Step 2: Design fair pricing
A practical framework for fair algorithmic pricing
How to achieve fairness?
Implement anti-discrimination strategies at pre-processing, in-processing, or post-processing, linked to the fairness criteria from Step 1 (Xin and Huang 2024).
Model–criterion mapping
| Model | Criterion | Approach |
|---|---|---|
| M0 (full) | baseline | Uses X_P and X_{NP} |
| MU (unawareness) | FTU | \hat{Y} = f(X_{NP}) only |
| MDP (debiased) | DP | Pre-process X_{NP} to remove dependence on X_P |
| MCDP | CDP | Legitimate X_{NP_{\text{legit}}} retained; others debiased |
| MC (control) | CPV | Fit M0; average over X_P at prediction |
Pre-processing (MDP, MCDP)
- Disparate impact remover: align group distributions of X_{NP}
- Orthogonal predictors: residualise X_{NP} on X_P
MCDP allows group differences only through legitimate variables (e.g. claims history, vehicle type).
Post-processing (MC)
\hat{Y}_{MC} = \frac{1}{N}\sum_{j=1}^N \hat{f}_{M0}(X_{NP}, X_P = x_{p_j})
Protected attributes used in training only; averaged out at prediction, giving better proxy control than MU alone.
Implementation
- Empirical comparison uses GLM and XGBoost on French auto data (Xin and Huang 2024)
- Retain baseline M0 for welfare and audit comparisons (Steps 3–4)
- Document pre- / in- / post-processing choice and reproducible pipeline
Checklist
References
Xin and Huang (2024)
References
Xin, Xi, and Fei Huang. 2024. “Antidiscrimination Insurance Pricing: Regulations, Fairness Criteria, and Models.” North American Actuarial Journal 28 (2): 285–319.