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

The paper investigates whether providing substantially larger loans to small and medium enterprises (SMEs) can improve firm outcomes and how lender credit allocation choices affect aggregate productivity. In settings with asymmetric information, lenders often act as de facto monopolists for larger loans, making their allocation decisions pivotal for efficiency. The authors partnered with a major Egyptian MFI (ABA) to randomize access to much larger loans (about 4x prior size) versus moderately larger loans (about 2x), enabling identification of heterogeneous treatment effects. They collected comprehensive baseline data including psychometric and cognitive measures to predict which entrepreneurs benefit or are harmed by larger loans. The study addresses the hypothesis that average effects mask substantial heterogeneity and that standard allocation practices, focused on minimizing default and relying on past small-loan success, misallocate capital. It also explores the mechanism that over-optimism and excessive risk-taking drive negative returns for some entrepreneurs when offered large loans.

The study builds on work on misallocation and productivity (Hsieh and Klenow, 2009, 2014) and research on loan officer decision-making and incentives (Dobbie et al., forthcoming; Fisman et al., 2017; Cole et al., 2015; Rigol and Roth, 2021). It contributes to literature documenting heterogeneous returns to capital and credit: Hussam et al. (2022), Beaman et al. (2020), Banerjee et al. (2019), Meager (2020), and Crépon et al. (2024), and contrasts with microcredit expansions showing modest average impacts (Banerjee et al., 2015; Meager, 2019). It relates to alternative credit margins studies, including Breza and Kinnan (2021) on a credit contraction, Banerjee and Duflo (2014) on preferential credit to larger firms, and Bari et al. (2021) on in-kind loans. Methodologically, it leverages Chernozhukov et al. (2023) for honest ML-based heterogeneous treatment effect inference and highlights limits of standard applicant data to predict entrepreneurial treatment effects (McKenzie and Sansone, 2019). The interpretation draws on psychology of optimism/overconfidence (Carver et al., 2010; Peterson, 2000; Weinstein and Klein, 1996; Frese and Gielnik, 2014; Hilary et al., 2016).

Setting and partner: Collaboration with Alexandria Business Association (ABA), a large Egyptian microfinance institution. Individual-loan clients with at least three successful prior loans were nominated by loan officers for eligibility in a new program (Tamouh) targeting much larger loans. Experimental design: Randomized controlled trial. Eligible borrowers, screened by loan officers and approved by a central credit committee, were randomized within officer strata into: Treatment—offer of a loan typically about 4x the size of previous loan; Control—offer of a loan about 2x the size of previous loan (standard policy then was 1.5x, but 2x was used to maintain perceived fairness). Borrowers could choose loan terms (6–24 months). Partial default risk was guaranteed by a philanthropic partner (unknown to loan officers) to allay lender concerns. Randomization occurred in 31 batches over 13 months (2016–2017), with 1,004 borrowers nominated by 168 officers. Data: Three in-person surveys: baseline (pre-randomization), first follow-up (≈20 months post-disbursement), second follow-up (≈30 months). Administrative loan data include detailed repayment histories for five prior loans and up to 30 months post-experimental loan. Baseline collected: standard business/owner data (sector, revenues, expenses, profits, employment, demographics, etc.), cognitive tests (Raven's matrices, digit span), financial literacy, risk preferences, and 50 psychometric items (Likert 1–5). Outcomes: Primary—monthly profits, revenues, expenses, wage bill, productivity (revenue-based TFP proxy), household expenditures. Secondary—business survival, employees, assets, mental and physical health, repayment behavior (late fees, days late), borrowing at 30 months. Estimation of average effects: ANCOVA (pooled across two endlines) with baseline outcome, loan officer fixed effects (strata), survey round indicator; clustering at individual level. Heterogeneous treatment effects: Following Chernozhukov et al. (2023), sample-splitting (50/50) repeated 100 times; ML models trained on training set to predict control outcome B(Z) and individual treatment effects S(Z) using covariates Z. Algorithms: elastic net, neural net, random forest, gradient boosting; select best by prediction score; Bonferroni correction for multiple ML methods. Test for heterogeneity via Best Linear Predictor (interaction of treatment with S(Z)); then estimate Sorted Group Average Treatment Effects (GATES) by quartiles of predicted ITE on the testing set. To study spillover to other outcomes, implement Conditional GATES (CGATES): assign quartiles based on profits ITE and estimate treatment effects on other outcomes by these groups. P-values adjusted for multiple testing and sample splitting. Loan officer perceptions: For a subset (≈293 borrowers, elicited pre-randomization), officers rated repayment ability and revenue growth under large vs smaller loans; used to assess alignment with observed heterogeneous impacts. First-stage: Documented take-up, loan sizes, terms, outstanding debt dynamics using administrative data.

