) 0.0727 0.0667 0.0508 0.4218 0.1632 0.6384 3.487 3.507 (0.076) (0.756) (0.814) (< 0.001) (< 0.001) (< 0.001) 0.0024 0.0295 0.3384 0.6989 3.230 3.243 (0.891) (< 0.001) (< 0.001) (< 0.001) 0.0080 0.2718 0.0316 0.3129 0.7171 3.166 3.182 VNI (0.728) (< 0.001) (< 0.001) (< 0.001) (< 0.001) 0.0076 0.0696 0.2264 0.0314 0.3147 0.7149 3.164 3.183 (0.406) (0.016) (< 0.001) (< 0.001) (< 0.001)

N otes: The f igures in parentheses are the p-values. Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 161

Normally, the interaction between two assets or markets is referred as correlation. Correlation is a measure used to describe how prices of two investments move in relation to each other. Positively correlated asset classes move in the same direction, whereas negatively correlated ones move in opposite directions. Since we focus on the market level, MGARCH estimations of the conditional correlations in the sample markets have been implemented to see how closely returns of the pair markets move together, using the CCC and DCC approaches. The DCC model was employed to capture the time-varying conditional correlations, contrary to the benchmark CCC model, which maintains the constant conditional correlations. The estimates of the constant conditional correlations for the CCC model in the selected market returns are presented in Table 3. It indicates that all the estimated correlations are positive. For the 6 six selected markets, there exist 15 pair correlations. The estimated results reveal that these correlations are statistically significant difference from zero in most of the cases, except two market pairs (Vietnam, Malaysia) and (Vietnam, Thailand). Generally, the conditional correlations between the selected market returns are below medium level (< 0.4), ranging from 0.0208 to 0.3846. The highest correlation is found in the market pair (Indonesia, Thailand), while the lowest one is in the case of (gold and Thailand). Especially, gold and Vietnam stock markets display very low correlations (< 0.1) with other sample markets in returns. In terms of international investment, it is very crucial to note that portfolio managers, who apply the top-down approach should target their investment allocations in different market groups that exhibit low correlations i.e., market groups (Thailand, Vietnam, gold), (Indonesia, Vietnam, gold), (Malaysia, Vietnam, gold) and (Philippines, Vietnam, gold). Picking Indonesia, Malaysia, Thailand and the Philippines markets together could not be advisable, since these 4 markets show more interdependence to each other via higher degrees of correlations (0.2417 to 0.3846) as compared to the remaining sample pairs.

Table 3. Constant correlations between the selected market returns in the CCC model.

Market GOLDFIX JKSE KLSE PSE SET returns JKSE 0.0595 (0.001) KLSE 0.0661 (0.005) 0.3767 (< 0.001) PSE 0.0648 (0.005) 0.3217 (< 0.001) 0.3231 (< 0.001) SET 0.0208 (0.354) 0.3846 (< 0.001) 0.3627 (< 0.001) 0.2417 (< 0.001) VNI 0.0364 (0.097) 0.0427 (0.093) 0.0320 (0.200) 0.0923 (< 0.001) 0.0264 (0.289)

N otes: The f igures in parentheses are the p-values. 162 Do and Nimanussornkul

The estimates of the DCC parameters for the dynamic conditional correlations in the sample markets, based on estimating univariate GARCH(1,1) models for each return series, are reported in Table 4. It indicates that the DCC parameters,

θ1 and θ2 are positive and significant different from zero, which make clear that the assumption of the constant conditional correlations between the sample market returns is not supported empirically, so the conditional correlations in the whole are time varying.

Table 4. Estimates of the DCC parameters for the selected market returns.

DCC parameters Estimates

θ1 0.0089 (< 0.001)

θ2 0.9742 (< 0.001)

θ1 + θ2 0.9831

N otes: The f igures in parentheses are the p-values.

Table 5 reports a descriptive statistics of the time varying conditional correlations, obtained from the DCC model. Generally, the sample means of the time varying conditional correlations in all the 15 market pairs exhibit low to medium interdependence over time. However, they lie in a wider range [-0.0093, 0.4224] as compared with the range [0.0208, 0.3846] of the constant conditional correlations given in Table 3. It is meaningful to realize that, on average, the time varying conditional correlations between the gold and 5 stock markets estimated in the DCC model are lower than those estimated in the CCC model (Figure 2), whereas the time varying conditional correlations in the 10 pairs between the 5 stock markets (Indonesia, Malaysia, Philippines, Thailand and Vietnam) estimated in the DCC model are all higher than those estimated in the CCC model. The DCC model also indicates a more critical role of gold as a diversified asset in the portfolios and a lesser attractiveness of Vietnam stock market in hedging strategies than the CCC model over the sample period. Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 163

Table 5. Basic statistics of the estimated dynamic conditional correlations in the market pairs.

Mean Maximum Minimum S td. d e v. C.V. GOLDFIX_JKSE 0.0450 0.1993 -0.1706 0.0553 122.87 GOLDFIX_KLSE 0.0518 0.2044 -0.1712 0.0470 90.68 GOLDFIX_PSE 0.0545 0.2082 -0.2003 0.0466 85.49 GOLDFIX_SET -0.0093 0.119 5 -0.3331 0.0561 -601.48 GOLDFIX_VNI 0.0132 0.1110 -0.118 0 0.0292 221.30 JKSE_KLSE 0.4224 0.6420 0.2540 0.0642 15.20 JKSE_PSE 0.3475 0.5172 0.2 211 0.0529 15.22 JKSE_SET 0.4164 0.5939 0.2547 0.0557 13.37 JKSE_VNI 0.0734 0.2948 -0.0345 0.0443 60.29 KLSE_PSE 0.3565 0.5876 0.1638 0.0614 17.23 KLSE_SET 0.3899 0.6313 0.2251 0.0582 14.93 KLSE_VNI 0.0681 0.2893 -0.0419 0.0416 61.09 PSE_SET 0.2617 0.5694 0.1422 0.0551 21.06 PSE_VNI 0.1284 0.4441 0.0234 0.0572 44.54 SET_VNI 0.0676 0.2982 -0.0836 0.0464 68.68

Figure 2. Plots of the correlations between the market pairs in the CCC and DCC models.

