Thesis MSc Finance
The Influence of Risk Behavior on Blacklisting
within the Field of Microfinance in Bolivia
Author: Harm Schievink
1Supervisors: F. Cecchi, S. S. H. Eriksen & B. W. Lensink
2June 2016
Abstract
This study evaluates the relationship between risk behavior and blacklisting within the field of microfinance. For this purpose a dataset collected from clients of a Bolivian microfinance organization over the time period October 2015-November 2015 has been analyzed. Different logistic regressions (credit-scoring models) are estimated by including several variable sets in order to test the robustness of the risk behavior variable. Three distinctive risk behavior variables are derived from a hypothetical risk game. These variables have never been included in a credit-scoring model before and therefore this method is completely new and innovative. The findings of this study help (micro) financial institutions in making better predictions on potential blacklisting of new clients by using this new type of indicator.
Key words: Microfinance, risk behavior, blacklisting, credit risk, risk game
1. Introduction
In the last decade microfinance institutions have grown rapidly. Summit Campaign (2012) states that between the years 1997 and 2010 the number of extremely poor families with a microloan increased from 7.6 million to 137.5 million (Banerjee et al., 2013). Moreover, private investors worldwide invested approximately 10 billion US dollars in micro financial projects in 2014 (Microfinance Market Outlook, 2016). Taking a closer look to Bolivia, one of the microfinance pioneer countries, microfinance is the biggest reason for the fast and steady economic growth in the last couple of years (Schipani, 2012).
It is assumed that micro financial institutions incur significantly low default rates of their clients (Khandker, 2005). However, this success does not yield for every micro financial institution. In reality several micro financial institutions went bankrupt lately due to liquidity issues (Sane and Thomas, 2011). Hence, it may be too optimistic to state that micro financing comes only with low default rates. Furthermore, Armendáriz de Aghion and Morduch (2000 and 2010) argue that microfinance shifts away from group lending, which is an important cause for low default rates in the past. Therefore, based on the shift towards more individual lending it may be valuable to investigate whether it is possible to develop new methods that help to indicate potential blacklisting.
Microfinance organizations already manage their risk exposure by using credit-scoring models to predict whether customers have a relatively high chance of loan defaults. Merton (1974) was the first researcher who introduced such a type of model to determine and minimize the credit- or default risk exposure of potential clients (Hull et al., 2004/05). When a borrower once defaults on his loan or faces loan repayment problems, he ends up ‘blacklisted’ (Opwaka and Wanyoike, 2012). Hence, all ‘blacklisters’ have a relatively high chance of not repaying in time or defaulting. Therefore, (micro) financial organizations preferably do not offer these blacklisted borrowers a loan or might require a higher compensation rate to overcome the additional risk (Opwaka and Wanyoike, 2012).
Several characteristics such as gender, age and loan duration are often included in credit-scoring models in order to predict blacklisting (Dinh and Kleimeier, 2007; Vogelgesang, 2003; Sharma and Zeller, 1997). However, this thesis investigates a completely new variable and its relation to blacklisting. This new risk behavior variable is derived from a
hypothetical risk game; interviewees may invest an amount of money in a lottery that can be
term in this study. Not only defaults and loan restructurings are included, but also borrowers who face repayment issues are incorporated as these borrowers possibly also face an above average chance on defaulting. When a new indicator for potential blacklisting can be derived, (micro) financial institutions can incorporate this in the screening process for new clients. Therefore, this study examines the possibility to obtain a risk behavior measure from a simple
risk game that predicts potential blacklisting, which can be used by financial institutions.
To answer this main research question, various logit models including different variable sets, are estimated. These models help to determine whether the risk behavior variables can be maintained and remain significant in order to prove the variables’ robustness. There is no clear answer yet for the relationship between risk behavior and blacklisting. On the one hand, it can be reasoned that more risk taking results in more investments (Dohmen et al., 2011), which lead to a higher income that might be used in order to repay a loan. Hence the default rate decreases and borrowers do not end up blacklisted. On the other hand, it can be argued that risk averse borrowers do not take excessive risks that might result in lower losses. Hence, these borrowers might be better able to repay their loans. Another reason why risk averse borrowers might be better repayers is that they do not wish to get penalized for not repaying in time or defaulting their loan (Block and Gerety, 1995).
Besides the innovative character of the risk behavior variables, no research regarding credit scoring has been done in the time period of the obtained data used in this study. The thesis’ academic relevance lies thus in using a new variable in order to predict blacklisting and providing new insights in credit-scoring by using data samples in the period October 2015 - November 2015. The business practice of this study lies in improving decision-making of micro financial institutions. From the results of this study is can be derived that the risk measures show evidence that risk loving borrowers have a higher probability of being blacklisted. Therefore, based on these measures, it can be concluded that blacklisting is indeed influenced by the risk behavior of a borrower within the microfinance field. Accordingly, the hypothetical risk game should be introduced in screening processes of potential clients. The outcome of the game can subsequently be used to decide on providing the loan or not.
2. Literature review
This section extensively discusses the most important literature for this thesis. Subsequently, the main research question and hypothesis are extracted from the debated literature. The first part of this section briefly introduces the credit-scoring model. Then, the section continues with describing the empirical research related to the main hypothesis and the control variables that are used in order to determine the robustness of the main hypothesis.
2.1. Credit scoring model
In 1974, Merton was the first researcher who introduced a model to determine credit risk. After this introduction, financial institutions started to use this model frequently to identify the credit risk exposure of their clients. In order to minimize their risk exposures in the optimal way, the so-called credit-scoring models were introduced (Hull et al., 2004/05). Van Gool et al. (2012) defines a credit-scoring model as “a model which combines collected historical data in a model that shows the chance of repayment by regressing several indicators of the credit provider, credit receiver and the loan in order to determine the probability of default and hence the credit risk.” Credit risk is the risk that the lender is exposed to, as there always is a chance the borrower does not repay his loan. Many research has been conducted on credit-scoring models that are used by banks and insurance companies. However, less research has been done on credit scoring models used within the field of microfinance. Nevertheless, some examples are present: Reinke (1998) studied credit-scoring models within microfinance in South Africa, Vogelgesang (2003) in Bolivia, Dinh and Kleimeier (2007) in Vietnam and Sharma and Zeller (1997) in Bangladesh. Likewise, Van Gool et al. (2012) conducted research on credit-scoring models in Europe; Bosnia.
higher risk of not repaying a loan (Freytag, 2008). Despite all the pitfalls described, Schreiner (2000) argues that credit scoring is a robust approach to predict default within the field of microfinance, confirmed by his results obtained from research done in Bolivia and Colombia. The conducted studies often focus on the same type of predicting variables such as gender, age, duration and number of dependents, which are already confirmed to have influence on blacklisting by many researchers (e.g. Dinh and Kleimeier, 2007; Vogelgesang, 2003; Sharma and Zeller, 1997). In this thesis, these variables are used as control variables in order to determine whether the new variable ‘risk behavior’ is robust. The main focus in this study lies in the question whether it is possible to obtain a risk measure from a simple hypothetical risk game that can be incorporated during the screening process of new clients to help indicate potential blacklisting. By answering the main question, the level of asymmetric information between lenders ([micro] financial institutions) and borrowers (micro entrepreneurs) can be reduced as the predictability of a ‘bad borrower’ increases (Luoto et al., 2007). To come up with a new method in order to predict blacklisting might be valuable, since recently several micro financial institutions went bankrupt due to liquidity issues (Sane and Thomas, 2011). In other words, relatively many clients did not repay their loans, which resulted in difficulties to sustain their daily operations. In addition, there is a shift going on from group lending and required collateral, which are important causes for low default rates in the past of microfinance (Armendáriz de Aghion and Morduch, 2000; Armendáriz de Aghion and Morduch, 2010). The shift towards individual lending may thus result in higher risk behaviors.
