What is synthetic data in actuarial ratemaking?

Synthetic data in actuarial ratemaking is artificially generated data that preserves the statistical properties of a real insurance dataset, including univariate distributions, bivariate correlations, and multivariate relationships between rating variables, while containing no actual policyholder personally identifiable information. The 2025 CAS Ratemaking Prize-winning paper demonstrated that kernel density estimation can generate synthetic datasets with high fidelity for pricing model development.

Has any state insurance department accepted a rate filing based on synthetic data?

As of June 2026, no state department of insurance has formally accepted a rate filing where synthetic data was cited as a primary modeling input. Rate filings submitted through SERFF are evaluated against adequacy, non-excessiveness, and non-discrimination standards, and regulators have not yet issued guidance on how synthetic data inputs should be documented or validated in the filing process. This regulatory acceptance gap is the primary bottleneck for production deployment.

How does ASOP No. 56 apply to synthetic data in actuarial models?

ASOP No. 56 (Modeling) requires actuaries to document data used as inputs to models and to assess whether data is consistent with the intended purpose. When synthetic data replaces or supplements real policyholder data, the actuary must document the generation methodology, validate that statistical properties are preserved, disclose the substitution to intended users, and assess whether the synthetic data introduces bias or distortion that affects model output. The standard does not prohibit synthetic data but requires the actuary to exercise professional judgment about its suitability.

Synthetic Data Wins CAS Ratemaking Prize: Privacy-Safe Pricing With Kernel Density Estimation

The CAS E-Forum published a prize-winning paper proving that kernel density estimation can generate synthetic insurance datasets that preserve statistical pricing relationships while eliminating policyholder PII. With 79% of carriers open to synthetic data and state privacy laws tightening, this is the first rigorous actuarial framework for deploying AI models without exposing real data. No state DOI has formally accepted a synthetic-data-based rate filing.

The Casualty Actuarial Society awarded its $15,000 2025 Ratemaking Prize to Noa Zamstein of Earnix for "Enhancing Actuarial Ratemaking with Synthetic Data for Privacy Preservation," published in CAS E-Forum Quarter 1, 2025 and presented at the RPM Seminar in Orlando on March 10, 2025. The paper shows kernel density estimation can generate insurance datasets that carry no policyholder personally identifiable information and still support pricing model development. No state department of insurance has yet accepted a rate filing citing synthetic data as a modeling input.

Key Takeaways

79% of carriers are open to using or already using synthetic data for privacy and data-quality problems, in SAS survey data cited by Roots Automation.
Three fidelity levels are tested in the prize paper: univariate distribution matching, bivariate relationship preservation and multivariate structure retention. GLM coefficient comparability is the one that decides whether the data is usable for ratemaking.
40,000 observations in a separate Earnix auto renewal database reproduced the result for gradient boosting: comparable feature importance rankings and price elasticity distributions close enough to support scenario testing.
24 states have adopted the NAIC Model Bulletin on AI use by insurers, whose data governance provisions cover quality, integrity, bias and privacy but do not name synthetic data as a distinct practice.

What the Prize Paper Demonstrates

Kernel density estimation treats each record in the original dataset as the center of a small distribution, typically Gaussian, and sums those kernels into a smooth estimate of the density the data came from. Synthetic records are sampled from that estimate rather than copied from the source, which is why none of them is a policyholder.

The extension that matters is multivariate. Ratemaking data is correlated by construction: driver age moves with vehicle type, which moves with territory, which moves with frequency. Synthesizing column by column reproduces every marginal distribution and destroys every relationship between them, which leaves a dataset that looks right and prices nothing.

Zamstein's approach preserves all three levels, and the third is the one with a practical test attached. A GLM fitted on the synthetic data returns coefficient estimates close to what the original data produces. That is the claim a pricing actuary can check directly, on their own book, without taking the method on faith.

Earnix reproduced it independently on a different dataset and a different model class, which is what separates a generalizable method from a dataset-specific artifact.

Where the Fidelity Requirement Binds

The comparison that matters is not synthetic against real. It is synthetic against the alternatives, because the privacy problem has several proposed solutions and they fail in different places.

