When model outputs inform decisions, bias detection checks whether they show systematic disparities across groups or patterns that could lead to unfair results. section 8.5.2
Check outputs in several ways:
- Compare trends and performance across groups or data subsets, and seek feedback from users or stakeholders. section 8.5.2
- Use disparate impact analysis, confusion matrices to inspect false positives and negatives, or residual analysis to find patterns in prediction errors. section 8.5.2
- Continue evaluating outputs as data and identified biases change; reassess approved use cases and update models when needed. section 8.5.4
Residential loan underwriting
A model provides default probabilities as advice to an underwriter, who makes the final lending decision. Evaluate its outputs for bias, including possible effects of skewed training data. Output-level adjustments such as reranking or resampling may help, but need careful review and do not replace checking whether the model fits the use case. section 8.5.2