found by Lejuez et al. (2003). They found a significant correlation between real world risk behavior and a lab experiment. Binswanger (1981) also found no significant differences in his study, which compares the results from playing with real money and a hypothetical game. To support this statement even more, evidence is found by Camerer and Hogarth (1999). They studied 74 experiments and concluded that financial incentives hardly improve accurateness compared to hypothetical incentives. The outcomes of these studies are taken into account when designing the methodology for this study, on which is elaborated on section 3.2.
Since the relationship between risk behavior, derived from a risk game, and blacklisting has never been investigated before, no literature is available on this specific topic. It can be argued that borrowers who dare to take more risk on average do more investments (Dohmen et al., 2011). These investments might yield in high returns and hence more income that can be used in order to repay loans. From this point of view, a higher repayment rate for risk loving borrowers is expected. However, borrowers who tend to be risk averse may not want to take any excessive risks that can lead to (big) losses. Since these borrowers have a smaller chance to lose a lot of money, they have more money available to repay their loan. Moreover, Block and Gerety (1995) state that risk averse individuals do not want to get penalized for their actions. Translating that to this study: borrowers do not want to get penalized by means of for example fines. Consequently, they might push themselves to always repay their loans in time. This can influence the loan repayment rate positively and results in less blacklisted borrowers. Accordingly, no clear relationship can yet be given on this link between risk behavior and blacklisting. However, microfinance is often seen as the path out of poverty and hunger, which might make a loan more essential to survive. Hence, it is assumed that micro entrepreneurs act as risk averse individuals (Littlefield et al., 2003). Therefore, this thesis builds on the arguments for risk averseness and tries to find more evidence by studying and testing recent data received from a Bolivian micro financial institution. Therefore the overarching hypothesis is:
Hypothesis 1: A higher level of risk aversion comes with less chance on blacklisting
existence of external influences (economical or environmental) and the extent of incentivizing the borrower to repay his loan (by for example penalties).
Non-repayment of the loan is not the only reason why organizations decide to place a customer on the blacklist. The Credit Reference Bureau (CRB) in Kenya applies more criteria for listing a borrower as blacklisted such as the occurrence of fraud and falsification. In addition, misuse of the borrowed money, bankruptcies and fake declarations are also reasons for the Credit Reference Bureau to mark borrowers as blacklisted (Opwaka and Wanyoike, 2012). However, in this thesis only borrowers who face defaults, loan restructurings or repayment issues are considered as blacklisters.
To test the robustness of the risk behavior variables, several control variables are included in the credit-scoring models, which reflect personal characteristics of the borrower and characteristics of the loan. The subsequent paragraphs elaborate on these characteristics by discussing empirical studies and literature.
2.2. Control variables
In contrast to almost none existing empirical research done on the relationship between risk behavior and blacklisting, many studies have been conducted on the link between other variables and blacklisting. In this thesis, these variables are used as a control variable in order to test whether risk behavior remains significant when different variables sets are included in the credit-scoring model. The control variables are distinguished in individual characteristics and loan characteristics and are discussed accordingly.
2.2.1. Individual characteristics
and thus a weaker repayment history than their feminine family members. Vigano (1993) comes with the insight that women have a more entrepreneurial attitude than men and thus have a better feel of responsibility to repay their loans. Another explanation is that women on average have a smaller loan size in comparison to men, which makes it easier to repay the loan and therefore decreases the chance of a default (Godquin, 2004). From this evidence it is thus expected that women less default on their loans.
Another predicting variable that often shows up in conducted research is the share of young and old people in a household that depend the income of the household head (dependency ratio). When a household consists out of many young and old people, this influences whether a borrower is a good or bad repayer. According to Crook et al. (1992) the more children a household consist of, the higher the chance that one or more payments are not paid in time. Hamilton and Khan (2001) also studied the effect of the number of children on blacklisting and found almost the same results as Crook did. The main difference, however, is the risk of a household that has one child. In the case of Crook the risk with one child is lower than the risk of a household without children. The risk increases again when the number of children is more than two. Dinh and Kleimeier (2007) explain this relationship as households with more dependents face more costs and fees related to for example school, clothes and food. Moreover, both children and elderly do not participate in the income of the household. Consequently, there is less excess money to repay outstanding loans.
Dinh and Kleimeier (2007) studied whether educational level can predict the default rate. In general, it is expected that higher educated people are better repayers than low educated people. Every additional year of education results in smarter people due to a ‘training effect’. On the other hand, IQ level can be more seen as how smart an individual is by nature (‘being effect’). Since the IQ level is strongly correlated to education; higher education in general means a higher IQ level, it is assumed that people who were in school for more years and possess a higher IQ level are probably better repayers. An explanation is that these people can probably make (better) decisions by logically reasoning on whether they are able to repay a loan and to see the consequence of not repaying a loan (Reiter, 1980).
conclusion is that he or she gained more (in-depth) knowledge on financial matters. Therefore, it is expected that he or she is better able to repay the loan in time.
Moreover, Dinh and Kleimeier (2007) provide findings that show that higher educated people normally earn a more stable and higher salary. The total income, hence not only salary, makes it easier and more effective to repay a loan. Besides regular salary, other types of income can be distinguished such as remittances, sales and rents. Dinh and Kleimeier also confirm, visa versa, that people who earn a relatively low salary are worse repayers than people having a high income. This fact is supported by evidence found by Vogelgesang (2003) who argues that more wealth and non-business income lead to better repayment rates. He reasons logically that when income is too low or the repayment requirements are to high, it becomes very hard to meet the repayments on time. Furthermore, Hamilton and Khan (2001) also conducted research on the income variable, however no significant effect of income on blacklisting was found in their scoring model.
Expectedly, higher income borrowers have more excess money for saving. Dinh and Kleimeier (2007), Crook et al. (1992) and Vigano (1993) all found a negative coefficient of having a savings account, since have a saving accounts initially means that they have excess money to repay their loans. As Opwaka and Wanyoike (2012) argue: “savings contribute to economic freedom and security, support in avoiding large costs of short-term credit and stimulate investments”. Therefore, the higher the amount of money on their savings accounts logically means a better repayment rate.
Van Gool et al. (2012) name a loan used for farming as the safest loans, since more group control exists. However, working within the agricultural industry comes with a lot of uncertainty and high volatility in harvest and income due to external influences (Vigano, 1993). Since the weather conditions are always unreliable the harvest can be above average in years with the most favorable conditions. On the other side, the harvest can be very limited due to perpetual drought. Because of the unreliabilities related to weather conditions, the default risk will increase, and the chance of ending up on the blacklist increases. Therefore, it is expected that borrowers who work as a farmer in the agricultural sector face more problems with repaying their loans and therefore have higher default rates.
and Kleimeier (2007) found contrary results. In their research the people between 18 and 24 years old were characterized by the lowest default rates, even though Schreiner (2003) categorizes these people as most risk loving.
