Improving Financial Access for Entrepreneurs in Developing Countries: Evidence from a Series of Experiments with Commercial Bank Loan Officers

Transcription

1 Improving Financial Access for Entrepreneurs in Developing Countries: Evidence from a Series of Experiments with Commercial Bank Loan Officers Shawn Cole Harvard Business School, Jameel Poverty Action Lab, and Innovations for Poverty Action Martin Kanz Harvard University Leora Klapper World Bank November 17, 2009 [Preliminary and Incomplete. Please do not cite or circulate without author s permission] ABSTRACT Access to finance is a critical ingredient for successful entrepreneurship. Yet, in many developing countries, weak institutions and poor financial infrastructure limit firms access to credit. In such environments, loan officers of banks play a key role in screening and monitoring loan applicants. We analyze lending decisions made by loan officers of a commercial lender in India. Working with a large publicly listed bank, we conduct two experiments using active loan officers and a database of historical loans to better understand determinants of access to finance. We first vary incentive schemes faced by loan officers to examine how highpowered incentives affect loan officers screening performance, effort and willingness to take risk. Second, we test whether the availability of soft information improves lending decisions.

2 I. INTRODUCTION Drawing new firms into the formal financial sector is particularly difficult in emerging markets, where weak institutional environments and limited historical credit bureau data make it difficult for lenders to gauge the risk profile of prospective clients. Especially in the market for small business and retail loans, emerging market lenders have to rely heavily on the judgment of loan officers. Loan officers themselves, however, may have limited experience, be subject to behavioral biases, have insufficient information or may not be appropriately incentivized to translate available information into a correct lending decision that optimizes risk and return. In this paper, we analyze the underwriting process of small-business loans in an emerging market, using data obtained in cooperation with two large commercial lenders in India. We make two substantial contributions: Through a set of randomized experiments using real loan officers and historical loan files, we first examine the effect of high-powered incentives on the quality of officers decisions. Second, we randomly vary the availability of soft information to measure the value and importance of soft information which is essential in small enterprise lending, but both expensive to collect and difficult to quantify. The objective of this exercise is twofold. First, we provide a rigorous test of theories of delegated monitoring and identify ways in which incentive schemes inside the firm can be designed to improve the efficiency with which available information is translated into a lending decision. Second, we identify which types of information are most essential to accurately assess the risk of a prospective borrower and suggest ways to streamline the underwriting process to make lending in markets with opaque information more commercially viable. In a set of pilot experiments, we find evidence that the availability of soft information leads loan officers to become more conservative in their assessment of certain components of the loan proposal. The introduction of a simple linear incentive scheme does not lead to a similar conservatism in rating loan files. However, we find no evidence that the availability of soft information leads to a statistically significant improvement in the ability to screen out bad loan files, suggesting much potential for streamlining the use and collection of soft information. The paper proceeds as follows. In the next section, we provide a brief description of small-enterprise lending in India, and motivate our work in the context of the existing literature. Section III describes the experiment and experimental design. Section IV describes the experimental results, and section V concludes.

3 II. MOTIVATION AND CONTEXT Our work builds on two separate literatures. The first analyzes the role of banks in alleviating information asymmetries between borrowers and suppliers of capital. The second, related, literature focuses on the role incentives inside the firm and ways to reduce frictions in delegated monitoring. Finally, we present what is to our knowledge the first randomized experiment identifying the effect of incentives and variations in available information on loan officer behavior. We discuss these contributions in turn. Information and Lending Developing countries have experienced substantial economic liberalization over the past several decades. This newfound freedom has spurred a wide variety of new businesses, even industries, for example in India and China, and greatly contributed to the wealth and welfare of these nations and the world. Yet, in many cases financial market development has not kept pace with the potential for growth, and many have argued that the conservative lending practices of financial intermediaries, particularly government-owned banks, have limited firms growth potential. (e.g., Banerjee and Duflo, 2008; and Demirguc-Kunt, Beck and Honohan 2008). Financial intermediaries ability to meet rising demand depends critically on their ability to screen and monitor borrowers. In particular, the absence of verifiable credit ratings for the majority of new enterprises, the loan officer and risk-management team play a critical role in screening decisions of the loan officer in the field. The difficulty of lending to small businesses in emerging markets has long been recognized. Much of this literature has focused on poor institutions and the difficult informational environment in emerging markets (see Kunt, Beck, and Honohan (2008) for a review of this literature). A small but growing literature has begun to recognize that incentive problems within the bank may limit lending institutions ability to issue credit. In the absence of reliable credit rating, financial intermediaries in emerging markets rely heavily on soft information. This is particularly true in the market for retail finance (personal loans, small business loans, car loans, education loans) where it is particularly difficult to verify a borrower s true financial position from standard documents. Financial intermediaries operating in the Indian retail lending market, for example, routinely supplement the documentation required by the country s central bank with additional soft information such as site visits to the home and office of a prospective applicant, interviews with business partners and references from employers or trade representatives in the area.