Through the sample means of the time varying conditional correlations given in 164 Do and Nimanussornkul

Table 5, we find that, on average, the gold and Vietnam stock markets express very low to negative correlations with the other sample markets (-0.0093 to 0.1284) over time, whereas the 4 remaining stock markets (Indonesia, Malaysia, Philippines and Thailand) exhibit relatively low to medium correlations with each other (0.2617 to 0.4224). These are consistent with the findings in the CCC model. Table 5 provides the coefficient of variation (C.V) in the dynamic conditional correlations through the 15 market pairs over the sample period. Actually, the dynamic conditional correlations in a market pair with smaller C.V are less dispersed from the mean than those in other pairs with larger C.V. The highest C.V is found in the market pair (Gold, Thailand), while the lowest one is the case (Indonesia, Thailand). Overall, high C.V in the dynamic conditional correlations are found between gold and the 5 stock markets (85.5% to 601.5%), while the medium cases (44.5% to 68.7%) occur in the pairs between Vietnam and other 4 stock markets, and the low ones (13.4% to 21.1%) exist in the remaining pairs. The dynamic paths of the conditional correlations in the 15 pairs of the sample markets obtained from the DCC model are plotted in Figure 3. The plots illustrate a varying pattern in the correlation dynamic paths, so the constant conditional correlations appear to be refuted as considerable variations occur in all the pairs over the sample period. Especially, large variations in the conditional correlations can be seen in all the pairs over the end period of the time path, which denotes the recent financial and economic crisis. Slightly upward trend in the conditional correlations appears in the market pairs (Malaysia, Indonesia), (Malaysia, Philippines), (Malaysia, Thailand), (Indonesia, Philippines), (Indonesia, Thailand) and (Philippines, Thailand) over the sample period, which reveals increasing interdependencies over time. Meanwhile, no clear trend seems to be observed in the plots of the remaining market pairs that consist of gold and/or Vietnam, over the sample period. However, during the end period of the time path, the pattern of correlations tends to change, as the sharp downward trend from low to negative levels occurs in the 5 pair correlations between the gold and 5 selected stock markets, whereas the rapid upward trend appears in the 4 pair correlations between Vietnam and the 4 remaining stock markets, meaning that Vietnam stock market could be more integrated with the other ASEAN emerging stock markets. Interestingly, these behaviors of the conditional correlations change suddenly in the end period of the time path and then they are adjusted gradually. This could be due to behavior of investors in restructuring the portfolio during the recent financial crisis. Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 165

Figure 3. Dynamic paths of the estimated conditional correlations.

5. Concluding remarks

The paper provides an insight on the interdependencies among ASEAN emerging stock and international gold markets through examining the conditional correlations between them. Multivariate GARCH models i.e., the CCC and DCC models were applied to see how closely returns of each market pair move together. The estimates of the CCC and DCC models, in general, exhibit medium to low interdependencies between the sample markets. Typically, gold and Vietnam stock markets display very low to negative correlations with the remaining sample markets over time, while Indonesia, Malaysia, Philippines and Thailand stock markets show medium correlations with each other. 166 Do and Nimanussornkul

The estimates of the DCC model show a varying pattern in all the pair correlation dynamic paths over the sample period. The interdependencies among Indonesia, Malaysia, Philippines and Thailand stock markets are growing over time. Moreover, during the recent financial crisis, the downward trend in the correlations between gold and ASEAN emerging stock markets highlights more attractiveness of gold as a diversified asset in the portfolios, while Vietnam stock market could be more correlated with other ASEAN markets, reducing its attractiveness in hedging risks when investing in ASEAN markets.

References

[1] R. T. Baillie, T. Bollerslev and M. R. Redfearn (1993). Bear squeezes, volatility spillovers and speculative attacks in the hyperinflation 1920s foreign exchange, Journal of International Money and Finance, 12, 511-521.

[2] T. Bollerslev (1990). Modelling the coherence in short-run nominal exchange rates: A multivariate generalized Arch model, Review of Economic and Statistics, 72, 498-505.

[3] T. Bollerslev (1986). Generalised autoregressive conditional heteroscedasticity, J ournal of Econometrics, 31, 307-327.

[4] A. Christofi and A. Pericli (1999). Correlation in price changes and volatility of major Latin American stock markets, Journal of Multinational Financial M anagement, 9, 79-93.

[5] R. F. Engle (2002). Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models, Journal of Business and Economic Statistics, 20, 339-350.

[6] R. F. Engle (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation, Econometrica, 50, 987-1007.

[7] Y. Hamao, R. W. Masulis and V. Ng (1990). Correlations in price changes and volatility across international stock markets, Review of Financial Studies, 3, 281-307.

] [8 Y. Hong (2001). A test for volatility spillover with application to exchange rates, J ournal of Econometrics, 103, 183-224.

[9] F. In, S. Kim, J. H. Yoon and C. Viney (2001). Dynamic interdependence and volatility transmission of Asian stock markets: Evidence from the Asian crisis, International Review of Financial Analysis, 10, 87-96. Examining Interdependencies among International Gold and 5-ASEAN Stock Markets through the Conditional Correlations 167

[10] G. A. Karolyi (1995). A multivariate GARCH model of international transmission of stock returns and volatility: The case of the United States and Canada, Journal of Business and Economic Statistics, 13, 11-25.

[11] C. Kearney and A. J. Patton (2000). Multivariate GARCH modelling of exchange rate volatility transmission in the European monetary system, Financial Review, 35, 29-48.

[12] M. A. King and S. Wadhwani (1990). Transmission of volatility between stock markets, Review of Financial Studies, 3, 5-33.

[13] G. Koutmos (1996). Modelling the dynamic interdependence of major European stock markets, Journal of Business Finance and Accounting, 23, 975-988.

[14] G. Koutmos and G. G. Booth (1995). Asymmetric volatility transmission in international stock markets, Journal of International Money and Finance, 14, 747-762.

[15] J. M. Lebo and J. M. Box-Steffensmeier (2008). Dynamic conditional correlations in political science, American Journal of Political Science, 52, 688-704.

[16] Y. A. Liu and M.-S. Pan (1997). Mean and volatility spillover effects in the U.S. and Pacific-Basin stock markets, Multinational Finance Journal, 1, 47-62.

[17] F. Longin and B. Solnik (1995). Is the correlation in international equity returns constant: 1960-1990? Journal of International Money and Finance, 14, 3-26.

[18] M. K. P. So, K. Lam and W. K. Li (1997). An empirical study of volatility in seven southeast Asian stock markets using ARV models, Journal of Business Finance and Accounting, 24, 261-276.

[19] R. Susmel and R. F. Engle (1994). Hourly volatility spillovers between international equity markets, Journal of International Money and Finance, 13, 3-25.

[20] P. Theodossiou and U. Lee (1993). Mean and volatility spillovers across major national stock markets: Further empirical evidence, Journal of Financial Research, 16, 337-350.

[21] Y. Tse and G. G. Booth (1996). Common volatility and volatility spillovers between U.S. and Eurodollar interest rates: Evidence from the futures market, Journal of Economics and Business, 48, 299-312.