2.2.2 Loan characteristics
Besides individual variables that predict blacklisting, also loan characteristics that influence the predictability of blacklisting can be distinguished.
Firstly, the duration of a loan is one of the most predicting variables in order to determine blacklisting. Solvability in the future is expected to be more uncertain and thus leading to higher defaults for long-term loans. Moreover, when there is unsatisfactory capacity in the short term, long-term loans will be used to repay those short-term loans, which eventually boost the effect of a non-payment (Van Gool et al., 2012). Van Gool et al. (2012) found a positive significant coefficient, which supports the abovementioned statement. Besides, Dinh and Kleimeier (2007) also expected that the duration and default rate be positively correlated. Their results show that loans with a short duration of less than thirteen months have a default rate of 2.37%. This default rate slightly increases to 7.35% for loans with a duration between three and four years. Loans that have a duration of more than four years have a contradictory default rate of 2.16%. Dinh and Kleimeier (2007) do not provide any explanation for this contrasting outcome. From the explanation provided by Van Gool et al. (2012) and the outcomes of Dinh and Kleimeier (2007), it is expected that borrowers who issued a loan with a longer duration have more chance on blacklisting.
Every borrower has its own reasons for issuing a loan. Why a loan is issued and what the purpose of the loan is, affects the chance of defaulting the loan and hence ending up on the blacklist. Several loan purposes can be distinguished as using the money for farming, repaying other debt or housing. Dinh and Kleimeir (2007) examined the default rate of five additional loan purposes in their study: business, house, collateralized, general credit and credit card. They conclude that loans with a collateralized purpose have the highest default rate (11.96%) and credit card loans the lowest (0.40%). Credit card loans are the lowest, since only people with a low risk profile are allowed to receive a credit card. The high default rate of those collateralized loans is explained by the fact that the collateral is already a sign for relatively high credit risk (Dinh and Kleimeir, 2007). Additionally, it can be expected that micro entrepreneurs that receive a money injection and use the money for their business, are better repayers than micro entrepreneurs that use the money for other purposes related to self-interest, such as food, repaying other debts and farming. According to Hermes et al. (2011) the granted credit by micro finance institutions is used in order to lower poverty and setting up own enterprises that generate income for living and thus not for self-interest purposes such as food consumption.
However, when not taking the purpose of the loan into consideration, but focusing on loans with a collateral to secure the loan, Vigano (1993) argues that these collaterals help to improve the repayment rate. Moreover, Dinh and Kleimer (2007) distinguished four collateralized types of loans; real estate, fixed assets, mobile assets or no collateral at all. It can be concluded that loans with no collateral have the highest default rate (7.30%) followed by mobile assets (7.13%), fixed assets (2.86%) and real estate (0.41%). In addition, Vigano (1993) also states that the type of collateral is very important. When personal belongings are given as a collateral, it significantly increases the repayment rate. Therefore, to conclude, as argued by Vigano (1993), providing collateral in order to receive a loan makes the borrower aware of the fact that he should repay his loan.
Table 1. Expected influence of the (control) variable(s) variables on blacklisting.
Variable Exp. influence on blacklisting
Risk behavior + Female (gender) - Dependency ratio + Years education - IQ - Financial literacy -
Total income (in Bolivianos) - Total savings (in Bolivianos) -
Agriculture + Age - Duration + Numbers of loans - Good purpose - Collateral -
A plus or minus means respectively a positive or negative expected influence on blacklisting.
3. Data
This section elaborates on the data collection, the data structure, the risk game, the variable descriptions and concludes with providing the descriptive statistics of the (sub) sample(s).
3.1. Data collection and structure
In order to answer the research question, a dataset that is collected between October 2015 and November 2015 in Bolivia is used. The data is obtained by taking direct interviews with micro entrepreneurial families across different communities and regions in Bolivia. From these interviews both quantitative and qualitative data is gathered. The data is collected in collaboration with the Bolivian microfinance institution ‘Sembrar Sartawi’ and Institute for Advanced Development Studies (‘INESAD’). In total 1,994 interviews have been conducted whereby information about 8,334 individuals is collected. This is the case as the interviewee, in practically all cases the household head, was asked for information on his whole household (wife, children and other family members). All family members were questioned about specifics such as personal characteristics (e.g. gender, age, education), belongings (e.g. cattle, house, phones, cars), loan characteristics and whether they had contact with Sembrar Sartawi before. At the end of every interview a risk game was played.
needs further explanation before continuing. At the end of every interview, a so-called risk game is played with the participant. The risk game is executed in order to measure the participant’s risk behavior. More in-depth details on the execution of the risk game are presented in the next paragraph.
Initially, the survey collected information on 8,334 individuals (1,994 households). As this study focuses on household heads that have or had some kind of loan at the time of the survey, and were the person who provided the answers and played the risk game, this initial number decreases to 693 observations. The remaining data is controlled for outliers (see section 3.3.3 for the applied method) in the (control) variables and adjusted for data that drop out due to missing values. Therefore, the total number of observations that is used in the logit model equals 674.
3.2. Risk game
As briefly mentioned in the data collection and data structure paragraph, at the end of every survey a so-called risk game is played from which the different risk behavior variables are derived. This paragraph explains the risk game and the methodology in more detail.
Participants of the survey were informed beforehand about a reward of 20 Bolivianos (the Bolivian currency) that they would receive for sure and up to 40 Bolivianos as a sign of appreciation for collaboration throughout the survey. Hence, after completing the survey the participant receives the 20 Bolivianos and again it is mentioned that the he or she may keep the money as a sign of appreciation.
However, then the risk game is notified. This risk game is based on the Gneezy and Potters method that is based on expected value (Gneezy and Potters, 1997). The expected value is used in order to measure the level of risk aversion. Participants are asked to decide whether they want to keep the money, or to try to gain even more money by investing any amount of the received 20 Bolivianos in a lottery that can be seen as a risky investment. The game works as follows: the interviewer tosses a coin. Before the coin is tossed, the participant is asked to choose one of the sides (head or tails) as the winning side. Logically, the participant has a fifty-fifty chance of winning or losing. This risk game is an incentivized behavioral game, as participants have to act in a way that cannot be pretended or faked. Therefore the bias in the provided answers is diminished.
all of his received Bolivianos, chooses head and it turns out to be head, the participant receives 40 Bolivianos. However, when the participant invests again all of his 20 Bolivianos, chooses head and it appears to be tails, the participant receives only 10 Bolivianos in return and hence looses 10 Bolivianos.
The participant is completely free to choose the amount of money that he or she wants to invest. Hence, this can be zero (the risk game is not played) up till 20 (all the received money is invested). Thus, when none of the Bolivianos is invested, the participant makes sure that he earns 20 Bolivianos. The interviewer gives many examples with different invested amounts of money in order to make sure the participant understands the game completely. Before the coin is tossed, the interviewer informs whether the participant completely understands the game before continuing. When the participant understands the game, he is asked whether he would play one (or more) trial games, which are marked as the hypothetical games.
