Advancing the Credit Ecosystem: Machine Learning & Cash Flow Data in Consumer Underwriting
This empirical white paper assesses the impacts on model predictiveness and credit access of both adopting machine learning techniques and incorporating electronic bank account information (often called cash flow data) in consumer underwriting models.
The evaluation was conducted by building a series of models based on an anonymized dataset that combines data from one of the main three nationwide credit bureaus with bank account information compiled by a data aggregator, and then comparing the models’ predictions against actual credit performance on new accounts opened in 2018-2019.
Across all models built for the study, the machine model that combined credit bureau data with cash flow data was the most predictive overall and across all subgroups. It also had the highest approval rates overall and for most subgroups at most risk thresholds, while also producing relatively low false positive rates (approvals of consumers who went on to default). The machine learning model built using only credit bureau data generally ranked second on both predictiveness and access.
This empirical white paper expands on FinRegLab’s prior quantitative research on the use of cash flow data for credit underwriting in both consumer and small business markets and managing explainability and fairness concerns in connection with machine learning underwriting models.
This empirical white paper is produced and published by FinRegLab.