4 The importance of soft information represents a particularly important constraint in emerging markets for two reasons: first, as levels of post-secondary education are low, the relative price of highly skilled individuals able to make credit judgments may be particularly limited. Second, the cost of softinformation collection is roughly invariant to loan size 1, making it particularly uneconomical to collect extensive soft information for small loans. Interviews with loan officers and risk management executives suggest that decision makers often prefer these non-traditional sources of client information to the standard documents provided by the applicant. Moreover, stylized evidence from the recent downturn suggests that financial intermediaries with a relationship-based lending model that places greater emphasis on the use of soft information suffered smaller losses due to (i) a more carefully vetted client portfolio and (ii) an improved ability to detect repayment risks and offer customized solutions, such as a more flexible rescheduling of installments that takes account of the client s true financial position. While the reliance on non-traditional or soft information has allowed financial intermediaries with a more decentralized lending model to mitigate default risk, the advantage offered by an increased reliance on soft information usually comes at a significant cost. It is not a coincidence that traditional lenders still tend to shy away from small enterprise loans and the retail segment for the precise reason that effective risk management in this segment requires a very relationship intensive lending model, with very high costs of verification and collection of client information beyond standard financials. As a result, the business of retail banks in emerging markets is still heavily skewed away from small enterprise lending and in favor of SME loans where standard hard information is more reliable and screening costs are lower. In the academic literature, the distinction between hard and soft information has received attention in the context of financial access in developed financial markets (Petersen 2004). Stein (2002) and others note that the size of a bank is negatively correlated with its ability to collect and use soft information, for the simple reason that the collection and assessment of soft information requires local expertise and a decentralized organizational structure. Stein (2002) and Berger et al. (2005) develop and test a model that relates the organizational form of a bank to its lending business. In the theory and evidence, large, hierarchical firms provide relatively weaker incentives for front-line loan officers to collect soft information, which provides a signal about borrower quality, but cannot be credibly transmitted to the 1 Indeed, it may take more time to evaluate small enterprises with poor documentation and credit history relative to larger firms, which have audited accounts and an established financial track record.

5 loan officer s superiors. Larger banks are shown to lend less to informationally opaque firms, and rely less on relationship lending. Incentives and Lending The design of optimal repayment incentives is a central lending issue that has received some attention by theorists. Diamond (1984) and Fuentes (1996) take a close look at the optimal design of delegated monitoring contracts. Common to both approaches is the assumption of costly state verification: a bank that lends to an entrepreneur with poor collateral can improve its assessment of the project risk by paying an employee to monitor the client. Without delegated monitoring, the optimal contract is a debt contract with costly bankruptcy: that is, a contract that provides the lender with an expected return equal to the prevailing interest rate and deters default with non-pecuniary penalties equal to any shortfall from the principal in the entrepreneur s repayment. With delegated monitoring, on the other hand, the financial intermediary can attain better outcomes and reduce its risk through diversification, while extending its loan portfolio. While all studies highlight the importance of performance-based incentives when the screening decision is delegated to an agent, the screening decision is not modeled formally. Moreover, there is very little empirical evidence on the comparative performance of different incentive contracts. We are aware of only three studies that link loan officer incentive schemes to actual lending behavior. Hertzberg, Liberti, and Paravisini (2009) study how loan officer rotation induces loan officers to reveal bad information about borrowers, suggesting that communication problems exist even within a bank. Banerjee, Cole, and Duflo (2009) use data on all bank loans in India from , and find that loan officers are very conservative. Moreover, this conservativeness appears driven by the threat of punishment for loans that go bad. Finally, Agarwal and Wang (2009) take advantage of a natural experiment within a large commercial lender in the United States. Once the lender instituted bonuses for loan origination, the share of loan applications approved increased, while the quality of lending decreased. While banks in emerging markets, particularly those that privatized, are beginning to recognize the importance of new business formation, they have yet to develop effective credit-scoring models to identify and screen high-risk, high-return firms. Loan officers typically face incentive schemes that offer modest rewards for originating new loans, but include either penalties or severe constraints for lending to risky borrowers. Encouraging loan officers to take appropriate risks, without making reckless decisions, is necessary to increase financing to entrepreneurs. Yet because experimenting with new lending models can be very expensive, banks are reluctant to innovate.