Table 2 presents all expected values per invested amount of Bolivianos. For example when 10 Bolivianos are invested, the outcome might be that one looses 5 Bolivianos, or win 10 extra Bolivianos. Thus the final outcomes are respectively 15 and 30, and hence the expected value equals 22.5 (average). It can be seen that the highest expected value can be reached when 20 Bolivianos are invested in the risk game. Gneezy and Potters (1997) argue that risk neutral or risk seeking individuals accordingly may invest all of their money in the game in order to obtain the highest expected value. On the other hand, risk averse individuals may invest less in the game.
Table 2. Expected value risk game
Investment (in Bolivianos) 0 2 4 6 8 10 12 14 16 18 20 Expected value (in Bolivianos) 20 20.5 21 21.5 22 22.5 23 23.5 24 24.5 25 The expected values per invested amounts in the risk game.
from now on, ‘risk game’ is meant as the hypothetical risk game that is played. For the completeness, the findings of the incentivized behavioral game: the real risk game, are added to appendices.
3.3. Variable description
This paragraph describes the dependent-, risk behavior- and control variables. Finally, the descriptive statistics of the whole sample and two distinctive subsamples are provided.
3.3.1. Blacklisting variable
The dependent variable in this study is blacklisting. Opwaka and Wanyoike (2012) define an excellent client as someone who repays his or her loan on time and is at the same time suitable to receive a new loan in the future again. By contrast, they define ‘blacklisted’ as borrowers that encounter problems in repaying their loans. The participants in the survey are first asked whether they have an outstanding debt at the moment of the survey. Subsequently, the household heads that confirm to have an outstanding loan are requested to answer this following question: Have you found it difficult to repay this debt? The survey presents the participants four different answer options, which are; not at all; yes, but I always met payments; yes, sometimes I was running behind schedule and yes, my credit is being restructured. Accordingly, a borrower is marked blacklisted in this study when his loan is restructured or when he was behind repayment schedule once or more than once.
3.3.2. Risk behavior variable
Table 3. Frequencies of invested Bolivianos in the hypothetical risk game and percentages of blacklisted and non-blacklisted borrowers per subsample
Invested amount Frequency Blacklisted Blacklisted (%) Non-blacklisted Non-blacklisted (%)
Risk behavior (a)
< 10 404 46 11.39 358 88.61 ≥ 10 270 57 21.11 213 78.89 Risk behavior (b) 0 39 3 7.69 36 92.31 2 210 22 10.48 188 89.52 4 66 11 16.67 55 83.33 6 65 6 9.23 59 90.77 8 24 4 16.67 20 83.33 10 205 48 23.41 157 76.59 12 3 0 0.00 3 100.00 14 1 0 0.00 1 100.00 16 2 0 0.00 2 100.00 18 5 0 0.00 5 100.00 20 54 9 16.67 45 83.33 Risk behavior (c) ≤ 6 380 42 11.05 338 88.95 > 6 294 61 20.75 233 79.25
The table shows the frequencies per invested amounts in the risk game. Furthermore, the number and percentages of blacklisted and non-blacklisted borrowers per invested amount are presented.
subsample. It can be seen that most of the not-blacklisters invested 2 or 10 Bolivianos. In addition, a notable percentage of the blacklisters invested 10 Bolivianos in the risk game.
Figure 1. Density per invested amount for the not-blacklisted and blacklisted subsample
Appendices C1-C3 present the findings for the outcome of the real risk game. The risk behavior are estimated in the same way as for the first trial risk game. The only difference is that the median for the real risk game is 0 instead of 6, hence risk behavior (c) is adjusted. What stands out in this real risk game is that there are hardly any differences in the risk behavior variables’ means, density and blacklist percentages between the blacklisted and not-blacklisted subsample. The results section elaborates more extensively on these findings.
3.3.3. Control variables
Following Vogelgesang (2003), several control variables are included in the logit model. The control variables are distinguished in personal characteristics of the borrower and the loan characteristics. Excess values within the control variables set, which fall below the 1 percentile or above the 99 percentile are appointed as outliers and therefore removed from the final dataset.
literacy training. On the other hand, a one means that the participant has received financial literacy training. Furthermore, the agriculture variable is also composted as a 0-1 dummy variable. A zero defines that the participant is not engaged in any agricultural activity. It follows that a one represents that the participant is engaged in any kind of agricultural activity. Likewise, the purpose variable is defined as a dummy. A one states that the loan is used in order to fulfill a good purpose, which equals a non-farming business purpose. A zero is given when the loan is used for bad purposes, which are farming, housing, food consumption or paying other debt. The last 0-1 dummy variable is collateral. When a collateral is required in order to receive a loan, a 1 is given. Accordingly, loans with no collateral are similar to a zero. The final variable set is shown in table 5. The first column shows the predictive characteristics, which are divided in borrowers characteristics and loan characteristics. Then, the different categories per variables are described in the second column. The last column defines the variable and briefly summarizes the expected outcome.
Table 4 shows an overview of the descriptive statistics3 of the sample. What stands out is that there is quite some variation in total income, total savings and duration, as is reflected by the standard deviation presented in the fourth column.
Table 4. Descriptive statistics of the data
Variable Obs. Mean Std. Dev. Min. Max.
Blacklisting 674 0.15 0.36 0 1
Risk behavior (a) 674 0.40 0.49 0 1
Risk behavior (b) 674 6.78 5.41 0 20 Risk behavior (c) 674 0.44 0.50 0 1 Gender 674 0.18 0.38 0 1 Dependents 674 0.37 0.24 0 1 Years of education 674 7.26 4.14 0 22 IQ 674 2.56 2.82 0 15 Financial literacy 674 0.17 0.37 0 1
Total income (in Bolivianos) 674 2,938.64 3,847.04 0 60,000 Total savings (in Bolivianos) 674 1,198.69 3,876.79 0 35,000
Agriculture 674 0.93 0.26 0 1 Age 674 43.14 11.60 18 81 Duration 670 123.71 87.91 4 672 Numbers of loans 674 1.18 0.55 0 4 Purpose 674 0.09 0.29 0 1 Collateral 674 0.86 0.35 0 1
The table presents the number of observations, mean, standard deviation, minimum- and maximum value for the dependent variable; blacklisting, independent variables; risk behavior (a), (b) and (c) and the control variables.
Table 5. Explanatory variables in more detail
The first column shows the predictive variables’ names that are used to create the credit-scoring model. The second column provides insight in how categories are established and finally the last column gives a brief description of all the variables on how the variable is defined and the expected influence on blacklisting.
Variables Categories Description
Borrower characteristics
Risk behavior (a) risk averse = 0, risk loving = 1 When 10 or more Bolivian (out of 20) are invested in the win game, the participant is labeled as risk loving. Participants who are risk averse are expected to be better repayers.
Risk behavior (b) 0 - 20 Amount invested in the win game. Participants that invest a higher amount are expected to be worse
repayers.
Risk behavior (c) risk averse = 0, risk loving = 1 Participants who invested more than the median (6 Bolivianos) are marked as risk loving. When more
than the median is invested, it is expected that the repayment rate declines.
