Big Loans to Small Businesses: Predicting Winners and Losers in an Entrepreneurial Lending Experiment

This paper investigates the impact of larger enterprise loans on small businesses in Egypt, focusing on the efficiency of lender credit allocation decisions. The research is motivated by the prevalence of asymmetric information in credit markets, where a specific lender often holds a de facto monopoly, influencing capital allocation and aggregate productivity. Inefficient credit allocation can lead to high default rates, unprofitable loans for borrowers, and suboptimal loan sizes. Existing research largely relies on aggregate data, lacking detailed insights into individual lender decisions. The study addresses this gap by collaborating with the Alexandria Business Association (ABA), a major Egyptian lender, to conduct a randomized controlled trial. The experiment overcomes three key challenges: (1) It offers large loans even to firms not typically eligible, extending beyond ABA's normal practice; (2) It collects novel psychometric data to measure heterogeneous treatment effects and employs machine learning to avoid overfitting; and (3) It employs a credible identification strategy, generating sufficient loans to firms who wouldn't typically receive them and determining which firms would have received loans under business-as-usual conditions. ABA selected a sample of firms believed to benefit from larger loans. Within this sample, treatment borrowers received loans four times their previous size, while controls received twice the size of their previous loan. The flexibility offered in loan size and duration aimed to reflect increased credit access. The study also examines two allocation theories: (1) Lenders favor firms demonstrating success with smaller loans; and (2) ABA heavily relies on loan officer opinions, strongly incentivized to avoid defaults. The authors collected baseline psychometric data, to assess entrepreneurial characteristics relevant to success and used this data with machine learning to identify group average treatment effects.

The paper builds upon a growing literature examining heterogeneous firm-level returns to capital. Studies like Hussam et al. (2022) show individuals can identify peers with high returns, and Beaman et al. (2020) find that borrowing farmers experience higher returns than non-borrowers. Banerjee et al. (2019) demonstrates higher returns for experienced entrepreneurs, while Meager (2020) highlights both positive and negative returns to credit. Crépon et al. (2024) demonstrates substantial heterogeneity in returns to cash drops and loans. McKenzie and Sansone (2019) find that neither experts nor machine learning can predict future entrepreneurial success using standard data alone. This study contrasts with previous research by examining a different margin of credit expansion (larger loans) and focusing on heterogeneity rather than average effects. It is also related to studies examining alternative margins of credit alteration in developing countries, such as Breza and Kinnan (2021) which studied the impact of a microcredit crisis in Andhra Pradesh and Banerjee and Duflo (2014) that studied an Indian program expanding preferential credit access to larger firms. Bari et al. (2021) examine the provision of in-kind loans four times larger than previous loans, finding positive average effects but not focusing on allocation decisions.

The study partnered with the Alexandria Business Association (ABA), a large Egyptian microfinance institution, to conduct a randomized controlled trial. ABA's loan officers nominated clients they deemed suitable for larger loans. Approved applicants (who had successfully repaid at least three prior loans) were randomized into two groups: a treatment group receiving loans four times their previous size and a control group receiving loans twice their previous size. The experiment was stratified by loan officer. Three rounds of in-person surveys were conducted: baseline (before randomization), first follow-up (20 months post-disbursement), and second follow-up (30 months post-disbursement). Data collected included standard business and borrower characteristics, cognitive data (Raven’s matrices, digit span recall, financial literacy), risk preference data, and psychometric data (50 statements designed to capture personality traits). Administrative data from ABA tracked loan performance. Loan officer surveys collected their expectations regarding business and loan performance for borrowers under treatment and control conditions. To assess heterogeneity, the authors used a machine-learning approach from Chernozhukov et al. (2023). The sample was split into training and testing sets, using the training set to train machine learning models to predict individual treatment effects. The testing set was then used to evaluate the predictive performance and estimate heterogeneous treatment effects using sorted group average treatment effects (GATES). To explore how heterogeneous effects on profits translate to other outcomes, the authors extended the GATES procedure to estimate conditional sorted group average treatment effects (CGATES), conditioning on the individual treatment effect for profits. They employed four default machine-learning algorithms (elastic net, neural net, random forest, and gradient boosting), selecting the best-performing algorithm. They implemented 100 data splits into training and testing sets to account for uncertainty. The study used a Bonferroni correction to adjust p-values for the four algorithms.

The study reveals substantial heterogeneity in treatment effects, despite largely null average impacts. Top-performers (top quartile of predicted treatment effects for profits) increased profits by 55% of the control group mean, while poor-performers (bottom quartile) experienced a 52% reduction. Similar patterns were observed for wage bills, productivity, and household expenditures. Targeted lending toward top-performers would significantly improve average treatment effects for productivity and profits (estimated increase in aggregate monthly profit of about 7 million EGP). However, top-performers are less likely to receive large loans. Analysis shows they exhibit smaller profit increases (or larger decreases) in the control group, suggesting they may be overlooked as successful firms by loan officers. Loan officer beliefs strongly bias against lending to top-performers, as they perceived a higher default risk for top-performers compared to poor-performers. This is attributed to loan officer incentives focused on minimizing defaults rather than maximizing overall returns. Analysis using only standard data failed to reveal statistically significant heterogeneity, highlighting the importance of psychometric and cognitive data in predicting treatment effects. This suggests that entrepreneur characteristics, rather than firm characteristics, are crucial in determining the impact of increased credit access. The study found that poor-performers were more likely to agree with statements suggesting over-optimism and a gung-ho attitude toward risk, implying that over-optimism combined with decreasing returns to risk-taking could explain the negative impacts observed for some borrowers.

The findings highlight the significant role of credit allocation in determining the impact of credit expansion. The substantial heterogeneity in treatment effects and the observation that existing practices lead to misallocation underscore the importance of moving beyond average treatment effects when evaluating credit interventions. The study demonstrates the potential of psychometric data and machine learning techniques to improve the allocation of credit to small businesses. The reversal of fortune observed – where those who perform well with smaller loans perform poorly with larger loans – is consistent with a model incorporating heterogeneous optimism levels and decreasing returns to risk-taking. This suggests that some entrepreneurs may be overly optimistic and take on excessive risk with larger loans, leading to negative outcomes. The results suggest that a change in loan officer incentives to emphasize firm revenues might lead to better capital allocation.

The study demonstrates that the allocation of credit within a lender matters significantly for firm profitability and productivity and that current practices lead to misallocation. Psychometric data show considerable promise in predicting heterogeneous treatment effects, but caution is needed regarding potential gaming and discrimination. Future research should focus on strategy-proof methods for identifying suitable candidates for large loans and investigating how to incorporate beliefs into the allocation process. The study highlights the importance of understanding the potential for over-optimism and its impact on risk-taking behavior among entrepreneurs.

The study's findings may not fully generalize to contexts with different penalty structures for default. The study acknowledges potential spillover effects on other firms not captured in the analysis. The discussion of optimism is ex-post, and replication with pre-specified tests based on validated measures of optimism is recommended. The study's relatively limited timeframe might not capture the full long-term consequences of the credit expansion. Finally, the sample is limited to existing ABA borrowers and the credit committee review may bias the result toward higher quality entrepreneurs.