International Journal of Intelligent Technologies and Appld ie Statt is ics

Vol.9, No.2 (2016) pp.169-189, DOI: 10.6148/IJITAS.2016.0902.06 © Airiti Press

Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas

Chanita Panmanee1* and Aree Wiboonpongse1,2

1Faculty of Agriculture, Chiang M ai University, Chiang M ai, Thailand 2Faculty of Economics, Prince of Songkla University, Songkla, Thailand

ABSTRACT This paper focuses on the determinants and practices of Arabica coffee growers in Pang Ma-O and Pamiang villages for planning and developing green cluster supply chains (GCSC). Samples selected in this paper are coffee growers in two villages, staff of the Highland Research and Development Institution (HDRI), and the Royal Project Foundation (RPF), and officers of the Pang Ma-O Extension Project, and the Pamiang Royal Project Development Center. Multivariate probit (MVP) model, cluster mapping and modified GEM model were used as tools for analyzing. The findings revealed that education, age and input cost have significantly positive influences on the green adoption whereas farm size has a negative impact on it. Considering cluster mapping, it is not complicated so the plan and development of GCSC is not too difficult. For the determinants involving in GCSC, the results not only showed the macroeconomic environmental determinants such as government policies, geographical position and total domestic demand, but also the microeconomic environmental determinants, namely, green relation, structure and strategies, and resources which are the strength factors confirming that it is possible to develop the GCSC. However, the external market is a relative weakness determinant that is implied as the drawback of GCSC development. The findings of this paper are useful for planning and developing GCSC to enhance the competitiveness of Arabica coffee farmers in the highland area bringing about the strength of community, the ability to be self-reliance, and the enhancement of income and well-being of farmers in the selected area, and as a helpful resource for expanding the similar projects in other areas.

K eywords: Arabica coffee; Green cluster supply chain; Adoption; Multivariate probit model; Modified GEM model

1. Introduction

Nowadays, the awareness of environmental friendly issues have continuously increased and brought about a high demand of green quality goods and services,

* Corresponding author: [email protected] 170 Panmanee and Wiboonpongse whereas many economic sectors such as agricultural, industrial, and service sectors have attempted to improve their products for supporting the environment. What is the idea of being environmental friendly? It involves healthy food and drinking safely, healthy ecosystems, toxic-free communities, an appropriate and safe waste management, and the restoration of contaminated sites [6]. Increasing green food consumption trend and green food competition have led to strategic planning agendas of various agencies. At the same time, various firms are integrating their supply chain processes to lower costs and serve the consumers better. These two situations are not independent. The firms have to be associated with the suppliers and consumers to meet and even exceed the environmental expectations of their consumers and their relevant agencies [47]. Arabica coffee is one of the many products that have been likely to face this kind of situation. Arabica coffee is one of the economic crops in the northern region of Thailand. Many agencies and highland development projects such as the Royal Project Foundation (RPF) and the Highland Research and Development Institute (HRDI) bring this crop for growing and promoting in the northern highland areas. The approach is useful for creating jobs and revenue, and reducing opium cultivation and forest invasion. Because of the appropriate geography and climate of the highland areas in the northern region, Arabica coffee is successful in production and market acceptance, in terms of quantitative and qualitative aspects. The Arabica coffee production trend in the northern of Thailand indicates that the cultivation areas of Arabica coffee, especially in Chiang Mai and Chiang Rai provinces, have expanded from 38,885 rais in 2012 to 59,151 rais in 2015; resulting in the increase of yields from 6,145 tons to 7,327 tons at the same time [30]. In the present, the proportion between Robusta and Arabica coffee yields in Thailand is approximately 85:15 percent which is higher than the proportion in 2008 which has been equal to 93:7 percent. The main cause is not only the promotion and development of government and private agencies, but also the increase in coffee prices which is the crucial motivation for farmers to expand the space for Arabica coffee growing. However, under the changes from the traditional trade systems to free trade systems, especially the entering into the ASEAN Economic Community (AEC) in 2015, it allows for high international trade and competition of various agricultural products, such as palm oil, soy bean oil, raw silk, sugar, fruit and vegetables, and tea and coffee. These changes will inescapably affect Arabica coffee growers, particularly the farmers who receive the incomplete information. The lesser the capability of farmers gaining access to information, the more severe they will become from being affected by the trade liberalization. Considering the context of Arabica coffee growers in the highland areas like the Pang Ma-O village in and Pamiang village in , , although these farmers have been promoted and developed by the HRDI and RPF, and their major earnings have come from tea and coffee cultivating, most of them are faced with the problems of coffee quality and a Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 171 lack of cooperation in the communities. These issues are considered as a significant weakness for the farmers’ competitiveness. With the philosophy of sufficiency economy and based-area economic development perspectives, the strength of the community will take place if the local economy is able to be associated with the industries by means of the supply chain. Agricultural communities not only supply raw materials and labors, but they also play a role as upgrading partners by adding the value on agricultural product, providing products that meet the needs of the industry, and using the industry as the economic gateway to link the outside areas. The key and common issue in establishing a linkage in supply chain based on the agricultural community is the industrial cluster which is defined as the geographic concentrations of linked agricultural organizations, such as farmers, agri-business, government and private agencies, etc., in both horizontal and vertical linkages for achieving the goal of sustainable competitiveness enhancement in their field [32-34]. In addition, the supply chain development taking into account the environmental friendliness is another important means for creating the value for Arabica coffee supply chain and enhancing the competitiveness of Arabica coffee growers in the highland areas [18, 38, 45, 47, 48]. There is a research question regarding on how to develop the Arabica coffee supply chain network in highland area by using the framework of the green cluster supply chain (GCSC), whereas the collaboration in cluster supply chains is concerned with being environmentally friendly. The precious way to form the GCSC is to know the conditions and environmental factors before the cluster development in the green supply chains of Arabica coffee of the highland farmers. Therefore, the aim of this paper is to explore the determinants of the green supply chain adoption and practice of Arabica coffee growers in the Pang Ma-O and Pamiang villages: these are the agricultural communities in the highland areas, for gathering the data-based and knowledge used in planning and developing the GCSC. The rests of this paper is structured as followed: Section 2 displays the theoretical framework. Section 3 presents the methodology used in this paper, Section 4 represents the empirical results, and Section 5 discusses and summarizes the study’s findings.

2. Theoretical f ramework

2.1 Green supply chain

The concern about the environment is the crucial issue affecting the economic and ecological systems. Various organizations have taken an interest in sustaining the environment by using the green supply chain management (GSCM) as a tool for improving their images, increasing their profit, and enhancing their competitiveness [35, 37, 43]. Green supply chain (GSC) refers to the supply chain that is focused on the processes and cooperation networks which begins with the suppliers and all the way towards to the final consumers, such as production, transportation, waste 172 Panmanee and Wiboonpongse disposal, etc., on which the environmental impacts occurring have been taken into accounts [15, 43]. The GSCM consists of a green design concerning with the designing of products that have the least impinge on the environment over the product life cycle [14, 17, 47, 51, 52]; green purchasing involving in environmentally-conscious practices in the procurement of factors or intermediate products including reducing resources, eliminating waste, recycling and reuse, decreasing non-contaminant, and substituting materials without affecting material property [8, 17, 26, 47, 51, 53]; green production and processing dealing with the green technology used in production focusing on resource efficiency and waste reduction for environmental benefit, and production capability development such as reduced, reused, reproduced, and recycled components to maximize the value gain [13, 51]; green logistics regarding transportation, inventory, and warehouse with minimizing on the cost, loss, and greenhouse gas emission [2, 51]; green recycle [51]; and green waste management [2]. Thus, the GSC concept used in this paper focuses on green production, green waste management, and green logistics.