Gender male = 0, female = 1 Gender of the interviewee. Females are expected to have a better repayment rate.
Dependents 0 - 1 Number of children under 16 years and elderly over 65 years old as percentage of total household
members. A higher dependency ratio is expected to have a worse repayment rate.
Years education 0 - 22 Number of years the participant receive education. It is expected that more years of education result in
better repayment rates.
IQ 0 - 19 The participants can answer a maximum of 19 questions. Every good answer is one point. Participants
who have more points are expected to be better repayers.
Financial literacy no = 0, yes = 1 Whether the participant ever received any financial literacy training. It is expected that more financial
literacy training result in less chance on blacklisting.
Total income 0 - 60,000 The income of the participant in Bolivianos for a regular month. Participants who earn a higher
income are expected to be better repayers
Total savings 0 - 35,000 The total amount of savings in Bolivianos a participant has. It is expected that more savings result in
less chance on blacklisting.
Agriculture no = 0, yes = 1 Whether the participant engages in any agricultural activities. It is expected that involvement in
agricultural activities lead to lower repayment rates.
Age 0 - 81 Age of the participant. Older people are expected to be better repayers.
Loan characteristics
Duration 0 - 672 The duration of the loan in weeks. Participants who issued loans with a longer duration are expected to
be worse repayers.
Number of loans 0, 1, 2, 3, 4 The number of loans the household head has. It is expected that more loans result in a better
repayment rate due to the learning effect.
Purpose bad = 0, good =1 Whether participant uses the loan for a good purpose or a bad purpose. Participants that use the loan
for a good purpose are expected to be better repayers.
Collateral no = 0, yes = 1 Whether the loan is collateralized, it is expected that the repayment rate is higher than for a
4. Methodology
This section elaborates on the conducted methodology used in this research. First, the model is described, then the section continues by explaining the selection procedure of variables. Then, the robustness of the risk behavior variables is elaborated. Finally, the reliability, validity and performance are specified in more detail.
4.1. Estimation model
Several credit-scoring models are built with the purpose of testing whether the risk behavior variable influences the chance on blacklisting and the robustness through the different models. As mentioned in the data section, blacklisting is defined as: ‘a household head that
sometimes was running behind repayment schedule once or more than once or a household head whose credit was restructured’. A logistic regression is set up in order to fulfill the
objective. This model is based on studies conducted by among others Schreiner (2000), Thomas (2000) and Dinh and Kleimeier (2007) who also created a logit model. The logistic regression is beneficial as it fits the best with the obtained data, since the determined dependent variable ‘blacklisted’ is dichotomous (blacklisted: yes or no). In addition, Ordinary Least Squares (OLS) regression gives blacklisting probability outcomes, which are smaller than zero in this study. This does not match reality, since the probabilities of blacklisting always lie between zero and hundred percent. A logit model overcomes this problem, and is therefore preferred. The logit model is estimated by formula (1) and (2).
Within this model, larger values for πi match with a higher probability of being blacklisted. Further, Dinh and Kleimeier (2007) argue that logitj is ex-ante unnoticeable and that therefore the probability of default only can be calculated as a dummy between zero and one. In order to overcome the heteroskedasticity problem and autocorrelation, all logit models showed in this study are estimated by using clustered robust standard errors4 (Thompson, 2011).
πi = E(Yi = 1) = 1
1+𝑒(−𝑙𝑜𝑔𝑖𝑡𝑗) (1)
logitj = β0 +β1xi1 + β2xi2 + ….. + βkxik (2)
Yi = binary dependent variable ‘blacklisted’ β0 = intercept
βi = regression coefficients
xi = (dummy coded) explanatory variable πi = probability of default
4.2. Selecting variables
In order to come up with the most suitable model, Hamilton and Khan (2001) argue that the number of predicting variables should be minimized to optimize the predictive power. Further, the variable set has to be comprehensive so that it is easy to use for people who want to work with the model. The maximum number of variables that is included in the model is fourteen, which in line with the upper boundary of twenty predicting variables given by Thomas et al. (2002) and Laitinen (1999) who argues that the number of explanatory variables should not exceed fifteen.
Different methods are available for selecting the predicting variables that can be used in the model. Hand and Henley (1997) examine three distinctive methods. The first one is based on common sense and expert knowledge. Another method is grounded on the forward step, backward step or a combination of both approaches. The final method uses a quota that shows the distinction amongst the defaulted and non-defaulted loans distribution on a particular variable. In this thesis the forward stepwise method is executed to estimate the credit-scoring model as has been executed by Dinh and Kleimeier (2007) and Vigano (1993). This method starts with zero variables in the model and adds gradually more variables (‘specific-to-general approach’) that might influence the dependent variable: blacklisting. Variables are left out from the ultimate variable set when they correlate with one or more other independent variables or do not contribute to the prediction (Dinh and Kleimeier, 2007).
4.3. Robustness of risk behavior variable
the general-to-specific approach where model one is the most general model and the remaining models are the more specific ones.
4.4. Reliability and validity
To create reliable models, no form of multicollinearity should be present. Therefore a cross-correlation matrix is estimated in which the cross-correlations between all the independent variables are shown. Variables that have a correlation exceeding 0.35 are marked as modest or high and are therefore not included in the model at the same time (Taylor, 1990). The cross-correlation matrix is added to the appendices, and can be found in appendix D.
In addition, several post estimation tests are conducted to determine whether a model is reliable and to test the fit of the estimates. In order to optimize the variable set and to obtain the ultimate model, Verstraeten and Van den Poel (2005) argue that the Receiver Operating Characteristic (ROC) curve should be estimated. The ROC curve represents the discriminatory power of the credit-scoring model. Van Gool et al. (2012) argue that the area under the ROC curve (AUC) should be between 0.70-0.90 in order to discriminate between blacklisters and non-blacklisters. The AUC value always lies between 0.50 and 1.0. When the AUC value is equal to 1, the model predicts all outcomes. It follows that the model does not predict anything when the AUC value equals 0.50.
Besides the ROC curve and the area underneath, it is also tested whether the model fits the data. This is done be executing a Hosmer-Lemeshow goodness of fit test. The null hypothesis in the test is that the model does not fit the data. Accordingly, the alternative hypothesis states that the model fits the data.
4.5. Performance measurement
There are four possible outcomes when the model predicts blacklisting. The first option is that the model predicts a non-blacklisted borrower as a non-blacklisted borrower (Xg), hence the model is right. The second possibility is that the model makes a wrong prediction as a non-blacklist borrower, however in reality it is a non-blacklist borrower (Yg). To continue, blacklisted can be predicted as blacklisted (Yb) and non-blacklisted can be predicted as blacklisted (Xb).
calculated. In order to obtain the optimal credit-scoring model, Yg en Xb should be as small as possible, since in this way more observations are correctly predicted. Important to note is that a trade-off exists between specificity and sensitivity. When one increases, the other decreases and visa versa (Baesens et al., 2003).
Furthermore, table 6 reports how the percentage correctly classified good loans (PCCgood), the percentage correctly classified bad loans (PCCbad) and the percentage total correctly classified loans (PCCtotal) are calculated. It follows that the higher these percentages are, the better the entire model is able to predict blacklisting.