6 Our work is closely related to the literature on incentives within the firms. Recent experiments, such as Bandiera et al. (2007), have shown that incentive structures can have substantial impact on productivity within the firm. However, there is little internal knowledge, and no academic evidence, on what sorts of incentive contracts are most effective at improving screening decisions, or aligning the behavior of loan officers with the financial intermediary s operational and strategic goals. Research Context In this project, we propose a series of experiments designed to test the ability of different incentive schemes to motivate loan officers to make prudent trade-offs between risk and return. Our study presents an empirical contribution to the literature on optimal incentive contracts. Moreover, it seeks to offer guidance to practitioners who seek to improve incentives for optimal screening that will allow them to expand their client portfolio. New firms in developing countries lack credit in part because lenders have little means of screening high risk applicants. Understanding which incentive contracts lead to an improvement of screening decisions could therefore significantly increase lending to firms with little or no verifiable credit history. Prior to the recent financial crisis, commercial banks in emerging markets, subject to government prudential norms and capital requirements, have often been criticized for playing it safe, demonstrating a reluctance to lend to start-ups or high-risk enterprises. In countries whose financial system is centered around banks, this can make it particularly difficult for new and growing entrepreneurs to access the credit they require. This problem can be particularly acute for businesses led by women or minorities. Lacking informal networks from which to raise finance, they may be even more dependent on bank financing, yet may face discrimination from loan officers unaccustomed to lending to women or minorities. To implement this study, we will collaborate with a large publicly listed bank in Mumbai, the financial capital of India. Members of our field staff will accompany loan officers on their daily client visits. They will videotape loan interviews with new applicants. The videos gathered at these field visits will then be matched with our partner organization s loan database and integrated in a series of computerized experiments. In these interactive experiments, which will be integrated into our partner organization s employee training program, loan officers will be asked to rate the creditworthiness of loan applicants based on the videos and a copy of the required pre-sanction information on each applicant, and ultimately make a

7 decision about whether they would lend to that borrower. The participants for this experiment will be loan officers drawn from the staff of our partner organization. Loan contracts can be complex. An important advantage of this setting is that we will observe perfectly, and indeed have complete control, over the information that the loan officer sees. Thus, to understand the importance of collateral in obtaining credit, we can vary the presence, and type of, collateral indicated in the loan application. Similarly, we can observe and control whether the loan application is supported by a co-signer, as well as what types of covenants are required on the loan. Such precision would not be possible in an observational study. Before being shown the borrower videos and accompanying documentation, each loan officer will be randomly assigned to one of several incentive schemes. This will be either a baseline contract, stipulating a fixed payment for participation in the experiment, or a more complex incentive scheme, offering monetary rewards in proportion to the repayment performance, as well as potential penalties for lending to firms that eventually default. Subsequently, each loan officer will be asked to make lending decisions for a pool of loan applicants. The officer will be given a choice to view answers to selected questions drawn from the interviews conducted in the field, subject to a time constraint or limit to the total number of questions. Based on this information the loan officer will then be asked to make a decision to grant or deny a loan to the client. Comparing the quality of screening decisions under different performance schemes will then allow us to identify which incentive contracts are best suited to improve the quality of delegated screening decisions. In a second set of interactive experiments (joint with Leora Klapper), we will match loan officers and clients based on gender, ethnicity and social background. An emerging empirical literature has documented the presence of network effects and group bias in credit and business transactions (See Ravina, 2008, for an example). From the empirical evidence, it is as yet unclear to what extent this serves merely to divert resources to low-return uses (Banerjee and Munshi, 2003) or to strengthen enforcement mechanisms (Karlan 2007). The experimental setting we propose allows us directly to identify the role of ethnic, social or gender effects of this kind. To the extent that they introduce distortions into loan officers lending decisions, we will be able to examine which contractual scheme is best suited to mitigate these effects.