2.2 Cluster supply chain

Cluster supply chain (CSC) concept is originated from the integrating of supply chain management and industrial cluster concepts dealing with two or more supply chain network systems on cluster based in the specific area [50]. In the present, CSC is used as a tool for competitive advantage building which plays an outstanding role in the cluster enhancement and competiveness improvement of various units in the c l us t e r [ 49, 50 ]. The key success factor in CSC development is collaboration. Supply chain collaboration is defined as two or more chain alliances working together to plan and improve performance for achieving the goals and benefits through information sharing [23, 41]; decision synchronization [7, 41, 44]; resource sharing [39]; joint logistics [7]; and joint knowledge creation [7]. Simatupang and Sridharan [40], Barratt [5] and Mangan et al. [22] separate supply chain collaboration into two aspects - horizontal and vertical collaborations. The prior refers to the cooperation among two or more alliances in the same level (such as inter-farmers collaboration, etc.) whereas the latter is the cooperation in the different levels of supply chain (namely, farmers and assemblers collaboration, etc.). According to Xue et al. [49], supply chain collaboration is classified in three aspects, vertical, pooled horizontal and reciprocal horizontal collaborations.

2.3 Formation of green cluster supply chain concept

Green cluster supply chain (GCSC) is the integration between cluster and supply chain concerning about environmental friendly practices [50]. This relationship plays a crucial role in cluster upgrading and competitiveness improving. Cluster supply chain is difined as a network system consisting of different actors or firms belonging Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 173 to the same industry in a specific agglomeration location, i.e. suppliers, producers, retailers, consumers, government and private agencies, R&D institutions, etc. These actors constitute multiple local supply chain via the relationship of suppliers and consumers by the means of formal contract or informal trust and commitment. Consequently, there is an internal cooperation among the different actors of each individual single supply chain and an external across-chain coordination of the actors from different single supply chains at the industrial cluster. However, the GCSC concept is not only defined as a network system of different actors in supply chains, but is also involved in environmental concerning practices, such as green production, green waste management, and green logistics, etc. via environmental collaboration and monitoring [45]. In the GCSC formation process, the conditions achieving the successfulness are the collaboration and coordination among the local firms, government, local R&D institutions, institution for collaboration and financial institutions [42] as well as the consideration of environmental factors before cluster development, namely, the macroeconomic and microeconomic environmental factors. The macroeconomic environmental factors consist of regulations and culture in the area, geographical position, government policy, and macroeconomic factors [42] whereas microeconomic environmental factors feature demand determinants, firm structural conditions, related and supported industries, and the market. Accordingly, the conceptual framework of GCSC is represented in Figure 1.

3. Methodology

The population of this paper is Arabica coffee farmers in Pang Ma-O village, Chiang Dao District, and Pamiang village, Doi Saket District, Chiang Mai Province. Therefore, 69 coffee growers in Pang Ma-O village and 119 coffee growers in Pamiang village are selected as the samples by purposive sampling method for analyzing the factors affecting the growers’ adoption of green supply chain practices. Moreover, 14 of them in both areas who are willing to join the GCSC development project, 2 staff of the Royal project foundation, 2 staff of the HRDI and 2 officers of the Pang Ma-O Extension Project are selected for analyzing the determinants and practices of growers in the GCSC Thus, the methodology used in this research consists of the multivariate probit (MVP) model and the modified GEM model.

3.1 Multivariate probit model

The MVP model is the binary response regression model that estimates the effects of independent variables on various dependent variables at the same time [25]. In this study, the dependent variables, the growers’ adoptions of green supply chain practices, can be simultaneously occurred as the three choices (adoption of green production, adoption of green waste management, and adoption of green logistics), so the MVP model is appropriate for determining the factors influencing on them. 174 Panmanee and Wiboonpongse GCSC development framework. GCSC Figure 1. Figure Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 175

The MVP model is based on standard random utility maximization [24] and normal distribution [16]. This model is suitable utilization when the dependent variables are very intimately linked and affected by the same independent factors. The MVP is specified around a latent variable whose level is influenced by explanatory variables. Thus, the coffee growers’ utility by choosing three alternatives of green supply chain practices are represented as follows:

U = β'X + e (1) where U is a vector of growers’ utility, X is a vector of independent variables, β is a vector of unknown parameters, and e is a vector of stochastic error term. The MVP model estimates the parameters, β, and the variance covariance matrix of the multivariate normal distribution of the error terms. For the ith growers (i = 1, 2, …, 169), if the expected utility of grower i on the alternative k is more than 0, then he selects the alternative k, and the dependent variable, Y ik, is equal to 1. On the other hand, if the expected utility of grower i on this alternative is less than 0, he does not choose it, and the dependent variable becomes 0 [4]. The value of dependent variables, thus, can be described as follows:

Y ik = 1 if U ik > 0 (2) = 0 otherwise

The three choice probability conditioned on parameters β, Σ and a set of independent variables, X, can be written as:

βφ Σ = Σ PrYik ,∫∫∫ (e123 , e , e 0, ) de 1 de 2 de 3 (3) RR123 R where φ is the density function of a multivariate normal distribution with mean vector 0 and the variance-covariance matrix, Σ, and Rk is the interval (–∞, βk'Xk) if

Y ik = 1 and ( βk'Xk, ∞) if Y ik = 0 [9]. Then, the maximum likelihood estimation (MLE) method is used to estimate the parameters, β, and the three correlations of the error terms. According to the theoretical concept above, the variables for multivariate probit model covering technology, socio-economic, institutional, environmental, and household specific perspectives [1, 3, 4, 12, 19, 20, 27-29, 36] are expressed in Table 1. 176 Panmanee and Wiboonpongse

Table 1. Variables and descriptions for multivariate probit model.

Variables Descriptions De pendent variable: 1 if the coffee grower willing to adopt green production; 0 GPA otherwise. 1 if the coffee grower willing to adopt green waste management; 0 GWA otherwise. GLA 1 if the coffee grower willing to adopt green logistics; 0 otherwise. Inde pendent variables: Gender 1 if the coffee grower is a male; 0 otherwise. Age Age of the coffee grower (years). Coffee grower education levels (1 = lower than the primary school Education level, 2 = the primary school level, 3 = the secondary school level, 4 = the bachelor degree, 5 = higher than the bachelor degree). Experience Farming experience of the coffee grower (years). Green attitude Attitude of green supply chain practices (scores). Farm size Farm sizes (rais). Input cost problems of the coffee grower (1 = lowest, 2 = low, 3 = Cost of input moderate, 4 = high, 5 = highest). Information accessibility of the coffee grower (1 = poor, 2 = fair, 3 = Information accessibility average, 4 = good, 5 = excellent).