Table 6. The model’s predictability Observation Prediction Non-default Default Non-default Xg Xb Default Yg Yb SENS Xg/(Xg+Yg) SPEC Yb /(Yb+Xb) PCCgood Xg/(Xg + Xb) PCCbad Yb/(Yg + Yb) PCCtotal (Xg + Yb)/( Xg + Xb + Yg + Yb)
The matrix shows four possible outcomes of predictions. Accordingly, sensitivity (SENS) and specificity (SPEC) can be calculated. Further, the percentages correctly classified good, bad and total loans can be derived.
5. Results
This section presents the main outcomes after analyzing the data. These outcomes are used to confirm or reject the proposed hypothesis in section two and to discuss whether the outcomes are in line with the literature. Moreover, the addition of several variable sets tests the robustness of the risk behavior variable. Finally explanations between the outcomes of the hypothetical - and incentivized behavioral game are discussed.
5.1. Risk behavior results
Bolivianos in the risk game have a higher chance on blacklisting. Thus it can be argued that, according to this variable, risk-loving people have a higher chance on blacklisting.
Appendix E2 reveals the outcomes of the logit models with the third risk behavior (b) variable. Again the coefficient is positive in a range between 0.0433-0.0531 and significant on a 1% significance level. Therefore, hypothesis 1 is again confirmed. Translating this to reality means that participants who invest more in the risk game; risk loving, have a higher chance on blacklisting than borrowers that invest a (very) small amount of Bolivianos in the game.
Finally, it can be derived from table 7 on page 27 that the risk behavior (c) coefficient is positive in a range of 0.764-0.832 and remains significant on a 1% significance level throughout the different logit models. Hence, the outcomes are in line with the findings of risk behavior (a) and risk behavior (b) and therefore hypothesis 1 is also accepted for this variable. This means that participants who invest more than the median have a higher chance of blacklisting than participants who do invest less than the median. Thus, according to this variable risk loving borrowers have a higher chance on blacklisting than risk averse borrowers.
Since the risk behavior (a), risk behavior (b) and risk behavior (c) variables are significant in all logit models, it can be concluded that these variables are highly robust and therefore indicators for potential blacklisting. The preferred indicator for potential blacklisting is risk behavior (c). Since this variable splits the entire sample in two at the median, which is the benchmark of the individuals’ preferences. Moreover, the median is a good reference point or benchmark in order to determine whether an individual shows excessive risk taking in comparison to peers (Fiegenbaum and Thomas, 1988).
The findings, shown in appendices F1-F3, derived from the real risk game show some very contradictive outcomes. None of the risk behavior variables are significant and robust, and hence they cannot be used in order to indicate potential blacklisting. Section 5.4 discusses these findings and comes with possible explanations for these differences.
5.2. Control variables results
Several coefficients of control variables are significant and thus it can be concluded that these variables also influence the chance of blacklisting. For example, the coefficients of the variable linked to gender are positive, which is in contradiction to the literature that women are better repayers than man (Dinh and Kleimeier, 2007; Schreiner, 2004; Godquin, 2004). A possible explanation might be that women in Bolivia are not seen as entrepreneurs but still as ‘the runner of the household’. Financial institutions therefore perceive women as more risky borrowers than man and consequently require a higher interest rate on their loans, which makes it more difficult to repay the loan (Alesina et al., 2013).
The coefficients related to financial literacy are negatively correlated to blacklisting in all logit models. This is in line with the literature that participants who receive financial literacy training, are better able to repay their loans.
Loan duration is according to the finding negatively correlated to blacklisting, which contradicts to literature stating that a long duration leads to higher default rates (Van Gool et al., 2012; Dinh and Kleimeier, 2007). This contradiction might be justified when the scheduled loan duration does not match to the size of the loan or to the wishes and requirements of the borrower (Godquin, 2004). A one-year loan can for example be far too large for a small Bolivian micro entrepreneur that actually needs way less money.
The positive coefficient linked to the number of loans of the household head variable is in contrast with the expectation that more loans leads to a smaller chance on blacklisting due to the learning effect. In this case, this learning effect seems not to be present. Dinh and Kleimeier (2007) state that keeping good track of borrowers’ credit record is very important, since borrowers who defaulted before now face difficulties in obtaining a new loan. The credit record by (micro) financial institutions in Bolivia might be unorganized, which makes it possible for borrowers who defaulted in the past to receive a new loan.
Table 7. Logistic regression outcomes of hypothetical game for risk behavior (c)
VARIABLES Blacklist (1) Blacklist (2) Blacklist (3) Blacklist (4) Blacklist (5) Risk behavior (c) 0.808*** 0.832*** 0.764*** 0.823*** 0.773*** (0.231) (0.229) (0.232) (0.228) (0.233) Gender 0.483* 0.473* 0.505** (0.266) (0.263) (0.234) Dependents 0.920* 0.671 0.773* (0.471) (0.459) (0.437) Years education -0.0196 -0.0341 -0.0221 (0.0276) (0.0265) (0.0271) IQ 0.0637 0.0443 0.0550 (0.0403) (0.0381) (0.0398) Financial literacy -0.960*** -0.859** -0.824** -0.859** (0.367) (0.353) (0.357) (0.357)
Total income 9.61e-06 -4.34e-06 5.18e-06
(2.06e-05) (2.07e-05) (1.88e-05)
Total savings -8.37e-07 -8.19e-06 -4.33e-06
(3.46e-05) (3.71e-05) (3.38e-05)
Agriculture -0.511 -0.542 -0.604 (0.403) (0.403) (0.395) Age 0.0126 0.00629 0.0110 (0.00991) (0.00940) (0.00962) Duration -0.00256* -0.00267* -0.00257* (0.00149) (0.00147) (0.00142) Number of loans 0.407** 0.401** 0.416** (0.201) (0.188) (0.185) Purpose 0.399 0.339 (0.364) (0.329) Collateral 0.612* 0.575 (0.366) (0.351) Constant -3.363*** -3.052*** -1.763** -2.228*** -2.313*** (0.904) (0.538) (0.713) (0.343) (0.700) Observations 670 670 674 670 674 AUC value 0.6871 0.6645 0.6290 0.6459 0.6603 SENS (%) 0.00 0.00 0.00 0.00 0.00 SPEC (%) 99.82 100.00 100.00 100.00 100.00 PCCgood (%) 0.00 - - - - PCCbad (%) 84.75 84.78 84.72 84.78 84.72 PCCtotal (%) 84.63 84.78 84.72 84.78 84.72 Hosmer-Lemeshow 0.9068 0.2086 0.7043 0.1313 0.2290 Pseudo R2 0.0697 0.0563 0.0325 0.0446 0.0498
Clustered robust standard errors in parentheses, clustered at the community level. *** p<0.01, ** p<0.05, * p<0.1
5.3. Validation, reliability and performance measurement
In order to check the validity, reliability and performance of the credit-scoring models, various tests have been executed. Table 7 and appendices E1 and E2 show all outcomes of these tests with the coefficients and standard errors of the variables beneath. The AUC values for all models are slightly below the range of 0.70-0.90 that Van Gool et al. (2012) recommend. However, the AUC values of this study lie between 0.6266-0.6834 and match reasonably with the value range that Van Gool et al. (2012) found in their own study. Overall, it can be concluded that the obtained AUC values do not satisfy the required value of 0.70 in order to discriminate enough between blacklisted and not blacklisted.
