What might explain the different outcomes of grants and loans? Where loans are made with the group-lending model, Fischer (2013) suggests that group members might pressure others in the group to make lower-risk investments that also bring lower returns. But many microlenders, including some of those in the studies summarised by Banerjee et al. (2015a), now make individual loans.
Field et al. (2013) suggest a different reason why borrowers may make investments that are sub-optimally risky. Most MFIs require borrowers to start repaying loans a week or two after the loan is dispersed, and the authors posit that this requirement for rapid repayment might limit the types of investment that entrepreneurs are willing to make. For example, they may look for investments that will quickly generate the additional revenue required to make loan payments, but that may not offer the highest return. Feld et al. test this hypothesis by offering a randomly selected group of borrowers the opportunity to delay the first payment for two months. Their data show that the grace period leads the entrepreneurs to make higher-return investments. They also show that these investments are higher-risk and lead to much higher default rates among borrowers. A structural model suggests that the contract with the grace period would not be profitable for the lender, even though the social returns to that contract exceed the social returns to the standard contract.
There are vast numbers of microenterprises in any urban or peri-urban area of lower-income countries, and it is safe to assume that these enterprises vary in their potential. If lenders were able to identify those with better investment opportunities, perhaps they could design more risk-tolerant products for that subset of borrowers. It is not immediately obvious how to distinguish such enterprises, however, particularly as few employ formal accounting or even keep any written records. In this context, Hussam et al. (2017) design a very clever experiment that allows them to ask whether peers have information about which businesses are most likely to succeed and, if so, whether we can extract that information in an unbiased manner. Tapping into peer networks, of course, has a long heritage in development entrepreneurship. The approach was integral to the group-lending model developed by Yunus with Grameen Bank (Yunus 1989). Hussam et al. (2017) carry out a project in Amravati, India that follows in this tradition. While they work with samples of subsistence businesses, their approach may have applications to more dynamic firms as well. The authors begin by dividing their sample into groups of five owners. They then ask each owner a series of questions about the characteristics of the other members of their group, including questions on education level, enterprise profits and, most importantly, how much profits would increase if they invested an additional US$100 in the business. The design incorporates random cash grants that allow the researchers to estimate the actual marginal returns to capital in the same sample of enterprises. Although we should not necessarily expect peers to be able to predict the marginal returns of other enterprises, Hussam et al. provide clear evidence that they are able to do so – enterprises ranked in the lower tercile of expected returns ex ante show no gain in profits after receiving the cash grant, while those in the top tercile of ex ante expected gains show an increase in monthly profits of more than 20% of the value of the grant. They also show that when peers know that their evaluations will affect who receives the grants, their reports are biased in favour of family and friends. Hussam et al. implement incentives for truthful reporting that, under certain conditions, do produce less biased and more accurate reports. Whether the conditions necessary for truthful reporting can be met in practice is unclear, but the work at least shows that peers have valuable information.
Hussam et al. then compare the prediction of peers to predictions based on ‘hard’ data from baseline surveys conducted with the entrepreneurs themselves.[1] They find that information from peers (‘soft’ data) has predictive power above and beyond any predictions that can be gleaned from the hard data. When comparing predictions from machine learning models with those of peers, they find that machine-learning predictions can also help to isolate the owners with higher marginal investment returns (those in the highest tercile of predicted returns realise monthly returns of 18% in the grants experiment). But the soft information of peers has predictive power even after controlling for the baseline survey responses. Peers have sustained personal interactions with those they are judging, and the findings suggest that this may be crucial to the ability of judges to add value
[1] Fafchamps and Woodruff (2017) and McKenzie and Sansone (2017) similarly compare predictions of business plan competition panels with results from regressions or machine-learning algorithms.
