Introduction

A central debate in corporate governance concerns the proper objective of the firm, weighing whether its sole responsibility is to maximise profits for shareholders or to consider broader shareholder welfare, including social impact. This debate has become more salient with the rise of impact investing and business initiatives to adopt environmental, social, and governance objectives. This research investigates whether firms face a genuine tradeoff between profits and the welfare of the people they serve. It focuses on the financial services sector, which is the largest within the impact investing space. The researchers utilize a machine learning approach to assess how targeting specific customers affects a double-bottom-line enterprise weighing profits against pro-social goals, such as reaching underserved populations or improving client welfare.

Data

The study relies on data from three randomized microcredit approval experiments conducted in South Africa, the Philippines, and Bosnia. These datasets include baseline information about borrowers alongside repayment data to assess performance. In the South Africa study, researchers partnered with a for-profit lender offering high-interest consumer credit. The Philippines study involved microloans for entrepreneurs, and the Bosnia experiment focused on extending credit to a poorer segment of the population typically deemed too risky. In each case, the experimental sample consisted of marginal applicants who were near the lender’s approval threshold. The primary outcomes measured are lender profits, derived from administrative repayment data, and borrower income, obtained through household and enterprise surveys.

Methodology

The methodology proceeds in three primary steps. First, the researchers estimate predicted loan profitability conditional on borrower characteristics using flexible machine learning models. They then calculate lender profits under a counterfactual policy of offering loans only to borrowers predicted to be in the top tercile of profitability. Second, they evaluate distributional impacts by comparing this profit-maximising policy against a balanced lending policy that ensures credit is distributed across baseline income quintiles. Third, they use an R-learner strategy to estimate individual treatment effects on borrower income, attempting to identify which individuals would experience the largest positive change in response to a loan.

Results

The analysis reveals that lenders in South Africa and the Philippines could have significantly increased their profit margins through algorithmic targeting. Specifically, the average loan profit margin would have increased by 7.4 percentage points in South Africa and 8.6 p.p. in the Philippines. However, in Bosnia, the data and algorithms were unable to reliably predict which marginal applicants would maximise lender profits. Pooled results across all three studies show an average 4.8 p.p. increase in profit margin.

Regarding social impact, the study finds a significant tradeoff in credit access. If lenders approved only the highest profitability clients, they would shift lending away from groups that often struggle to access credit, such as female, less-educated, and lower-income borrowers. The research also estimates the effect of profit-maximising policies on borrower income, finding increases of 18% and 47% for South Africa and the Philippines, respectively; however, these estimates are noisy, with standard errors of 31% and 37%, and do not reach conventional levels of statistical significance.

To gauge the magnitude of the tradeoff, the researchers examined a balanced lending policy requiring lenders to lend to the most profitable 33% of borrowers within each income quintile. This constraint would lower the profit gains of targeting by about half in South Africa (reducing the gain to 3.6 percentage points) and by 36% in the Philippines (dropping to 5.5 p.p.). On average, the balanced approach sees a drop from 4.8 p.p. to 2.5 p.p. in potential profit gains.

Discussion and conclusion

This work highlights the complex tradeoffs lenders face when incorporating machine learning into loan allocation. While these algorithms can significantly boost lender profits, they also change the composition of borrowers in ways that could exacerbate socioeconomic inequalities. The study suggests that while profit-maximising lending could widen gaps by characteristics like gender or income, restricted targeting strategies, such as the balanced lending policy, offer pathways to achieve profitability gains while lessening impacts on borrower diversity. However, these restricted strategies are less effective and thus costly to the lenders.

The findings indicate that predicting lender profits is easier than predicting borrower impact with current data. This suggests that the profit side of the frontier will likely advance faster than the impact side, potentially making it harder for lenders to credibly maintain a double bottom line. Consequently, regulators, impact investors, or mission-driven capital may need to actively enforce social targeting objectives as algorithmic targeting becomes more prevalent.