AI-based consumer credit underwriting

Abstract

Consumer credit providers increasingly rely on algorithmic credit underwriting, deploying Artificial Intelligence (AI) tools such as Machine Learning (ML), to predict consumers' credit risk. Lenders’ turn to algorithmic underwriting was initially facilitated by credit data sharing regimes, originally established in the United States and which have now become prevalent in many other jurisdictions as well. A key player in these credit data sharing regimes is the credit bureau - a commercial entity that collects, processes and sells credit data and various services based on this data. ‘Credit scoring’ stands at the core of these services. Under a ‘commercial model’ of data sharing, private credit bureaus operate independent credit databases, develop scoring models and sell credit scores to lenders who weigh them in their underwriting processes. Under a ‘hybrid model’ of credit data sharing regime, all credit data is centrally held on a national, supervised and strictly secured database, to which authorized credit bureaus enjoy exclusive access. The credit bureaus develop scoring models under strict regulatory supervision of the central bank, and sell these scores to lenders. This chapter explores how the implementation of a hybrid data sharing model may impact the benefits and risks of AI-based consumer underwriting. The Israeli regime, which has recently implemented a national database, serves as a case study. The analysis of the hybrid model against the background of AI-based scoring practices demonstrates that a hybrid model may sustain most of the advantages of AI-based underwriting, while possibly reducing some of the risks it creates. This analysis could provide important insights for policy makers and legislatures from other jurisdictions who consider adopting a similar model. Specifically, a supervised database that includes financial data on allconsumers together with a mandatory data sharing requirement which applies equally to all reporting bodies, may increase the accuracy of credit scores, and turn them into a fair and efficient benchmark for predicting consumers’ creditworthiness. Moreover, the hybrid model may also increase data security and consumers’ privacy, while better detecting discriminatory inputs. At the same time, however, the hybrid model might contribute to the exclusion of financially weak populations from the credit market. This concern could be mitigated through public and private initiatives that facilitate lending to consumers with low credit scores.