Governing fairness and compliance in pricing models

Set up the audit: get the data you are missing

Here is the catch. You cannot measure disparate impact by race if you never collected race. Most insurers do not, and often cannot, ask.

The standard workaround is proxy estimation. The most cited method is Bayesian Improved Surname Geocoding (BISG), which estimates the probability that a policyholder belongs to a racial group using their surname and geographic location. The CFPB published its BISG methodology as a public reference.

BISG is imperfect. It gives you a probability, not a fact. But it lets you estimate group-level impact, which is what fairness testing needs.

Step 1: Test each variable for proxy behavior

Start with the inputs. For every rating variable, ask: how strongly does it predict a protected class?

A simple, defensible first pass is to regress your estimated race probabilities against each feature and check the strength of the relationship.

A high R-squared means the variable carries a lot of protected-class signal. That does not automatically make it illegal, but it flags the variable for justification. Credit-based insurance scores and geographic tiers are the usual suspects.

Step 2: Measure disparate impact on the output

Now test the price the model produces, not just the inputs.

Two common metrics:

Disparate impact ratio. Compare the rate at which each group receives an unfavorable outcome (for example, being placed in the top pricing tier). A ratio far from 1.0 signals imbalance. Regulators and courts sometimes reference a "four-fifths rule" (a ratio below 0.80) borrowed from employment law, though it is a rough guide, not a legal standard in insurance.

Group premium comparison. Weight premiums by BISG probabilities and compare average predicted premium across estimated groups. If group A pays materially more, you need to show that group A also has materially higher expected losses.

The key test is conditional: does the price gap remain after you control for legitimate risk factors? If two drivers have identical loss expectations but different estimated-race groups pay different prices, that is the smoking gun.

Step 3: Decide what to do about a flagged variable

You have three honest options for a variable that shows proxy behavior:

1. Justify and keep it. Show that the variable predicts loss for a real, causal reason, and that its predictive power does not collapse once you account for race. Document the actuarial rationale.

2. Modify it. Replace a coarse proxy with a more direct measure. Example: instead of raw ZIP code, use garaging-based catastrophe risk (hail, flood) that has a physical cause and less racial correlation.

3. Drop it. If a variable adds little accuracy but carries heavy proxy signal, removing it is often the cleanest path.

Beware proxy leakage through interactions. Removing race does nothing if five other variables jointly reconstruct it. Test combinations, not just single features.

Step 4: Fairness-aware modeling (use with caution)

You can also adjust the model itself. Techniques include reweighting training data, adding fairness constraints during training, or post-processing scores to equalize outcomes across groups.

Big caveat for insurance: deliberately altering prices by protected class can itself be illegal in some states, even if the goal is fairness. Adjusting a Black applicant's price because they are Black is disparate treatment, which is generally banned outright. This is a genuine legal tension. Involve counsel before applying any correction that uses protected-class estimates in pricing.

For most filings, the safer route is Step 3: fix the inputs, not the output.

Step 5: Document to survive the rate filing

A rate filing is the submission an insurer makes to a state insurance department seeking approval for its rates. Most states require rates to be "not unfairly discriminatory," among other standards. The NAIC (National Association of Insurance Commissioners) has issued model bulletins on AI use that many states have adopted, pushing insurers toward documented governance.

Your documentation package should include:

• Variable inventory. Every input, its source, and its actuarial justification.
• Proxy analysis. The tests from Step 1, with results and your rationale for keeping or dropping each flagged variable.
• Disparate impact results. Metrics from Step 2, including the conditional test.
• Model card. A plain-language summary of what the model does, its data, its limits, and its intended use.
• Governance trail. Who reviewed it, when, and what changed. Regulators want to see a repeatable process, not a one-time cleanup.

Write it so a non-technical examiner can follow it. Clarity is a compliance asset.

A worked mini-example

Suppose your model puts 22 percent of estimated-white applicants and 41 percent of estimated-Black applicants into the top premium tier. Ratio: 0.22 / 0.41 = 0.54. That is well below 0.80 and will draw scrutiny.

You investigate. After controlling for prior claims, annual mileage, and vehicle safety rating, the gap shrinks to a ratio of 0.91. Most of the disparity was explained by legitimate loss factors.

But a residual gap remains, and you trace it to credit-based insurance score, which carries strong proxy signal in your data. You test dropping it. Accuracy falls only slightly, and the ratio rises to 0.97.

Decision: drop or replace the credit factor for this product, document the tradeoff (small accuracy loss, large fairness gain), and file with the full analysis attached. That story is defensible.

Key takeaways

• Blindness is not fairness. A model that excludes race can still discriminate through proxies like ZIP code and credit score. Test inputs and outputs both.
• Measure with estimated groups. Since insurers rarely hold race data, use documented proxy methods like BISG to estimate group-level impact, and acknowledge their limits.
• Run the conditional test. The question is whether price gaps survive after controlling for legitimate loss factors. Unexplained residual gaps are the real risk.
• Fix inputs, not outputs. Adjusting prices directly by protected class can be illegal disparate treatment. Prefer justifying, modifying, or dropping proxy variables, and consult counsel before any correction.
• Documentation is the deliverable. A clear variable inventory, proxy analysis, impact metrics, model card, and governance trail are what actually gets a filing approved.