The question pricing teams should be asking is not “does my model use a protected characteristic?” — it does not, because that would be illegal. The question is “does my model use factors that are correlated with a protected characteristic in a way that a court would call indirect discrimination under Section 19 of the Equality Act 2010?”

That is a harder question, and the standard approach to answering it is not adequate.

What most teams currently do

The conventional proxy check is a Spearman rank correlation between each rating factor and the protected characteristic. Set a threshold — typically r > 0.25 — and flag anything above it. It is fast, it produces a number, and it fits in a single column of a compliance spreadsheet.

We benchmarked this against insurance-fairness across 50 random seeds on 20,000 synthetic UK motor policies with a known postcode-ethnicity proxy relationship — the structure replicating the Citizens Advice (2022) finding that postcode loading drives a roughly £280/year ethnicity penalty in UK motor.

The benchmark result

Metric Manual Spearman (threshold r > 0.25) Library proxy R² + MI
Postcode flagged as proxy No ( r ≈ 0.10, below threshold) Yes (proxy R² ≈ 0.62, RED)
Factors correctly flagged (out of 6) 0/6 1–2/6    
Detection rate (50 seeds) 0/50 50/50    
Mean postcode Spearman r   ~0.10 N/A
Mean postcode proxy R² N/A ~0.62 (std ≈ 0.02)    
Captures non-linear proxy No Yes    
Quantifies financial impact No Yes (£/policy)    

The Spearman check detects nothing, across all 50 seeds, at any reasonable threshold. The library detects the proxy with a mean R² of 0.62 in every single run.

This is not a failure of the threshold. The Spearman r of 0.10 is genuinely small. The issue is that postcode area encodes ethnicity non-linearly: Inner London postcodes have high ethnic diversity, outer postcodes lower, rural postcodes the lowest — but within each postcode area the relationship is not monotone, and the Spearman statistic, which measures rank correlation, cannot detect non-linear categorical-to-continuous relationships where rank correlation is near zero despite strong predictive power.

The library fits a CatBoost model predicting the protected attribute from each rating factor in isolation. Proxy R² of 0.62 means postcode area explains 62% of variance in the diversity score. That number provides quantitative evidence relevant to a Section 19 assessment — establishing the substantial disadvantage element — and is the kind of finding a regulator can act on.

The financial impact

The benchmark also quantifies the premium differential. High-diversity policyholders pay roughly £70–90 more per year than low-diversity policyholders, driven through the postcode area loading. This is the number a pricing committee needs to decide whether to act.

A Spearman correlation of 0.10 does not produce this number. A proxy R² of 0.62 demands it.

The API

from insurance_fairness import FairnessAudit

audit = FairnessAudit(
    model=None,                   # None to use pre-computed predictions in data
    data=policies_df,
    protected_cols=["diversity_score"],
    prediction_col="annual_premium",
    outcome_col="claim_cost",
    factor_cols=["postcode_area", "vehicle_group", "ncd_years", "age_band", "mileage", "payment_method"],
    exposure_col="exposure",
)

report = audit.run()
report.summary()
# Per-characteristic demographic parity, calibration disparity, proxy flags

# Per-factor proxy scores (list of ProxyScore objects, sorted by R²):
for s in report.results["diversity_score"].proxy_detection.scores:
    print(f"{s.factor}: proxy_r2={s.proxy_r2:.2f}, mi={s.mutual_information:.2f}, flag={s.rag}")
# postcode_area: proxy_r2=0.62, mi=0.41, flag=RED
# vehicle_group: proxy_r2=0.08, mi=0.06, flag=GREEN

report.to_markdown("fairness_audit.md")  # FCA-ready report with regulatory mapping

The to_markdown() output includes explicit regulatory cross-references to PRIN 2A, Consumer Duty (PS22/9), and Equality Act Section 19. It is structured for a pricing committee pack — one document that can go to legal, compliance, and the pricing team simultaneously.

The action/outcome distinction

One finding from the benchmark is less obvious but more important. Minimising premium disparity (action fairness) does not minimise loss ratio disparity (outcome fairness). On a 20,000-policy TPLI portfolio, the policy with the most equal premiums produced the most unequal loss ratios.

This is the compliance gap the FCA has highlighted in its Consumer Duty supervisory work. Firms were auditing at the point of quoting — checking whether premiums are equal across groups — and missing the Consumer Duty Outcome 2 (Price and Value) obligation, which is a post-sale value question. A firm paying claims at the same rate across groups regardless of risk has cross-subsidised protected groups, which looks fair but is not economically sustainable. A firm pricing to true risk produces premium differentials that may look unfair but are the correct actuarial answer.

The DoubleFairnessAudit class in the library computes the Pareto front across action fairness and outcome fairness. A pricing committee can see every operating point along the trade-off, with quantified evidence of the choice being made. A firm that can only show a single demographic parity ratio cannot demonstrate the same level of considered decision-making to the FCA.

The FCA context

The FCA’s 2024 multi-firm review of Consumer Duty implementation found most insurers’ Fair Value Assessments were “high-level summaries with little substance”. The FCA has signalled that fair value enforcement is active and ongoing. This is live enforcement risk, not theoretical.

Citizens Advice (2022) put the ethnicity penalty at £280/year and £213 million per year in aggregate for UK motor — and this is driven by postcode factors that every standard motor pricing model uses. The question is whether your book has the same structure. A Spearman check at r = 0.10 will not tell you.

Limitations

The proxy detection works on the factors in isolation — it tests whether postcode area is a proxy for ethnicity when considered alone. It does not test for indirect proxying through combinations of factors. If vehicle group and payment method are individually clean but jointly correlated with a protected characteristic, the single-factor proxy test misses it. The SHAP proxy score addresses this partially by tracing price impact back through the full model structure.

Computation time for proxy R² is 12–15 seconds per factor on 50,000 policies (80 CatBoost iterations). For a 20-factor tariff, run on Databricks. The mutual information check is a useful first pass — under 5 seconds for the full factor set — to identify candidates before running CatBoost proxy detection.

Verdict

Use proxy discrimination testing if:

The Spearman correlation check at r > 0.25 is not adequate for demonstrating absence of proxy discrimination. A CatBoost proxy R² of 0.10 or above on any rating factor warrants investigation. At 0.62, you have a documented problem and a £/policy number that quantifies it.

When a RED proxy flag fires, there are three responses available: remove the factor from the rating structure; retain it and mount an actuarial justification under the Section 19(2)(d) defence (the factor reflects genuine risk differentiation and the means used are appropriate and reasonably necessary); or seek external legal advice if the exposure is material and the actuarial justification is contested. None of these decisions is quick, which is why the documentation needs to be in place before a regulatory enquiry, not after.

uv add insurance-fairness

Source and benchmarks at GitHub. The Monte Carlo sensitivity benchmark (benchmarks/benchmark_sensitivity.py) runs 50 seeds and produces the detection rate table.

fairnessproxy-discriminationfcaconsumer-dutyequality-actpricingpython

Related reading