8 III. EXPERIMENTS The experiments we propose address two key questions concerning the evaluation of loan proposals in a market with opaque information. First, we examine the importance of soft information, which is typically costly to collect, in arriving at a correct assessment of credit proposals. Second, we ask which type of monetary incentive schemes can improve screening performance. Both treatments are motivated by the aim of identifying ways to streamline the underwriting process of small enterprise loans to realize reductions in the fixed cost of lending. An essential first step in streamlining the underwriting process is to arrive at an accurate assessment of the marginal information content of each part of the loan proposal. The collection of borrower information accounts for a significant share of the fixed cost of lending. This is particularly true in markets where hard information is notoriously unreliable, forcing lenders to collect additional soft information to assess the merit of the proposal. While this type of relationship-specific information can be of significant value to the lender, its collection comes at a high cost which often leads financial intermediaries to shy away from informationally opaque markets, such as start-up and small enterprise lending. A second strategy to reduce the fixed cost of lending aims at improving the efficiency with which available information is translated into a lending decision. This approach is the focus of the monetary incentive treatments we turn to in the second part of our analysis. While economists have long recognized the problem of moral hazard of bank equity owners, who enjoy access to finance because of deposit insurance, the incentive problems within the firm have received much less attention. The recent financial crisis has brought great attention to the importance of getting incentives for bankers correct. Allegations of loan officers looting assets of the banking, by making loans they think unlikely to be repaid in order to collect bonuses, have highlighted the limitations of incentive systems. In this paper, we test the most common incentive schemes offered for loan officers: a regime that mimics a salary, in which loan officers face no marginal incentive for making loans or maintaining portfolio quality; and incentive schemes in which officers are rewarded for making good decisions.

9 V. EXPERIMENTAL DESIGN AND RANDOMIZATION To study loan officer decision-making, we use a simple computer-based experiment in which loan officers approve or reject loan applications. The experiment is carried out at the central training facility of a large Indian bank, and complements the bank s regular credit officer training courses. Participants are drawn from all levels of the bank s staff and have a minimum of one year of experience in sanctioning small ticket loans. While the decisions are hypothetical in the sense that the loans have already been made and their performance has been observed, the loan officers face incentive schemes that depend both on their decisions and the eventual outcomes of the loans. The software presents loan officers with the presanction information of small enterprise loans drawn randomly from a representative sample of 600 loan files taken from the client database of one of India s largest retail lenders. Each file contains the full presanction information available at the time of approval and has at least 12 months of repayment history. We follow standard conventions in classifying asset quality: files are classified as either good (never overdue), marginal (more than 0 but less than 60 days overdue) or bad (more than 60 days overdue) and the loan officer s performance is measured by comparing her decision to the historical performance of each application. FIGURE 1

10 The software allows the loan officer to select which parts of the loan proposal she wishes to review. Each loan file contains hard information (financial statement, credit history, required documentation) and soft information (site visit report home, site visit report business, trade reference check, internal assessment). The loan officer reviews this information and is then asked to assess the merit of the proposal along various criteria such as business risk, project risk, management risk and collateral. Finally, the loan officer decides whether to sanction or reject the proposal. Using this framework, we exogenously vary (i) the type and extent of information available to loan officers, and (ii) the incentive scheme offered to loan officers, and compare the screening performance under each treatment. A crucial feature of soft information is that it is costly to collect. We match this feature in the experiment in the following manner. Upon arrival, loan officers are endowed with 200 points. In the course of the study, they may choose to purchase particular pieces of soft information with points. Alternatively, at the end of the experiment, they may trade the points in for cash compensation (at Rs. 2 / point, this represents a non-trivial share of the loan officers compensation). To put the amount in context, a typical loan officer in an Indian bank earns approximately RS. 100/hour, so, from the loan officer s perspective, this represents real money. By randomly varying the price of soft information (including offering some or all the information free, and making some information impossible to obtain), we can evaluate both the value of soft information to improved lending decisions, and measure the extent to which loan officers optimally collect costly information. A basic insight from principal-agent models is that when the principal s actions are unobservable, incentives are necessary to generate effort. To that end, we will cross the information treatments with a series of financial incentive schemes. These incentive schemes will include a piece rate, whereby the loan officer is paid a bonus for each loan application he or she approves, as well as a range of incentive programs, in which good decisions are rewarded, and poor decisions penalized. Our first outcome measure is the quality of loan decisions made by loan officers. Because we know the ex-post performance of each loan, we can evaluate the profitability of each loan officer s total portfolio. 2 Our second measure of outcome is the level of effort expended, as measured by the points paid by participants to view soft information. Of course, as banks cannot make incentive payments based on the 2 We choose to use the actual outcome of loans, rather than an independent rating of the quality of the loan, as it is the most objective measure of outcome, and because it can be communicated clearly and unambiguously to loan officers participating in the experiment.