3.2 Modified GEM model

The analyses of determinants and practices in GCSC consist of three aspects: the interdependence among the actors in cluster supply chain, the macroeconomic environmental factors, and the microeconomic environmental factors. Consideration of the interdependence among the actors in cluster supply chain of Arabica coffee is featured in both horizontal and vertical clusters. The actors in horizontal cluster are identified as the production unit, government agencies, local R&D institutions, financial institutions, collaborative institutions, and Arabica coffee growers, whereas the actors in vertical cluster are difined as the growers, and their stakeholders in the Arabica coffee supply chain. The efficient tool used for analyzing is cluster mapping as seen in various studies such as Condliffe et al. [10] and Villanueva et al. [46]. In terms of macroeconomic environmental factors, the analyzing is composed of regulations and culture in the area, geographical position, government policy, and macroeconomic factors [42]. For the microeconomic environmental factors perspective, modified GEM model Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 177 is used for analyzing. This model is expanded into the traditional GEM (Grounding, Enterprise, and Market) model of Padmore and Gibson [31] by adding Porter's five competitive forces and green concepts. This model successfully turned qualitative variables into quantitative indicators and was applied to evaluate the agriculture or food cluster [21]. In terms of modified GEM model analysis, the steps for evaluating are as follows:

Step 1: Interviewing the HRDI officers, the relevant experts and committees of the grower group to find the specific determinants and determinant weights. The Analytic Hierarchy Process (AHP) is adopted to evaluate the relative importance of each indicator. There are 3 microeconomic dimensions, 8 determinants and 37 indicators shown in Table 2.

Table 2. Microeconomic environmental indicators in GEM model.

Dimension Determinants Indicators natural resources, capital, human resources, techniques, Resources technology information. Grounding transport and communication, business environment, Infrastructures policy, associations, R&D institutes and university. supplier strength, quality of suppliers, cooperation Suppliers and related with suppliers, related agencies strength, quality of agencies related agencies, cooperation with related agencies. unanimous in group, ownership, production and Structure and strategies strategies plan. Enterprise bargaining power of suppliers, bargaining power Competition of buyers, intensity of rivalry, threat of substitute products, threat of new entrants. green production management, green transportation, Green relation green disposal of waste. scale of local markets, local market share, growth and Local markets opportunity, buyer boundaries, specific demand. Market distances to external markets, scale and growth External markets rate, external market share, characteristics of final consumers, entry to external markets.

Step 2: Scoring individual determinant by using 1 ~ 10 rating scores, from the highest score to the lowest score. Step 3: Calculating the average scores of each determinant and total modified GEM score. The formulas are shown in the following Equations (4) and (5): 178 Panmanee and Wiboonpongse

N N N ii i DETDETDET. . SCORE SCORE. SCOREin in =in = = ∑ ∑ ∑ S S /S /N N / N (4) n=n1=1n=1

3 3 3 2/32/32/3 GEMGEMGEM SCORE SCORE SCORE = = = 2.5 2.5 2.5 (DET (DET (DET . SCORE . SCORE . SCORE ) ) ) (5) (∏(∏(∏i=1i=1i=1 i i) i) )

i where Sn refers to individual indicator score in each determinant, n is sets of indicators of each determinant, n = 1, …, N and i is sets of determinants of each dimension, i = 1, 2, 3.

Step 4: Plotting the radar graph and interpreting the effect of environmental factors on GCSC development.

4. Empirical results

4.1 Green cluster supply chain adoption of the cof fee growers

The empirical results obtained from the MVP model estimation are represented in Table 3. The McFadden’s R2 of the model is 0.1803. For the green production aspect, education of the coffee growers and the problems of input cost have significantly positive influences on the decision to adopt green production practice. The growers who have higher education levels can improve their skills and increase their ability to obtain, process, and use information dealing with green production. In the same way, the growers who highly face the problems of input cost, such as high wage of labors, high cost of chemical pesticide and fertilizer, etc., tend to change their productions to reduce cost via adopting the green production practice. Considering green waste management adoption, the result in Table 3 shows that the older coffee growers are likely to adopt this practice. Older growers have gained experience and knowledge over time, so they have the better ability to manage the waste than the younger ones. However, farm size has negative impacts on the decision to adopt green waste management practice. In the small land, the waste and pollution are lower than numerous space. Thus, the growers with small farm are able to control and manage the wastes easily. In case of green logistics practice shown in Table 3, the finding also indicated that higher education levels of coffee growers positively impinges on green logistics adoption. Higher education influences coffee growers’ attitudes and thoughts leading them to increasingly being open minded, making more rational decisions, and being able to analyze the benefit of applying the new practice. Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 179

Table 3. Multivariate probit analysis of Arabica coffee growers’ adoption of green supply chain practices.

Adoption of Adoption of Adoption of Variables green production green waste management green logistics Coefficients t-value Coefficients t-value Coefficients t-value Constant -1.6795 ** -2.277 -0.4202 -0.533 -0.0574 -0.089 Gender -0.0001 -0.271 -0.0001 -0.163 -0.0002 -0.417 Age 0.0062 0.707 0.0334*** 3.112 0.0009 0.098 Education 0.1536 ** 1.981 -0.1117 -1.473 0.1349 * 1.759 Experience -0.0233 -0.916 -0.0135 -0.454 -0.0050 -0.203 Attitude of green 0.0173 1.405 0.0113 0.818 -0.0130 -1.127 Farm size 0.0216 1.399 -0.0274** -2.026 0.0076 0.433 Cost of input 0.1462 * 1.655 -0.0381 -0.364 0.0922 1.093 Information -0.0670 -0.673 0.0 511 0.463 -0.0260 -0.277 accessibility Correlation: *** ρ12 = 0.6874

ρ13 = 0.1086

ρ23 = -0.1812 Number of observations: 188 Log likelihood [log L] = -320.4389

Log likelihood [log L0] = -390.8982 McFadden’s R-squarea = 0.1803

N ote: ***, **, * denote statistical significance at 1%, 5% and 10% levels, respectively. a 2 McFadden’s R is calculated by 1 – [log L / log L0] where log L and log L0 are the maximum value of log-likelihood function estimated at the full model and at only constant term, respectively.

4.2 Cluster mapping

Cluster mapping of Arabica coffee in Pang Ma-O and Pamiang villages is shown in Figure 2. The focal actor is Arabica coffee growers in this area. The inputs used in their production consist of land, water, labors, young plants, fertilizer and capital. The land in this area is fertile and suitable for growing Arabica coffee whereas the major source of water used for cultivating is natural water such as rain water. For the labor force perspective, most of all growers have used household labors as a core in the production process and hired nearby village labors in the harvesting period. In case of capital, the growers have constructed the farmer group for saving and lending in the agricultural purposes. Moreover, the farmer group has launched the coffee bean pawn project and the members have been able to bring this money for investing in their farms. In terms of young plant, the growers bought them 180 Panmanee and Wiboonpongse from external village, such as the Royal Project Foundation, Wawee village, etc., in the past, but now the young plants are bred by themselves and distributed in their village. These are resulted from the promotion and support of the extension project of the Royal Project Foundation.

