Step 1: Define fairness
A practical framework for fair algorithmic pricing
What is fair?
Agree which fairness metrics apply, under which regulations, and on which outputs (premium, markup, approval) before building or auditing a model.
Principles of fair differentiation
Insurance pricing is built on differentiation. Frees and Huang (2023) asks which differentiations are appropriate:
| Lens | Emphasis | Examples |
|---|---|---|
| Social good | Solidarity, access, mandatory cover | Compulsory third-party auto |
| Economic commodity | Efficiency, adverse selection | Voluntary life |
Rating factor principles
| Principle | Question |
|---|---|
| Control | Can the policyholder influence the factor? |
| Mutability | Does it change over time or stay fixed? |
| Statistical discrimination | Does it predict risk? |
| Causality | Does it cause insured events? |
| Past discrimination | Does use reinforce historical injustice? |
| Socially valuable behaviour | Does pricing discourage beneficial actions? |
Proxy discrimination and disparate impact are central grey areas (Frees and Huang 2023).
Regulatory landscape
Xin and Huang (2024) maps regimes to modelling choices:
- Restriction or prohibition on protected variables
- Proxy restrictions
- Disparate impact tests
- Community rating
Fairness criteria
Let X_P be a protected attribute, X_{NP} other features, \hat{Y} the predicted premium (Xin and Huang 2024).
| Type | Criterion | Idea |
|---|---|---|
| Individual | FTU | \hat{Y} does not use X_P |
| Individual | CPV | Discrimination-free price averaging over X_P |
| Group | DP | Equal distribution of \hat{Y} across groups |
| Group | CDP | Parity within legitimate subgroups |
| Group | Separation / sufficiency | Equalised errors or calibration |
Krafcheck, Balnozan, and Huang (2026) distinguishes individual (actuarial) fairness from group fairness; achieving both is often impossible, so document the trade-off.
Checklist
References
Frees and Huang (2023)
Xin and Huang (2024)
Krafcheck, Balnozan, and Huang (2026)