11 amount or quality of soft information collected, we intentionally do not consider incentive contracts that condition on the collection of soft information. Experimental Design The randomization involves two steps. First, we assign a random sequence of loan files drawn from the three asset quality bins to each loan officer. Second, we randomly assign experimental treatments in which we vary the type and amount of borrower information available to the loan officer and the incentive scheme under which the file is rated. The information treatments are designed to estimate the marginal information content of the different sections of a typical loan file. In delineating the subsections of the loan file, we are helped by the fact that all financial intermediaries are required to adhere to minimum information criteria and a basic application format by the Reserve Bank of India. While, in practice, the level of detail of the data collected differs quite significantly between banks, the common format allows us to classify the information content of a typical loan file into standardized sub-sections, which can be further grouped under the headings of soft and hard information. For the purpose of this experiment, we follow Petersen (2004) and define hard information as information that can be verified and passed on without loss of content. Examples include a prospective borrower s biographical data, credit history or audited financial statements. In contrast, we define soft information as information that requires significant subjective judgment on the part of the individual collecting it. In the context of the typical EM retail or small enterprise loan, this includes items such as a third party reference check, report on a visit to the client s home or business or direct conversation with the applicant. The loan files in our database contain all of this information, and will be supplemented by videotaped interviews with a subset of borrowers, allowing us to compare screening performance under a wide variety of restrictions to the information set. By varying the price and availability of each type of information for the each loan, we will be able to estimate the causal effect of each information item on the accuracy of underwriting decisions. In practical terms, this will allow us to (i) quantify the value of soft information and (ii) advise which parts of the loan proposal transmit valuable information and which elements are redundant and can be trimmed or eliminated to reduce the cost of lending. 3 3 To be clear, any given loan may be assigned to any of the information treatments below.

12 TABLE I: INFORMATION TREATMENTS Client and Co- Signer Profile Financial Statement Credit History CIBIL Report Site Visit Trade Ref Check Second Opinion Video Information I [Full Information] Information II [Hard Information] Information III Information IV Information V Information VI The incentive treatments, by contrast, are designed to identify ways to improve the efficiency with which information is translated into a lending decision holding the information environment constant. The incentive treatments we propose are informed by the current debate regarding loan officer incentives. In particular, we propose four treatments that mirror the incentive schemes currently used by lenders in the market we are examining: I. Salary: Loan officers will simply be remunerated on a per-hour basis, regardless of how many applications they process. The goal of this is to understand the baseline (intrinsic motivation) of participants. II. Piece rate: Loan officers will receive a bonus of Rs. 30 for each loan they originate, regardless of whether the loan performs III. Low Incentives: Loan officers will receive a bonus of Rs. 30 for each loan they originate, and no bonus for each loan that defaults IV. High Incentives: Loan officers will receive a bonus of Rs. 30 for each loan they originate, and a penalty of Rs. 60 for each loan they originate which defaults.