Figure 2. GCSC of Arabica coffee in Pang Ma-O and Pamiang villages.

There are the government agencies and NGOs promoting and supporting the production and marketing. In addition, R&D institutions and universities, namely, the Royal Project Foundation, HRDI, Chiang Mai university, Maejo university, etc. have helped to create and transfer the knowledge for Arabica coffee growers. The outputs, Arabica coffee beans, are sold by directly distributing channel to the Royal Project Foundation, the middlemen, the processing companies and farmer group. When the middlemen and Arabica coffee farmer group purchased these products, they would be sent them to the other processing companies.

4.3 Macroeconomic environmental determinants

For government policy, the government agencies provide support for Arabica coffee growing in the highlands via R&D and trade barriers such as income tariff. For example, the commitments in the WTO are the import quotas which are equal to 5.25 tons of coffee beans with import tariff rate of 30 percent and out of quota with an import tariff rate of 90 percent. For the macroeconomic factors aspect, total domestic demand in Thailand shows that the overall trend of demand for coffee intake in the country has steadily increased. Moreover, the amount of café has also expanded resulting in the increase of demand for coffee beans to be used as raw material. In terms of import and export, Thailand imports the coffee beans from Vietnam in the highest proportion 94.99 percent and followed by Laos and France, Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 181

respectively. While the highest export source of coffee beans is Canada with the proportion about 47.66 percent and followed by Japan and USA, respectively [11].

4.4 Microeconomic environmental determinants

The microeconomic environmental determinants are received by interviewing 20 stakeholders consisting of HRDI staff, the Royal Project Foundation staff, Pang Ma-O Extension Project officers and committee of farmer group in Pang Ma-O and Pamiang villages. The Analytic Hierarchy Process (AHP) is used to determine the weights of individual indicators in each determinant. The results are expressed in the the following three dimensions, namely, grounding, enterprise and market. The first perspective, grounding, refers to the determinants of the supply side, such as resources and infrastructures. The result reveals that in the average score of research determinant in Pang Ma-O village including five indicators, e.g. natural resources, capital, human resources, technique and technology, and information, is equal to 6.660 whereas the same determinant in Pamiang village is equal to 6.393. The mean score of infrastructures determinant in Pang Ma-O and Pamiang villages comprising of transport and communication, business environment, policy, associations, and R&D institutes and University are equal to 5.401 and 6.597, respectively, as shown in Table 4.

Table 4. The scores of microeconomic environmental determinants and indicators in Pang Ma-O and Pamiang villages.

Pang Ma-O village Pamiang village Aver. Aver. Determinants Indicators score of Weights Scores score of Weights Scores indicators indicators Grounding dimension Natural resources 7.30 0.380 5.95 0.11 Capital 5.90 0.200 6.50 0.19 Human resources 7.00 0.190 5.55 0.18 Resources 6.660 6.393 Technique and 6.20 0.105 6.90 0.31 technology Information 5.80 0.125 6.50 0.21 Transport and 4.40 0.085 7.05 0.15 communication Business environment 5.30 0.170 6.80 0.14 Infrastr- Policy 4.80 0.135 5.401 6.00 0.14 6.597 uctures Associations 5.60 0.300 6.35 0.25 R&D institutes and 5.80 0.310 6.75 0.32 university 182 Panmanee and Wiboonpongse

Table 4. The scores of microeconomic environmental determinants and indicators in Pang Ma-O and Pamiang villages. (continued) Enterprise dimension Supplier strength 5.10 0.165 5.85 0.13 Quality of suppliers 6.00 0.230 6.05 0.25 Cooperation with 5.30 0.135 5.80 0.11 suppliers Suppliers Related agencies and related 6.20 0.155 5.880 6.20 0.15 5.983 strength agencies Quality of related 7.10 0.145 5.95 0.12 agencies Cooperation with 5.60 0.170 5.95 0.24 related agencies Unanimous in group 7.40 0.430 6.45 0.32 Structure and Ownership 7.10 0.320 6.20 0.41 7.029 6.078 strategies Production and 6.30 0.250 5.45 0.27 Strategies plan Bargaining power of 4.90 0.140 5.10 0.15 suppliers Bargaining power of 4.80 0.210 5.95 0.18 Competition buyers 5.525 5.902 Intensity of rivalry 5.50 0.290 6.25 0.3 Threat of substitute 6.50 0.230 6.10 0.25 products Threat of new 5.70 0.130 5.55 0.12 entrants Green production Green 7.80 0.470 4.60 0.3 management 7.114 4.628 relation Green transportation 5.80 0.190 4.80 0.38 Green disposal of 6.90 0.340 4.45 0.32 waste M arket dimension Scale of local markets 6.00 0.150 6.10 0.16 Local market share 5.80 0.180 6.40 0.21 Local Growth and 6.50 0.310 6.339 6.35 0.32 6.155 markets Opportunity Buyer boundaries 6.10 0.130 5.90 0.18 Specific demand 6.90 0.230 5.70 0.13 Distances to external 3.90 0.130 5.95 0.13 markets Scale and growth rate 4.70 0.350 5.80 0.29 External market External 4.30 0.120 5.40 0.19 share 4.656 5.468 markets Characteristics of 5.20 0.170 4.90 0.14 final consumers Entry to external 4.80 0.230 5.20 0.25 markets Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 183

In terms of the second aspect, enterprise, it represents the structural determinants of firms such as suppliers and related agencies, structure and strategies, competition, and green relation. Interestingly, the highest score of these indicators in Pang Ma-O village is green relation, around 7.114 average score, and following by structure and strategies, suppliers and related agencies and competition with 7.029, 5.880, and 5.525 average scores, respectively. On the other hand, the highest score of these indicators in Pamiang village is structure and strategies accounted for 6.078 average score, and following by suppliers and related agencies, competition and green relation with 5.983, 5.902, and 4.628 average scores, respectively. The last determinant, market, points to demand determinant being comprised of local markets and external markets. The result in Table 4 indicates that the average score of local markets indicators is higher than external markets indicators in both areas. The mean scores obtained from the above calculations are plotted as a radar graph to compare eight indicators, shown in Figure 3. The findings indicated that in general, the Arabica growers in Pang Ma-O village has quite the competitiveness, especially in an environmental friendly perspective that has the highest score, and in the structure and strategies of group with the second highest score. The two issues are the strengths and usefulness for the GCSC development of Arabica coffee. However, the external markets indicator is in the lowest score. It implies the weakness in development. Considering the Arabica growers in Pamiang village, the findings represented that the growers have a better strength in the resource aspect with the highest score, especially in the suitable geographic, whereas the perspective of green relation has the lowest score. The latter view is the weakness for the GCSC development of Arabica coffee in Pamiang village.