Furthermore, the Hosmer-Lemeshow goodness of fit statistic has been calculated. According to the outcomes, all null hypotheses, in table 7 and appendices E1 and E2, can be rejected. Consequently, it can be concluded that the credit-scoring models fit the data in all cases.
Finally, the SENS, SPEC and percentages correctly classified loans are calculated. The SPEC is for all models very high (>99%). However, the SENS is close to zero for all models, which means that blacklisted borrowers are hardly identified as blacklisted. The PCCtotal for all models fluctuates around 84 percent and exceeds the values found by Diallo (2006). Nonetheless, to determine the robustness of the risk behavior variables is the main objective of this study. Therefore, no major consequences should be derived from these performance findings.
5.4. Discussion of differences between hypothetical- and real risk game
From the logistics regressions, significant differences between the outcomes of the hypothetical- and the real risk game are notified. As (micro) financial institutions cannot provide ‘free money’ to every new client in order to play the risk game during the screening process, the outcomes of the hypothetical risk game are more valuable for this study. However, it might be very interesting to compare the significant differences and examine possible causes.
In order to test these possible explanations, an OLS regression is run. Since the majority (≈ 90 percent) played only one trial risk game, the amount invested in this first trial game is also included in the OLS regression, next to the number of trial games and the winning outcomes of these trial games. From the outcomes of this regression, which are added to appendix G can be seen that the coefficient (0.88) of the total number of trials is positive and significant at a 5% level. Hence the amount invested in the real game increases with 0.88 Bolivianos for every trial risk game played. Furthermore, the coefficient of the amount invested Bolivianos in the first trial game (0.51) is also positive and significant at a 1% significance level. This means that for every Boliviano invested in the first risk game, the amount invested in the real risk game increases with 0.51 Boliviano. To conclude, it does not matter whether the participant won or lost the trial risk game, as this variable is insignificant.
In addition, the correlation between the real amount invested and the amount invested in the first trial risk game is calculated at 0.5032. As this correlation is moderate (Taylor, 1990) and significant (P=0.0000), it can be concluded that the first trial risk game outcomes (‘lab experiment’) are an acceptable indicator for the outcomes of the real risk game. Therefore, the findings are in line with Lejuez et al. (2003), Camerer and Hogarth (1999) and Binswanger (1981) and contradict with the statement by Vroom et al. (2012) that a hypothetical game does not reflect reality.
6. Conclusion
This section reflects on the research question and presents the main findings. Furthermore, the findings are evaluated in the light of academic and managerial relevance. Finally, limitations are discussed and recommendations for further research are suggested.
6.1. Main findings
This study tries to determine whether risk behavior affects the chance of blacklisting within the field of microfinance. In order to answer this question, a dataset of 1,994 Bolivian households is analyzed over the time period October 2015 - November 2015 by means of estimating different logit models. The main findings are presented in this paragraph.
financial institutions should choose to play a hypothetical risk game, since it is unrealistic for these institutions to provide ‘free money’ in order to execute the risk game. When a participant invests more than the median of the invested amount, he has a significantly higher chance on blacklisting. As this variable holds in all the logit models, it can be argued that the variable is robust and therefore an acceptable indicator for potential blacklisting. This conclusion is even more supported, since significant evidence is also found for a relationship between the invested amount in the risk game on a continuous scale and blacklisting, and investing more than half of the reward and blacklisting. Likewise, these variables remain significant in all logit models, and therefore they can also be used as indicators for potential blacklisting.
The logit models are further tested for their validity, reliability and performance. The calculated areas under the ROC curve are rather in line with the literature, the models fit the data (goodness of fit) and the correctly classified observations show positive and applicable results for the models that contain all fourteen variables and all personal characteristic variables.
The findings contribute to existing literature in several ways. First, since this topic has never been studied before in the context regarding place and time-period. Secondly, and most importantly, using a hypothetical risk game as an indicator for potential blacklisting has never been examined before. Hence, it is completely new and innovative within the field of microfinance. Therefore, it provides some very interesting insights for decision-making in micro financial organizations in developing countries. If micro financial organizations are able to incorporate the hypothetical risk game in the screening process for new clients, without clients knowing the purpose of the game, the micro financial institutions can make better predictions on potential blacklisting. Since the outcome of the hypothetical risk game becomes the leading indicator in order to decide whether a loan will be granted, participants should not know the purpose to avoid that they will behave in a risk averse way. Eventually, the micro financial institutions can decrease their credit risk exposure as better predictions can be made and therefore their profitability might increase.
6.2. Limitations
estimate which participants receive a loan. Subsequently, the second step identifies which of these participants are blacklisted (Heckman, 1976). In this study, we work with a sample of households that took loans and hence did not consider the first step, which may bias the coefficients.
Moreover, the data concerning blacklisting might be biased. One can imagine that participants feel ashamed to inform the interviewer concerning repayment problems and defaulted loans in the past. Therefore, these participants might have given desired instead of true answers.
Finally, one might argue that the incentivized real risk game should be used as the independent variable in the model as this game may show the real behavior of the participants instead of pretended behavior.
6.3. Further research
As mentioned before, the conducted methodology is from an econometrical point of view not the correct one. Therefore, it is recommended to test the main research question again by using the Heckman two-step approach and see whether the same outcomes are found.
In addition, research should be conducted on how the risk game should be executed in practice, since participants should not know that the outcome of the game determines whether they will receive a loan or not. Otherwise participants may decide to invest a very small amount in order to increase the probability of receiving the loan.
Moreover, it would be highly interesting to conduct more research on possible (interaction) effects that might cause the insignificant outcomes of the incentivized risk game. And how these effects correlate to the outcomes of the hypothetical risk game.
Furthermore, another suggestion to point out for further research is to test whether the findings are applicable in countries outside of Bolivia. In first instance to check whether the same outcomes yield for other countries as well. Moreover, it might be interesting to establish a benchmark among other countries that are also actively involved in the field of microfinance.
References
Alesina, A. F., Lotti, F., & Mistrulli, P. E., 2013. Do women pay more for credit? Evidence from Italy. Journal of the European Economic Association 11(s1), 45-66.
Armendáriz de Aghion, B., & Morduch, J., 2000. Microfinance beyond group lending. Economics of transition 8(2), 401-420.
Armendariz de Aghion, B., Morduch, J., 2010. The Economics of microfinance. The MIT press, Cambridge.
Baesens, B., Van Gestel, T., Viaene, S., Stepanova, M., Suykens, J., & Vanthienen, J., 2003. Benchmarking state-of-the-art classification algorithms for credit scoring. Journal of the operational research society 54(6), 627-635.
Banerjee, A. V., Duflo, E., Glennerster, R., & Kinnan, C, 2013. The miracle of microfinance? Evidence from a randomized evaluation. Unpublished working paper. Massachusetts Institute of Technology (MIT), Cambridge.
Binswanger, H., 1981. Attitudes toward risk: theoretical implications of an experiment in rural India, Economic Journal 91, 869–890.
