0 of 4
alternatives beat the tuned baseline by more than noise · three fairness fixes (reweighing, ExponentiatedGradient, group thresholds) and a tabular foundation model
README
Disparate impact with confidence intervals for five tracks, T0 to T4, on German Credit. Every interval crosses the dashed line at 0.8.
Disparate impact on German Credit, with intervals. T0 is the tuned baseline, T1 to T3 are reweighing, ExponentiatedGradient and group thresholds, T4 is a tabular foundation model. Every interval crosses the 0.8 line.original

Problem

Claim tested: that standard bias-mitigation methods reduce disparate impact in credit scoring.

Evaluation

0.7263
baseline disparate impact on German Credit; its interval crosses the 0.8 fairness line
README

On German Credit the baseline disparate impact was 0.7263 and every confidence interval spanned the 0.8 four-fifths line; the 200-row test block, with 62 women, was underpowered. On the Taiwan set the baseline was already 0.9767; reweighing and ExponentiatedGradient moved it by under 0.004, and group-specific thresholds made it worse. The README says it in four words.

What failed and what changed

  1. No intervention improved fairness

    Result
    A 1.6B tabular foundation model did not distinguishably beat a tuned GBDT either.
    Lesson
    Many published fairness improvements rest on samples too small to support them. Reporting the null result is the finding.

Stack

  • Fairlearn
  • XGBoost
  • scikit-learn
  • FastAPI
  • Streamlit

235 testsCI run · Oct 2026

Contact / Rourkela, Odisha, India

Write to me