13 V. Non-linear Incentives: Loan officers will receive a bonus of Rs. 30 for each loan they originate. The penalty will increase in the number of loan defaults: 0-1 defaults: No penalty 2 Defaults: Rs. 60 penalty 3 Defaults: Rs. 120 penalty 4 Defaults: Rs. 240 penalty 5 or more Defaults: Rs. 500 penalty As in the case of the information experiments, we can compare the historical performance of the file to the loan decision of the loan officer to estimate the profitability of each officer s lending portfolio. In addition to objective screening performance, we have several ways of measuring the effect of monetary incentives on screening effort. In the full experiment, we can measure the amount of credit points spent to view soft information, the time spent reviewing each file and each sub-section of the file. In addition, we ask loan officers to provide a subjective rating of each file along several dimensions, such as business risk, financial risk, management risk and collateral. This allows us to estimate how different incentive schemes affect both objective screening performance and subjective risk assessment. VI. PRELIMINARY RESULTS Due to the timing of our partner organization s credit officer training programs, we have not yet been able to conduct the lending experiment at the bank s loan officer training facility. In this section, we report results from a preliminary pilot experiment, run with employees of the Center for Microfinance. Appendix A provides precise details on the experimental design and sample size. The first experiment looks at the effect of available information on screening performance. In the second experiment, we take the information set as given and examine how a simple monetary incentive (INR 25 for each correct rating) affects loan officers ability to distinguish good from bad applications. The outcome variables we consider are measures of (i) objective decision quality (a dummy variable that takes on a value of 1 for each good file that was sanctioned and each bad file that was rejected by the loan officer) and (ii) the loan officers subjective assessment of different aspects of the loan file such as business risk, management risk and financial risk of the proposal. In the full experiment we will also

14 collect extensive data on screening effort, including the time spent reviewing each proposal and subsection of the proposal. Information: We consider the simplest assessment of the value of hard versus soft information. Each subsection of the loan file was classified as either hard or soft information. The control group was assigned loan files with the complete information set, while the treatment group was restricted to view only soft information. Incentives: The incentive treatment consists of a simple linear incentive scheme with a Rs 25 bonus for each successful decision, defined as the sanctioning of a successful loan or the rejection of a loan that subsequently turned delinquent. We analyze the outcome of each experiment relative to the full information no incentive baseline using treatment regressions of the form: y T [1] i where is a constant intercept, T is a treatment dummy, i is a loan officer fixed effect and is a stochastic error term. We first consider the effect of each treatment on objective screening performance and then look at the effect on loan officers subjective ratings of the proposal. In the pilot experiments, we do not find an effect of either the incentive or information treatment on the share of successful decisions overall. This does not change when we distinguish between correctly accepted and correctly rejected loans (see Tables A1 and A2). With regard to the information treatment, this suggests that much of the information necessary to arrive at a correct judgment is already contained in the basic customer data and financials soft information, at least of the type used in the present exercise, seems to do very little to improve loan officers ability to screen out high risks. It should be noted that we did not use the full range of soft information in the pilot (e.g. we did not use videotaped borrower interviews). Nonetheless, if this result holds in a larger sample, it would suggest significant potential for streamlining the collection of non-standard borrower information such as site visit reports and trade reference checks. While so far we cannot link the availability of soft information to an improvement in objective screening performance, there is some evidence that loan officers subjective assessment of risk changes with the availability of additional information. Table A3 breaks down the ratings for each file into several subcategories, such as business risk, management risk and financial risk and even further, recording loan officers assessment of items such as the honesty and character of the management, business cycle risk

15 and quality of the management. Interestingly, the availability of more information seems to make loan officers more conservative in their assessment of these risks, particular along the dimension of business and financial risk. To study the effect of incentives, we run the regression identified above, with an indicator variable for whether the loan file was covered by an incentive scheme. Table A4 repeats the exercise for the incentive treatment and, interestingly, does not find an analogous effect of the linear incentive scheme on loan officers subjective assessment of credit risks. While the results indicate that much of the conservatism in subjective risk assessments documented in Table A3 is driven by a loan officer fixed effect, the discrepancy between subjective risk assessment and the quality of loan officer decisions points to an interesting puzzle that warrants further exploration. VII. Conclusion The experiments we propose present (to our knowledge) the first randomized evaluation of incentive schemes in a setting where the monitoring decision is delegated to an agent. The question of how financial intermediaries can devise contracts to minimize agency problems and default rates in this setting is a question of considerable theoretical and practical importance. The results will be of academic interest as they will improve our understanding of how the structure of an incentive contract and the availability of different types of information affect the quality of delegated monitoring decisions. At the same time, our research will provide practical guidance on how banks can reduce credit risk by offering reward schemes that improve the screening decisions of their loan officers in the field. At a minimum, our work will impact practice at the commercial bank with which we are working. We plan to disseminate the results widely, both within India and at international conferences, and are hopeful that this study may spur banks around the world to experiment with more effective incentive schemes.