Block, M. K., & Gerety, V. E., 1995. Some experimental evidence on differences between student and prisoner reactions to monetary penalties and risk. The Journal of Legal Studies 123-138.
Camerer, C. F. and R. M. Hogarth, 1999. The Effects of Financial Incentives in Experiments: A Review and Capital-Labor-Production Framework, Journal of Risk and Uncertainty 19, 7– 42
Collins, M., Harvey, K., & Nigro, P., 2002. The Influence of Bureau Scores, Customized Scores and Judgmental Review on the Bank Underwriting Decision-Making Process. Journal of Real Estate Research, 24(2), 129-152.
Cubiles-De-La-Vega, M. D., Blanco-Oliver, A., Pino-Mejías, R., Lara-Rubio, J., 2013. Improving the management of microfinance institutions by using credit scoring models based on Statistical Learning techniques. Expert Systems with Applications 40(17), 6910-6917. Dennis, W. L., 1995. Fair Lending and Credit Scoring: There are many legal questions to consider before adopting a credit-scoring system. Lenders looking for a fair-lending safe harbor need to know the issue is nowhere near that simple. Mortgage Banking 56, 55-60. Diallo, B., 2006. Un modele de ‘credit scoring’pour une institution de micro-finance. Africaine: le cas de Nyesigiso au Mali.
Dinh, T. H. T., Kleimeier, S., 2007. A credit scoring model for Vietnam's retail banking market. International Review of Financial Analysis 16(5), 471-495.
Dohmen, T., Falk, A., Huffman, D., Sunde, U., Schupp, J., & Wagner, G. G., 2011. Individual risk attitudes: Measurement, determinants, and behavioral consequences. Journal of the European Economic Association 9(3), 522-550.
Fiegenbaum, A., & Thomas, H., 1988. Attitudes toward risk and the risk–return paradox: prospect theory explanations. Academy of Management Journal 31(1), 85-106.
Gächter, S., & Renner, E., 2010. The effects of (incentivized) belief elicitation in public goods experiments. Experimental Economics 13(3), 364-377.
Gneezy, U., Potters, J., 1997. An experiment on risk taking and evaluation periods. Quarterly Journal of Economics 112 (2), 631–645.
Feldman, R. J., 1997. Small business loans, small banks and big change in technology called credit scoring. The Region, 19-25.
Hamilton, R., Khan, M., 2001. Revolving credit card holders: Who are they and how can they be identified?. Service Industries Journal 21(3), 37-48.
Heckman, J. J., 1976. The common structure of statistical models of truncation, sample selection and limited dependent variables and a simple estimator for such models. In Annals of Economic and Social Measurement 5(4), 475-492.
Herger, N., Hug, C., Janett, U., 2016. Microfinance Market Outlook 2016: Developments, forecasts, trends. Responsibility Investments AG, Zurich.
Hermes, N., Lensink, R., & Meesters, A., 2011. Outreach and efficiency of microfinance institutions. World Development 39(6), 938-948.
Hull, J., Nelken, I., & White, A., 2004. Merton’s model, credit risk, and volatility skews. Journal of Credit Risk 1(1), 1-27.
Khandker, S. R., 2005. Microfinance and poverty: Evidence using panel data from Bangladesh. The World Bank Economic Review 19(2), 263-286.
Laitinen E., 1999. Predicting a corporate credit analyst’s risk estimate by logistic and linear models. International Review of Financial Analysis 8(2), 97–121.
Lalkhen, A. G., & McCluskey, A., 2008. Clinical tests: sensitivity and specificity.Continuing Education in Anaesthesia, Critical Care & Pain 8(6), 221-223.
Lejuez, C.W., Aklin, W.M., Zvolensky, M.J., Pedulla, C.M., 2003. Evaluation of the Balloon Analogue Risk Task (BART) as a predictor of adolescent real-world risk-taking behaviours. Journal of Adolescence 26 (4), 475–479.
Littlefield, E., Morduch, J., & Hashemi, S., 2003. Is microfinance an effective strategy to reach the Millennium Development Goals?. Focus Note 24(2003), 1-11.
Luoto, J., McIntosh, C., & Wydick, B., 2007. Credit information systems in less developed countries: A test with microfinance in Guatemala. Economic Development and Cultural Change 55(2), 313-334.
Mason, W., & Watts, D. J., 2010. Financial incentives and the performance of crowds. ACM SigKDD Explorations Newsletter 11(2), 100-108.
Microcredit Summit Campaign (2012). State of the campaign report. Technical report, Resuts Educational Fund.
Min, J. H., & Lee, Y. C., 2008. A practical approach to credit scoring. Expert Systems with Applications 35(4), 1762-1770.
Opwaka, R. P., & Wanyoike, D., 2014. An Assessment of the Attitudes of Blacklisted Bank Clients on Their Bank Use: A Case of Blacklisted National Bank Clients in Nakuru. International Journal of Science and Research 3(10), 2156-2160.
Reinke, J., 1998. How to lend like mad and make a profit: A micro-credit paradigm versus the start-up fund in South Africa. Journal of Development Studies 34(3), 44-61.
Reiter, R., 1980. A logic for default reasoning. Artificial intelligence 13(1), 81-132.
Sane, R., & Thomas, S., 2011. A policy response to the Indian micro-finance crisis. Available at Social Science Research Network: 1842926.
Sayinzoga, A., Bulte, E., Lensink, R., 2014. Financial Literacy And Financial Behavior: Experimental Evidence From Rural Rwanda.
Schreiner, M., 2000. Credit scoring for microfinance: Can it work? Journal of Microfinance/ESR Review 2(2), 105-118.
Schreiner, M., 2003. Scoring: the next breakthrough in microcredit. Consultative group to assist the poorest (CGAP). Occasional Paper 7, 1-127.
Schreiner, M., 2004. Benefits and pitfalls of statistical credit scoring for microfinance. Savings and Development, 63-86.
Sharma, M., Zeller, M., 1997. Repayment performance in group-based credit programs in Bangladesh: An empirical analysis. World Development 25(10), 1731-1742.
Taylor, R., 1990. Interpretation of the correlation coefficient: a basic review. Journal of diagnostic medical sonography 6(1), 35-39.
Thomas, L. C., 2000. A survey of credit and behavioral scoring: Forecasting financial risk of lending to consumers. International Journal of Forecasting 16, 149−172.
Thomas, L. C., Edelman, D. B., & Crook, J. N., 2002. Credit scoring and its applications. Siam, Philadelphia.
Thompson, S. B., 2011. Simple formulas for standard errors that cluster by both firm and time. Journal of Financial Economics 99(1), 1-10.
Van Gool, J., Verbeke, W., Sercu, P., Baesens, B., 2012. Credit scoring for microfinance: is it worth it?. International Journal of Finance & Economics 17(2), 103-123.
Verstraeten, G., Van den Poel, D., 2005. The impact of sample bias on consumer credit scoring performance and profitability. Journal of the operational research society 56(8), 981-992.
Vogelgesang, U., 2003. Microfinance in times of crisis: The effects of competition, rising indebtedness and economic crisis on repayment behavior. World Development 31(12), 2085-2114.
