Fair credit

Credit scoring with a smaller gender gap

2026MIAX master's, Instituto BME

Figure 1. As the fairness penalty grows, the gender gap falls while AUC barely moves

Summary

A neural network for credit risk that cuts the gender gap by 72% while losing only 0.6 points of AUC, and estimates its own uncertainty. Team of 3.

Keras on Home Credit data, with MC Dropout for uncertainty.

The task

A neural network for credit risk on Home Credit data, focused on applicants with little credit history, that must not discriminate by gender. Four pieces: a custom layer that computes the debt-to-income ratio with a trainable saturation; a fairness loss that adds a penalty for dependence between the prediction and gender, tried with three measures; a Keras Tuner search that maps the trade-off between AUC and fairness; and an uncertainty estimate for every prediction.

Results

The chosen compromise cut the gender gap on the test set by 72% (from 5.6 to 1.6 points) while AUC fell only from 0.7404 to 0.7345. Monte Carlo dropout returns a class and a variance; the variance rises 47% only when all three external credit scores are missing.

Figures

**Figure 2.** Pareto front from Keras Tuner: accuracy against fairness

**Figure 3.** A custom trainable layer that computes the debt ratio inside the network

Built with

PythonKeras 3 + TensorFlowKeras Tunerscikit-learnpandasMC Dropout