Figure 3. GEM score evaluation of Arabica coffee cluster in Pang Ma-O and Pamiang villages. 184 Panmanee and Wiboonpongse

In the rest of the analysis, the GEM score is calculated by using Equation (5). In generally, if this score is greater than 250, it indicates that there is a high level of competitiveness [21]. The results of the calculations in this paper showed that GEM score in Pang Ma-O and Pamiang villages are 355.25 and 449.01, respectively. That means the Arabica growers in both areas have the ability to compete at a high level.

5. Discussion and conclusion

The continuously increasing rate in green commodity consumption and the severe impacts of trade liberalization lead to the concernment about the competitiveness of all economic sectors, especially the agricultural sector in which the farmers are inescapably affected. Arabica coffee is one of various agricultural products influenced from these phenomena. The purpose of this paper is to explore the determinants and practices of Arabica coffee growers in Pang Ma-O and Pamiang villages for gathering the data-based and knowledge used in planning and developing the GCSC. The results of this paper are able to conclude in the following three aspects.

(1) The cluster mapping of Arabica coffee in Pang Ma-O and Pamiang villages helps for knowing the interaction among the actors in the cluster as a whole. The results reveal that the focal actor is Arabica coffee growers in this area. The inputs used in their production consist of land, water, labors, young plants, fertilizer and capital. Most of them are available in the communities. There is the cooperation in this area as such as establishing a farmer group for saving and lending in the agricultural purposes, and launching the coffee bean pawn project to help their members’ investment. Moreover, there are the government agencies and NGOs promoting and supporting the production and marketing and R&D institutions and universities helping to create and transfer the knowledge for Arabica coffee growers. For the outputs, Arabica coffee beans are sold by directly distributing channel to the Royal Project Foundation, the middlemen, the processing companies and farmer group. This cluster mapping indicates that the structure of Arabica coffee cluster in Pang Ma-O and Pamiang villages is not complicated, thus, the plan and development of the GCSC is not too difficult.

(2) The macroeconomic environmental determinants analysis is useful for farmers to prepare and adapt themselves from macro external conditions affecting Arabica coffee cluster. The result represents that government policies, geographical position and total domestic demand are the support determinants for Arabica coffee growing in the highland. Accordingly, in the GCSC development project, not only the government, non-government agencies and domestic consumer are the prominent actors in developing the GCSC but the geographical position and total domestic demand are the determinants confirming that the GCSC development is possible to happen. Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 185

(3) The microeconomic environmental determinants analysis helps the farmers to understand their identity in terms of both strength and weakness. The relative strength determinants consist of green relation, structure and strategies, and resources. These factors indicate that the focal unit, Arabica coffee farmers, have great conditions supporting the GCSC development. On the other hand, the external market is relative weakness determinant with the lowest score. The key reason may be due to the major market of coffee beans; in this area the local market is an everyday fact of life so the farmers do not pay much attention to the external market. Furthermore, the products, coffee beans, are sold via auction among the middlemen. The advantage of this method is that the farmers have more bargaining power, but they may also be an “occurring collusion phenomena” which can lead to a low price for the farmers’ to receive. As a matter of fact, the external market aspect is an very important issue for developing, especially when entering to AEC in the future. The continuous increase in coffee yields, the standard and quality improvement of products, and the eco-friendly image creation have to need expanding its market to the outside areas. Therefore, it is a significant determinant to take into account in developing the GCSC of Arabica cof f ee.

The findings of this paper are useful for planning and developing the GCSC to enhance the competitiveness of Arabica coffee farmers in the highland area, thus bringing about the strength of the community, the ability to have self-reliance, and the enhancement of income and well-being of farmers in the selected area, and being resource when expanding a similar project in other areas.

References

[1] M. A. Akudugu, E. Guo and S. K. Dadzie (2012). Adoption of modern agricultural production technologies by farm households in Ghana: What factors influence their decisions? Journal of Biology, Agriculture and Healthcare, 2(3), 1-13.

[2] E. Andiç, Ö. Yurt and T. Baltacıoğlu (2012). Green supply chains: Efforts and potential applications for the Turkish market, Resources, Conservation and Recycling, 58, 50-68.

[3] P. Arellanes and D. R. Lee. (2003). The determinants of adoption of sustainable agriculture technologies: Evidence from the hillsides of Honduras, Proceedings of the 25th International Conference of Agricultural Economists ( I AAE), Durban, South Africa.

[4] R. Baskaran, S. Managi and M. Bendig (2013). A public perspective on the adoption of microgeneration technologies in New Zealand: A multivariate probit approach, Energ y Policy, 58, 177-188. 186 Panmanee and Wiboonpongse

[5] M. Barratt (2004). Understanding the meaning of collaboration in the supply chain, Supply Chain Management: An International Journal, 9, 30-42.

[6] B. M. Beamon (1999). Designing the green supply chain, Logistics Inf ormation M anagement, 12, 332-342.

] [7 M. Cao and Q. Zhang (2011). Supply chain collaboration: Impact on collaborative advantage and firm performance, J ournal of O perations M anagement, 29, 163-180.

[8] C.-C. Chen, H.-S. Shih, H.-J. Shyur and K.-S. Wu (2012). A business strategy selection of green supply chain management via an analytic network process, Computers and M athematics with A pplications, 64, 2544-2557.

[9] S. Choo and P. L. Mokhtarian (2008). How do people respond to congestion mitigation policies? A multivariate probit model of the individual consideration of three travel-related strategy bundles. Trans portation, 35, 145-163.

[10] K. Condliffe, W. Kebuchi, C. Love and R. Ruparell (2008). K enya Cof f ee: A Cluster Analysis, Microeconomics of Competitiveness, Harvard Business School, Boston, MA.

[11] Department of Foreign Trade (2012). Import and Export Statistics of Cof fee Beans, Available at http://www.dft.go.th/Default.aspx?tabid=164 (In Thai)

[12] P. I. Devi, S. S. Solomon and M. G. Jayasree (2014). Green technologies for sustainable agriculture: Policy options towards farmer adoption, Indian Journal of Agricultural Economics, 6, 414-425.

[13] C. M. Dües, K. H. Tan and M. Lim (2013). Green as the new Lean: How to use Lean practices as a catalyst to greening your supply chain, Journal of Cleaner Production, 40, 93-100.

[14] T. K. Eltayeb, S. Zailani and T. Ramayah (2011). Green supply chain initiatives among certified companies in Malaysia and environmental sustainability: Investigating the outcomes, Resources Conservation and Recycling, 55, 495-506.

[15] K. Green, B. Morton and S. New (1996). Purchasing and environmental management: Interactions, policies and opportunities, Business Strategy and the Environment, 5, 188-197.

[16] W. H. Greene (2003). Econometric Analysis (5th ed.), Prentice-Hall, Upper Saddle River, NJ.