16 APPENDIX: Description of Pilot Experiment In November, 2009, we carried out a pilot experiment using employees of the Center for Microfinance. The experiment involved 20 subjects (whom we call loan officers, although their primary occupation involves evaluating microfinance programs), in a custom-designed computer lab in Ahmadabad, India. Subjects were given an introductory training session on the software interface and were then asked to evaluate 20 files on the basis of each application s pre-sanction information for a total of 400 observations. The historical loan files used for the pilot comprised a subset of 40 historical loan applications with at least 12 months of repayment history. All loan files are from self-employed borrowers in the state of Gujarat and are for small ticket business loans with a total loan amount ranging from 120,000 to INR 1,000,000 (US$ 2,500 to US$ 22,000). For the purpose of the initial experiment, we restricted our attention to 20 files with a good repayment track record (never overdue) and 20 files with a bad record (more than 60 days past due). We randomized within subject, such that in each of the two sessions of the experiment, subjects were asked to rate loan files under two different information or incentive treatments, allocated in random order. The following incentive and information schemes were used: In the first of two sessions, half of the rated files contained full information, including soft information items such as a trade reference check and a site visit report. The remaining files were restricted, such that loan officers could view only hard information, such as the client profile, documentation and financials. The sequence of good and bad files as well as the sequence of basic information/full information ratings was randomized. In the second of the two sessions, loan officers always had access to the complete information for each loan file. However, in 50% of the cases (chosen in random order), the loan officer was offered a linear incentive scheme, rewarding a correct decision with Rs 25 for a maximum reward of Rs 125 per session. The reward amount was chosen to correspond to a typical loan officer s hourly wage.

17 References Agarwal, Sumit, and Faye Wang, Perverse Incentives at the Banks? Evidence from a Natural Experiment, working paper, Federal Reserve Bank of Chicago. Banerjee, Abhijit, Shawn Cole, and Esther Duflo (2009). Default and Punishment: Incentives and Lending Behavior in Indian Banks, manuscript, Harvard Business School. Banerjee, Abhijit, and Esther Duflo, Growth Theory Through the Lens of Development Economics, Elsevier Handbook of Macroeconomics. Berger, Allen, Nathan Miller, Mitchell Petersen, Raghuram Rajan, and Jeremy Stein, 2005, Does function follow organizational form? Evidence from the lending practices of large and small banks, Journal of Financial Economics 76(2): Demirguc-Kunt, Asli, Thorsten Beck, and Patrick Honohan, Finance for All? Policies and Pitfalls in Expanding Access, World Bank: Washington, DC. Diamond, Douglas Financial Intermediation and Delegated Monitoring Review of Economic Studies pp Fuentes, Gabriel The Use of Village Agents in Rural Credit Delivery The Jurnal of Development Studies 33 (2) pp Hertzberg, Andrew, Jose Liberti, and Daniel Paravasini, Information and Incentives Inside the Firm: Evidence from Loan Officer Rotation, forthcoming, Journal of Finance. Petersen, Mitchell (2004). Information Hard and Soft Nothwestern University, Kellog School of Management. mimeo. Ravina, Enrichetta (2008). Love & Loans. The Effect of Beauty and Personal Characteristics in Credit Markets, working paper, Columbia University. Stein, Jeremy (2002). Information Production and Capital Allocation: Decentralized versus Hierarchical Firms, Journal of Finance, LVII:

18 APPENDIX A SUMMARY STATISTICS: LOAN FILE SAMPLE, PILOT EXPERIMENT All Files Mean StDev Min Max Loan Amount (INR) Total Income (INR) Total Debt (INR) Years in Business Age Employees N 40 Good Mean StDev Min Max Loan Amount (INR) Total Income (INR) Total Debt (INR) Years in Business Age Employees N 20 Bad Mean StDev Min Max Loan Amount (INR) Total Income (INR) Total Debt (INR) Years in Business Age Employees N 20

