Black-Box Credit Scoring and Data Sharing

Abstract

Should credit scoring algorithms be transparent or opaque? I study this question in a model where the lender uses data shared by borrowers for pricing and rationing credit, and is privately informed about the data-generating process, on which he tailors his algorithm. I show that revealing the algorithm's parameters makes it vulnerable to gaming in the form of strategic withholding of unfavorable information. Under opacity, data withholding emerges as a hedging strategy against the risk of credit rationing and averse price discrimination. This leads to equilibrium no-disclosure when the borrowers' hedging motives are strong, typically when the lender has greater bargaining power. Conversely, conclusive evidence is disclosed in more competitive markets where borrowers can secure better interest rates. The lender's optimal transparency regime maximizes the algorithm's efficacy in rationing credit (while forgoing price discrimination) and is socially inefficient.