[17] A. A. Hervani, M. M. Helms and J. Sarkis (2005). Performance measurement for green supply chain management, Benchmarking: An International Journal, 12, 330-353. Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 187

[18] R. P. Kampstra, J. Ashayeri and J. L. Gattorna (2006). Realities of supply chain collaboration, The International Journal of Logistics Management, 17, 312-330.

[19] D. Knowler and B. Bradshaw (2007) Farmers’ adoption of conservation agriculture: A review and synthesis of recent research, F ood Policy, 32, 25-48.

[20] D. Läpple and T. V. Rensburg (2011). Adoption of organic farming: Are there differences between early and late adoption? Ecological Economics, 70, 1406-1414.

[21] D. Li and Y. Zhou (2006). Cluster competitiveness and strategy based on modified GEM model -- An analysis on Changsha engineering machinery cluster in Center China. Proceedings of 2006 International Conf erence on M anagement Science and Engineering, Lille, France.

[22] J. Mangan, C. Lalwani and T. Butcher (2008). Global Logistics and Su pply Chain M anagement, John Wiley & Sonsm, Chichester, UK.

[23] V. Manthou, M. Vlachopoulou and D. Folinas (2004). Virtual e-Chain (VeC) model for supply chain collaboration, International Journal of Production Economics, 87, 241-250.

[24] D. McFadden (1973). Conditional Logit Analysis of Qualitative Choice Behavior. Available at http://eml.berkeley.edu/reprints/mcfadden/zarembka.pdf

[25] C. P. Milioti, M. G. Karlaftis and E. Akkogiounoglou (2015). Traveler perceptions and airline choice: A multivariate probit approach, Journal of Air Transport M anagement, 49, 46-52.

[26] H. Min and W. P. Galle (2001). Green purchasing practices of US firms, International Journal of O perations & Production Management, 21, 1222-1238.

[27] N. Mzoughi (2011). Farmers adoption of integrated crop protection and organic farming: Do moral and social concerns matter? Ecological Economics, 70, 1536- 1545.

[28] M. Mwangi and S. Kariuki (2015). Factors determining adoption of new agricultural technology by smallholder farmers in developing countries, J ournal of Economics and Sustainable Develo pment, 6(5), 208-216.

[29] G. B. Nkamleu and A. A. Adesina (2000). Determinants of chemical input use in peri-urban lowland systems: Bivariate probit analysis in Cameroon, Agricultural S ystems, 6 3 , 111-121.

[30] Office of Agricultural Economics (2016). Agricultural Production Data, Available at http://www.oae.go.th/ewt_news.php?nid=13577 188 Panmanee and Wiboonpongse

[31] T. Padmore and H. Gibson (1998). Modelling systems of innovation: II. A framework for industrial cluster analysis in regions, Research Policy, 26, 625- 641.

[32] A. L. Patti (2006). Economic clusters and the supply chain: A case study, Su p pl y Chain Management: An International Journal, 11, 266-270.

[33] M. E. Porter (1998). Cluster and the new economics of competition, Harvard Business Review, 76(6), 77-90.

[34] M. E. Porter (2003). The economic performance of regions, Regional Studies, 37, 549-578.

[35] P. Rao and D. Holt (2005). Do green supply chains lead to competitiveness and economic performance? International Journal of O peration & Production M anagement, 25, 898-916.

[36] J. Saltiel, J. W. Bauder and S. Palakovich (1994). Adoption of sustainable agricultural practices: Diffusion, farm structure and profitability, Rural Sociolog y, 59, 333-349.

[37] J. Sarkis (2003). A strategic decision framework for green supply chain management, Journal of Cleaner Production, 11, 397-409.

[38] J.-B. Sheu, Y.-H. Chou and C.-C. Hu (2005). An integrated logistics operational model for green-supply chain management, Transportation Research Part E: Logistics and Transportation Review, 41, 287-313.

[39] C. Sheu, H. R. Yen and B. Chae (2006). Determinants of supplier-retailer collaboration: Evidence from an international study, International Journal of O perations and Production M anagement, 26, 24-49.

[40] T. M. Simatupang and R. Sridharan (2002). The collaborative supply chain, The International Journal of Logistics Management, 13(1), 15-30.

[41] T. M. Simatupang and R. Sridharan (2005). The collaboration index: A measure for supply chain collaboration, International Journal of Physical Distribution & Logistics M anagement, 35, 44-62.

[42] Ö. Sövell, G. Lindqvist and C. Ketels (2003). T he Cluster Initiative Greenbook, Ivory Tower, Stockholm, Sweden.

[43] S. K. Srivastava (2007). Green supply-chain management: A state-of-the-art literature review, International Journal of Management Reviews, 9, 53-80.

[44] T. P. Stank, S. B. Keller and P. J. Daugherty (2001). Supply chain collaboration and logistical service performance, Journal of Business Logistic, 22(1), 29-48. Determinants of Green Cluster Supply Chain Adoption and Practice of Arabica Coffee Growers in Pang Ma-O and Pamiang Areas 189

[45] S. Vachon and R. D. Klassen (2006). Extending green practices across the supply chain: The impact of upstream and downstream integration, International J ournal of O perations & Production M anagement, 26, 795-821.

[46] L. Villanueva, F. Maradiaga-Blandon, K. Goh, L. Gerwin and P. D. Broughton (2006). T he Nicaraguan Cof f ee Cluster: Histor y, Challenges and Recommendations for Improving Competitiveness. Microeconomics of Competitiveness Course, Harvard Business School, Boston, MA.

[47] S. V. Walton, R. B. Handfield and S. A. Melnyk (1998). The green supply chain: Integrating suppliers into environmental management processes, International Journal of Purchasing and Materials Management, 34(1), 2-11.

[48] H.-J. Wu and S. C. Dunn (1995). Environmentally responsible logistics systems, International Journal of Physical Distribution & Logistics Management, 25(2), 20-38.

[49] X. Xue, B. Huang and T. Xiao (2009). The study of inter-organizational collaboration by cluster supply chain, Proceedings of 2009 IEEE International Conf erence on Automation and Logistics, Shenyang, China.

[50] B. Yan and L. Wang (2009). Industrial upgrading based on global value chain and cluster supply chain integration: Case study of a toy industrial cluster in Guanyao, Proceedings of the International Symposium on Inf ormation Engineering and Electronic Commerce ( I EEC ’09 ), Ternopil, Ukraine.

[51] J. Ying and L.-J. Zhou (2012). Study on green supply chain management based on circular economy, Physics Procedia, 25, 1682-1688.

[52] Q. Zhu, J. Sarkis and K.-H. Lai (2007). Green supply chain management: Pressures, practices and performance within the Chinese automobile industry, Journal of Cleaner Production, 15, 1041-1052.

[53] G. A. Zsidisin and S. P. Siferd (2001). Environmental purchasing: A framework for theory development, European Journal of Purchasing & Supply Management, 7, 61-73.