Effort-Centric Fairness in Lending Decisions
创建于 更新于
摘要
The study introduces an effort-centric fairness framework for credit scoring that measures the minimum weighted effort for rejected applicants to reach approval, distinguishing feature-independent recourse from causal downstream effects [page::6][page::7]. It derives tractable local surrogates for general differentiable classifiers and exact expressions for logistic regression, enabling in-processing training that embeds effort parity into model fitting and yields explicit pathways to approval [page::9][page::11]. Empirical analysis on a mortgage dataset (HMDA+Freddie Mac) reveals masked inequality: even when predictive parity holds, rejected female applicants face higher recourse burdens; feature-independent regularisation reduces the gap by over 50%, while causal regularisation can achieve reductions above 90% but with larger predictive and risk-return trade-offs [page::21][page::23][page::24][page::25]. Under feature-independent parity, portfolio risk remains broadly stable and RAROC stays positive; under causal parity, losses rise but revenue remains near baseline and RAROC stays positive, illustrating the trade-offs involved [page::23][page::25]. The framework also provides actionable, applicant-level pathways to improve credit outcomes and supports transparent governance and regulatory discussions, with potential by-products of improved predictive parity [page::21][page::26][page::27].