Method	Fidelity	Privacy Guarantee	Complexity	Insurance Fit
Kernel Density Estimation (KDE)	High across univariate, bivariate, and multivariate structure	Empirical (no formal epsilon bound)	Low to moderate; standard statistical libraries	Strong for tabular rating data; CAS prize-validated
MICE (Imputation-Based)	High; preserves GLM coefficient comparability	Empirical	Low to moderate; well-understood in actuarial practice	Strong for tabular data with mixed variable types
Conditional Tabular GAN (CTGAN)	High for complex non-linear relationships	Empirical; DP-CTGAN variant adds formal bounds	High; requires substantial data and extensive tuning	Better for large datasets with non-linear structure
Variational Autoencoder (VAE)	Moderate to high; encodes to latent space	Empirical; DP-TVAE variant adds formal bounds	High; challenges applying text-based architectures to tabular data	Emerging; less validated on insurance-specific data
Differential Privacy (noise injection)	Reduced; noise degrades statistical relationships	Formal mathematical guarantee (epsilon-delta)	Moderate; parameter tuning affects utility tradeoff	Treats features independently; can miss dependencies

Differential privacy is the instructive failure. It offers the only formal guarantee on the list, a mathematical bound on how much any single record can influence the output, and it purchases that guarantee by injecting noise feature by feature. Treating features independently is precisely what a ratemaking dataset cannot survive. The dependency between age, territory and frequency is not a side effect of the data; it is the object of the analysis.

That is the shape of the whole tradeoff. Tighten the privacy bound and the correlations degrade; loosen it and the guarantee stops meaning anything. An actuary who accepts a distorted correlation structure to avoid a privacy exposure has exchanged privacy risk for model risk, and the second one lands in the rate.

Which is why the validation battery has to run past marginal distributions. Kolmogorov-Smirnov or density plots for continuous variables and chi-squared for categoricals establish only the first level. Correlation matrices and cross-tabulations establish the second. The third is the direct comparison of coefficients, standard errors and predicted values between a model fitted on synthetic data and the same specification fitted on the original, and on the privacy side, membership inference tests that should perform no better than random guessing.

ASOP No. 56 already reaches this work, and it reaches it twice. Section 3.2 asks whether the data is appropriate for the intended purpose; Section 3.7 covers reliance when the generation was done by a data science team or a vendor whose method sits outside the actuary's expertise. The documentation load is higher than for real data because the actuary is justifying both the dataset and the process that produced it.

The Filing Has Nowhere to Put It

The bottleneck is not technical. A rate filing is evaluated against rates being adequate, not excessive and not unfairly discriminatory, and the reviewing actuary at the department assesses whether the data, methodology and assumptions supporting the indication are sound. An exhibit reading "model trained on synthetic data generated via KDE from the company's policyholder records" gives that reviewer nothing to work from.

There is no precedent filing to compare it against and no bulletin to cite. The NAIC's AI evaluation tool, piloted across 12 states, addresses model transparency, bias testing and outcome validation, and does not treat synthetic training data as a category needing its own documentation. The Model Bulletin's data governance controls in 24 states cover quality, integrity, bias and suitability without naming the practice.

So the actuary can demonstrate that the model performs comparably to one built on real data and still cannot represent that fact in a filing in a form the reviewer is equipped to evaluate. The gap is between a validated method and an institution with no field for it.

ASOP No. 56 does not close that gap either, for the same reason it does elsewhere: it binds the actuary's professional judgment, not the department's review standard. It predates practical synthetic generation, and the parallel gap in machine learning applied to loss reserving shows the pattern is not specific to this technique. Absent a practice note or a state pilot that establishes what adequate synthetic-data documentation looks like, each actuary who files one is proposing the standard as well as meeting it.

Synthetic Data Wins CAS Ratemaking Prize: Privacy-Safe Pricing With Kernel Density Estimation

Key Takeaways

What the Prize Paper Demonstrates

Where the Fidelity Requirement Binds

The Filing Has Nowhere to Put It

Further Reading

Sources