Original title: Effort-Centric Fairness in Lending Decisions
Shiqi Fang, Zexun Chen, Jake Ansell · July 2026
Lending algorithms often judge fairness only at the exact moment of a decision. Because of this, they can miss whether rejected people face unequal burdens to get approved later. The researchers call this problem masked inequality. To fix it, they built an effortThe minimum cost, in time or money, a rejected applicant must spend changing their situation to qualify.-based framework. It measures effort as the lowest weighted cost of real-life changes a rejected applicant needs to cross the approval line. It looks at simple feature actions as well as shifts that ripple through a causal modelA map of how changing one factor, like income, can change others, like debt, rather than treating each factor in isolation.. It compares the average minimum effort between protected groups. Testing this on mortgage data, the team found rejected female applicants had to work harder for approval even when standard fairness tests passed. Using feature-independent training reduced this effort gap by more than 50%, while a causal method cut it by over 90%. This process also provides clear paths to approval while tracking financial risk.
Imagine two students are told they both failed a driving test. One only needs to redo a parallel park; the other must take months of night classes because the test is given in an unfamiliar city. A rule that only counts who passes on the first try treats them as equal, but the second student faces a much higher hidden cost to reach the same result. In this paper, the approval line is the driving test, the minimum cost of changes needed to cross it is what the authors measure, and standard fairness checks miss exactly this kind of hidden difference between groups.
Lenders and regulators usually check fairness by comparing approval outcomes, but that can hide a subtler injustice: rejected applicants from one group may have to work much harder to ever qualify. This framework measures that hidden burden and shows how to shrink it, while also exposing the cost to the lender's risk and return. For rejected borrowers, that is the difference between a dead end and a visible path back to credit.
The abstract reports results on mortgage data, so the size of the effort gap and the effects of regularisation may not generalise. Causal regularisation cuts the gap by over 90% but increases expected and unexpected losses and lowers RAROC, so it is not free. Exact expressions are derived for logistic regression; other classifiers rely on local approximations. The paper presents effort parity as a complement to predictive fairness, not a replacement.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Effort-Centric Fairness in Lending Decisions — Shiqi Fang, Zexun Chen, Jake Ansell
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.