19 Table A1 Information Treatments Dependent Variable Overall Rating Correct Decisions Coefficient [1] [2] [1] [2] Full Information (2.24) (1.17) (0.08) (0.08) Constant (1.37) (1.67) (0.06) (0.12) Loan Officer FE No Yes No Yes Observations R Dependent Variable Correct Aprovals Correct Rejections Coefficient [1] [2] [1] [2] Full Information (0.10) (0.09) (0.07) (0.05) Constant (0.09) (0.10) (0.07) (0.11) Loan Officer FE No Yes No Yes Observations R Robust Standard Errors in Parentheses, *** significant at the 1% level, ** significant at the 5% level, * significant at the 10% level

20 Table A2 Incentive Treatments Dependent Variable Overall Rating Correct Decisions Coefficient [1] [2] [1] [2] Full Information (2.06) (1.85) (0.07) (0.07) Constant *** 0.25 (1.67) (1.68) (0.06) (0.22) Loan Officer FE No Yes No Yes Observations R-Squared Dependent Variable Correct Aprovals Correct Rejections Coefficient [1] [2] [1] [2] Full Information (0.07) (0.06) (0.03) (0.05) Constant (0.06) (0.01) (0.05) (0.22) Loan Officer FE No Yes No Yes Observations R-Squared Robust Standard Errors in Parentheses, *** significant at the 1% level, ** significant at the 5% level, * significant at the 10% level

21 Table A3 Information Treatments Business Risk Management Risk Dependent Variable Competition Seasonality of Demand Dependence on Business Cycle Quality of Management Experience of Management Management Honesty and Character Coefficient [1] [2] [1] [2] [1] [2] [1] [2] [1] [2] [1] [2] Full Information -0.61*** *** -0.46*** *** (0.17) (0.13) (0.17) (0.20) (0.19) (0.17) (0.22) (0.16) (0.20) (0.20) (0.21) (0.16) Constant 2.51*** 3.04*** 2.09*** 2.47*** 2.74*** 3.13*** 2.6*** 3.21*** *** *** (0.11) (0.13) (0.12) (0.26) (0.16) (0.15) (0.15) (0.24) (0.10) (0.20) (0.13) (0.24) Loan Officer FE No Yes No Yes No Yes No Yes No Yes No Yes Observations R Financial Risk Collateral Dependent Variable Debt Burden Interest Coverage Ratio Strength of Financial Position Collateral Coefficient [1] [2] [1] [2] [1] [2] [1] [2] Full Information *** *** (0.16) (0.16) (0.15) (0.11) (0.19) (0.12) (0.13) (0.14) Constant 1.93*** 2.86*** 1.85*** 2.80*** 2.77*** 2.86** 0.71*** 1.47*** (0.11) (0.15) (0.10) (0.22) (0.13) (0.18) (0.07) (0.28) Loan Officer FE No Yes No Yes No Yes No Yes Observations R Robust Standard Errors in Parentheses, *** significant at the 1% level, ** significant at the 5% level, * significant at the 10% level

22 Table A4 Incentive Treatments Business Risk Management Risk Dependent Variable Competition Seasonality of Demand Dependence on Business Cycle Quality of Management Experience of Management Management Honesty and Character Coefficient [1] [2] [1] [2] [1] [2] [1] [2] [1] [2] [1] [2] Full Information (0.16) (0.13) (0.18) (0.16) (0.17) (0.18) (0.14) (0.13) (0.13) (0.11) (0.18) (0.18) Constant 1.9*** 3.25*** 2.06*** 3.25*** 2.28*** 3.00*** 2.05*** 2.75*** *** *** (0.13) (0.23) (0.13) (0.69) (0.17) (0.26) (0.16) (0.23) (0.16) (0.23) (0.14) (0.44) Loan Officer FE No Yes No Yes No Yes No Yes No Yes No Yes Observations R Financial Risk Collateral Dependent Variable Debt Burden Interest Coverage Ratio Strength of Financial Position Collateral Coefficient [1] [2] [1] [2] [1] [2] [1] [2] Full Information 0.36*** (0.16) (0.17) (0.18) (0.14) (0.15) (0.15) (0.16) (0.09) Constant 1.68*** 3.00*** 1.3*** 2.25*** 2.27*** *** 1.75 (0.11) (0.27) (0.12) (0.23) (0.14) (0.26) (0.10) (0.57) Loan Officer FE No Yes No Yes No Yes No Yes Observations R Robust Standard Errors in Parentheses, *** significant at the 1% level, ** significant at the 5% level, * significant at the 10% level.