First stage and repayment:

• Take-up: 85% treatment vs 76% control. Loan size: treatment +10,768 EGP over control; loan term longer (19.7 vs 13.2 months). Monthly installment +314 EGP in treatment.
• Repayment: 100% eventual repayment; treatment had more repayment frictions—perfect on-time repayment 63% (treat) vs 76% (control); total days late +14.2; late fees within 24 months +193 EGP. Despite more late payments, increased late fee revenue likely offset opportunity cost; program expanded post-study, suggesting lender profitability. Average intent-to-treat (ITT) effects (pooled endlines):
• Profits: +1,294 EGP/month (~+9% of control mean 15,649), not statistically significant.
• Revenues: +5,312 EGP; Expenses: +4,958 EGP; Wage bill: +147 EGP; TFP: +0.04 SD; none statistically significant.
• Household expenditures: +446 EGP (~+7.9%); marginally significant before multiple-testing adjustment, not after. Heterogeneous treatment effects (using psychometric/cognitive data):
• Strong heterogeneity detected (heterogeneity coefficient ~0.85; p=0.002). Standard business/demographic data alone did not detect significant heterogeneity; including them with psychometrics worsened predictive performance.
• GATES on profits: Top quartile (“top-performers”) +8,611 EGP/month (+55% of control mean); Bottom quartile (“poor-performers”) −8,180 EGP/month (−52%); both significant at 5%. Conditional effects by profit-ITE groups (CGATES):
• Top group vs bottom group:
  • Revenues: +50,942 vs −43,058 EGP.
  • Expenses: +38,770 vs −30,943 EGP.
  • Wage bill: +2,657 vs −1,557 EGP.
  • TFP: +0.63 SD vs −0.48 SD.
  • Employees: +2.05 vs −2.15.
  • Business survival: bottom group −7 pp (closure risk rises); top group +5 pp (ns).
  • Household expenditure: +2,182 EGP (top) vs −440 EGP (bottom, ns).
  • Mental health: bottom −0.20 SD (ns), top +0.09 SD (ns). Allocation quality and misallocation evidence:
• Performance with small loans vs large: In control (2x loans), top-performers show smaller gains and lower endline TFP; with treatment (4x loans), they show largest gains—implying that targeting based on success with small loans would misallocate larger loans.
• Loan officer perceptions: Officers believed large loans increase default for 28% of poor-performers and by an additional 18 pp for top-performers, despite no evidence top-performers actually default more. Officers did expect larger revenue gains more often for top-performers (38% vs 24% poor-performers), suggesting potential for improved allocation if incentives shifted from default-minimization to value creation. Mechanism and traits:
• Psychometric correlates of doing well with big loans: higher cognitive scores (digit span, Raven’s), better financial literacy, more risk-averse (invest less in hypothetical risky asset), and lower agreement with optimism/impetuosity statements (e.g., “act first…”, “gut feeling”, “several solutions to any problem”, “flexible schedule”, “always ahead of time”, “work seven days”, “spend a lot of time planning”). Poor-performers display stronger optimism/overconfidence. Policy counterfactual magnitude:
• Targeting likely high performers would move average productivity effect from ~0 to ~0.5 SD and raise monthly profit treatment effect by about 46 percentage points; across 1,000 firms, aggregate monthly profits would increase by ~7 million EGP (~USD 400,000).

Findings show that the allocation of large enterprise loans critically determines impacts on firm profits, productivity, and employment. Average effects are modest because large positive effects for top-performers are offset by sizable losses for poor-performers. Psychometric and cognitive measures, not standard business and demographic data, predict who benefits. Existing lender practices—emphasizing minimizing default and extrapolating from small-loan success—likely misallocate large loans: top-performers appear less successful with small loans and are perceived as riskier by loan officers. Adjusting incentives to focus on portfolio profitability or revenue growth could better align officer choices with firm gains. A behavioral mechanism—heterogeneous optimism/overconfidence combined with decreasing returns to risk-taking—explains the reversal of fortune: over-optimists take excessive risks when given larger capital, reducing expected profits, while realists deploy larger loans prudently for expansion. These insights suggest that collecting entrepreneurial-psychometric data and modifying incentive structures can substantially improve capital allocation and aggregate outcomes.

The study demonstrates that large loans to SMEs can generate substantial gains for a subset of entrepreneurs while harming others, making lender allocation choices pivotal. Psychometric and cognitive profiling combined with honest ML methods can identify top-performers far better than standard loan application data. Current practices, driven by default-averse incentives and reliance on small-loan performance, are prone to misallocation. Policy and practice implications include:

• Integrating psychometric/cognitive assessments into underwriting for large loan upgrades, while safeguarding against gaming and discrimination.
• Realigning loan officer incentives toward portfolio profitability or value creation, not solely default rates.
• Considering targeted expansion strategies to amplify aggregate profit and productivity gains.
• Advancing belief elicitation and validated measures of optimism/overconfidence to pre-specify targeting rules. Future work should replicate across markets, test incentive redesigns experimentally, develop robust and non-manipulable assessments, and examine general equilibrium and spillover effects of large-loan expansions.

• External validity and context: Conducted with one MFI in Egypt with a partial guarantee; results may differ where default penalties or legal environments vary.
• Control group received 2x loans (vs status quo 1.5x), limiting direct comparison to prior policy and potentially attenuating differences.
• Selection into eligibility is endogenous (nominated clients with ≥3 prior loans), restricting generalizability to all SMEs.
• Psychometric measures may be manipulable if used for lending decisions; stability over time is imperfect (some responses changed between baseline and follow-ups).
• Heterogeneity mechanism: Decreasing returns to risk-taking not directly tested; evidence on optimism is correlational.
• Loan officer perception data covers about a third of the sample and was not incentivized; measurement error possible.
• TFP estimation uses a revenue-based residual with limited inputs and no baseline assets; subject to endogeneity and measurement concerns.
• Limited horizon (≈30 months) and few firm closures constrain assessment of longer-run dynamics and reallocation effects; spillovers not measured.
• Differences in loan term lengths imply combined changes in capital and repayment duration, complicating comparisons to fixed-term interventions.