Measuring Fairness in Credit Scoring

EasyChair Preprint 10481

Abstract

We propose a general methodology framework for eXplainable credit scoring to provide interpretability of each individual variable and measure fairness. Specifically, it is able to detect important variables and quantifies their individual impact on a firm's credit classification via the Shapley-Lorenz metric; and it quantifies the degree of discrimination, conditional on the endogenous effects generated by the variables, via the Kolmogorov-Smirnov test. In the experiment on a panel dataset of $119,857$ credit records for approximately $20,000$ small and medium-sized enterprises (SMEs) in four European countries and $21$ industry sectors for the period 2015 to 2020, we showcase the application of the eXplainable credit classification. We find that Leverage and P/L are the most important variables in credit scoring. In contrast there is marginal discrimination in terms of Country and Sector. The fairness tests show consistent results.

Keyphrases: Artificial Intelligence Credit Scoring, Fairness Test, Shapley-Lorenz

@booklet{EasyChair:10481,
  author    = {Ying Chen and Paolo Giudici and Kailiang Liu and Emanuela Raffinetti},
  title     = {Measuring Fairness in Credit Scoring},
  howpublished = {EasyChair Preprint 10481},
  year      = {EasyChair, 2023}}