Does DML Causal Inference Actually Work for Insurance Pricing?

The honest answer is: yes, but only if your current approach is a GLM coefficient, and only in the scenarios where confounding is structural rather than incidental.

We spent time running benchmarks on insurance-causal against a synthetic 50,000-policy UK motor book with a known data-generating process. The numbers are uncomfortable for anyone who has ever calibrated a telematics discount from a GLM coefficient.

What we actually tested

The setup mirrors a standard UK motor telematics portfolio. Treatment is a continuous normalised telematics score. True causal effect: a one-standard-deviation improvement in score raises log-odds of renewal by 0.08. The confounding is a multiplicative driver-risk term — age × NCB × region interaction — that drives both telematics scores and renewal probabilities through the same underlying risk profile.

A naive logistic GLM with age, NCB, and region dummies as main effects tries to control for this. The problem is functional form: the confounders interact multiplicatively in the true DGP, but the GLM absorbs them additively. CatBoost in the DML nuisance step naturally discovers “young driver, zero NCB, high-risk region” as a leaf. The GLM bias is structural, not a sample size problem. It persists at n = 100,000.

The numbers

The GLM bias is worst exactly where it matters most commercially. Young drivers in high-risk regions are the primary target for telematics pricing, and the GLM-calibrated discount in this segment is the most inflated.

Headline comparison:

The DML estimate converges to the true value. The GLM does not, at any sample size.

What it costs to get this wrong

A telematics discount calibrated to the GLM coefficient (0.12) rather than the causal coefficient (0.08) is set roughly 50% too aggressively. If you give a 5% discount to 40% of a 10,000-policy book expecting 8% retention uplift per unit telematics score, but the true causal uplift is only ~5% per unit, you are paying for retention that would have happened without the discount.

On renewal pricing, the same benchmark finds a GLM price-sensitivity estimate of −0.045 against a causal estimate of −0.023 — the GLM number is wrong by roughly 96%. A rate-change optimisation built on that elasticity will systematically undercharge price-inelastic customers and overprice elastic ones.

The API

from insurance_causal import CausalPricingModel

model = CausalPricingModel(
    treatment="telematics_score",
    outcome="renewal",
    confounders=["driver_age", "ncb_years", "region", "vehicle_group"],
    outcome_type="binary",
    cv_folds=5,
)
model.fit(df)
ate = model.average_treatment_effect()
print(ate)
# AverageTreatmentEffect(estimate=-0.023, std_error=0.004, ci_lower=-0.031, ci_upper=-0.015, p_value=0.0001)
bias_report = model.confounding_bias_report()
print(bias_report[["naive_estimate", "causal_estimate", "bias_pct"]])
# naive_estimate: 0.045, causal_estimate: -0.023, bias_pct: 0.49

The confounding_bias_report is the number to put in front of the pricing committee. It shows the naive estimate alongside the causal estimate, with the fraction attributable to confounding.

When DML is not the right tool

The GLM bias is a function of how nonlinear the confounding is. If treatment assignment was genuinely randomised — a properly designed A/B test, for example — the GLM is unbiased and DML is unnecessary overhead.

It also does not help with thin data. The five-fold CatBoost nuisance step needs enough policies in each fold to fit a credible propensity and outcome model. Below roughly 2,000 policies, the nuisance estimates are too noisy to improve on a well-specified GLM. The library uses sample-size-adaptive model capacity to handle this — shrinking to Ridge when n is small — but be cautious with books under 5,000 policies.

For heterogeneous effects — identifying which customer segments are most price-sensitive — the causal forest extension is worth running. On 20,000 policies with six pre-defined elasticity segments, causal forest achieves 100% confidence interval coverage versus 67% for a GLM with full interaction terms. The per-policy CATE from the forest is what allows segment-level targeting: identifying the 20% of customers who would lapse at any price increase from the 20% who will accept significant rate without flinching.

Verdict

Use DML if:

• Your treatment effect estimate comes from a GLM on observational data where treatment assignment was correlated with risk (telematics pricing, renewal pricing, channel effects)
• You are calibrating a discount or loading that will be applied commercially and you need the number to be right, not approximately right

Do not use it if:

• You have a properly randomised A/B test — a well-specified GLM is fine
• Your book is below ~5,000 policies — the nuisance models will not be credible
• You want heterogeneous segment effects without per-policy CATEs — a GLM interaction model is faster and close enough

The technique is real. The overhead is real too: 60 seconds versus under one second for a GLM. That is a one-time cost per rate review cycle, not a production scoring cost. For calibrating a discount that will apply to tens of thousands of policies, the 60 seconds is not the constraint.

uv add insurance-causal

Source and benchmarks at GitHub. Run benchmarks/run_benchmark.py (seed=42) for reproducible output.

• Double Machine Learning for Insurance Price Elasticity
• Does Constrained Rate Optimisation Actually Work?