Algorithmic Fairness in Credit Scoring
Against the backdrop of growing adoption of algorithmic decision-making, a team of researchers from the Financial Conduct Authority simulates the transition from logistic regression credit scoring models to ensemble machine learning models using credit file data for 800,000 UK borrowers. They find that machine learning credit models are more accurate and that machine learning models neither amplify nor eliminate bias where fairness criteria focus on overall accuracy and error rates for subgroups defined by race, gender, and other protect class characteristics.
Teresa Bono, Karen Croxson & Adam Giles, Oxford Rev. of Econ. Pol’y
September 2021