Is TabPFN the Silver Bullet for Insurance Pricing?

Technical Summary: Is TabPFN the Silver Bullet for Insurance Pricing?

Problem Statement
Non-life insurance pricing traditionally relies on Generalized Linear Models (GLMs) for modeling claim frequency and severity, often augmented by Gradient-Boosted Machines (like XGBoost) as the leading machine learning alternative. While these methods are transparent and well-understood, they require dataset-specific preprocessing, fitting, and hyperparameter tuning. Recently, Tabular Foundation Models (TFMs), specifically TabPFN, have emerged as a paradigm shift. Trained on millions of synthetic datasets using Prior-data Fitted Networks (PFN), TabPFN utilizes in-context learning (ICL) to approximate posterior predictive distributions without dataset-specific training. However, existing benchmarks for TabPFN focus on general prediction accuracy and lack evidence regarding its suitability for the specific actuarial requirements of insurance pricing, such as distributional fit and operational latency.

Methodology
This paper provides the first empirical evaluation of TabPFN for motor insurance pricing, benchmarking it against standard GLM and XGBoost models.

• Datasets: The study utilizes two publicly available Motor Third Party Liability (MTPL) datasets: the French freMTPL2 and the Belgian beMTPL97.
• Models:
  • Baselines: A Poisson GLM for frequency and a Gamma GLM for severity; XGBoost for both tasks.
  • TabPFN: Two versions were tested: TabPFN-v2.6 (scaling to 100k samples) and TabPFN-v3 (scaling to 1M samples, provided features < 200).
• Experimental Design:
  • Preprocessing: TabPFN operates on raw inputs; GLM and XGBoost received ordinal encoding for ordered categoricals and one-hot encoding for nominal categoricals.
  • Validation: 5-fold cross-validation was used.
  • Context Sizes: For frequency modeling, TabPFN was evaluated with varying context (training) sizes: 2,000, 5,000, 10,000, 50,000, 100,000, and the full training set. Severity modeling used the maximum context size.
  • Metrics:
    • Predictive Accuracy: Root Mean Squared Error (RMSE), exposure-weighted for frequency and unweighted for severity.
    • Distributional Fit: Poisson deviance for frequency and Gamma deviance for severity.
    • Operational Cost: Inference time (for TabPFN) and combined training/inference time (for baselines).

Key Results

• Predictive Performance (Frequency): TabPFN does not consistently outperform established baselines. At large context sizes (≥50,000), TabPFN-v2.6 achieved lower RMSE than XGBoost on both datasets but failed to surpass GLM. TabPFN-v3 performed similarly to baselines on French data but worse on Belgian data.
• Distributional Fit (Frequency): GLM consistently achieved the best Poisson deviance. TabPFN-v2.6 showed high sensitivity to context size; deviance performance was non-monotonic, worsening at intermediate sizes (2k–10k) before improving at larger sizes (50k–100k). TabPFN-v3 exhibited relatively stable but high (poor) deviance across all context sizes. Fold-to-fold variance for TabPFN was significantly higher than for GLM or XGBoost, particularly with smaller context sizes.
• Severity Modeling: Both TabPFN versions underperformed compared to GLM and XGBoost. GLM performed best on Belgian data, while XGBoost excelled on French data. TabPFN models exhibited larger variance in Gamma deviance across folds.
• Computational Cost: Inference time for TabPFN grows steeply with context size. Even at small context sizes, TabPFN's inference cost substantially exceeds the combined training and inference time of GLM and XGBoost. At capacity limits, inference takes several minutes, rendering it unsuitable for real-time, high-throughput pricing environments. TabPFN-v3 showed lower inference times than v2.6 but sacrificed accuracy on smaller datasets.

Key Contributions

1. First Empirical Assessment: This paper offers the initial benchmarking of TabPFN specifically for non-life insurance pricing, moving beyond general regression benchmarks to address actuarial-specific metrics (deviance) and constraints (latency).
2. Context Size Sensitivity: The study highlights that TabPFN's performance is highly sensitive to the size of the in-context training set, with intermediate sizes potentially yielding unreliable distributional fits.
3. Operational Reality Check: The paper demonstrates that despite the "no-training" promise of foundation models, the computational cost of in-context learning creates a bottleneck for real-time insurance applications, where GLM and XGBoost remain superior in efficiency.

Significance and Claims
The paper concludes that while Tabular Foundation Models represent a promising direction, particularly for data-scarce settings, TabPFN is not currently a viable replacement for established actuarial methods in insurance pricing. The model fails to consistently outperform GLMs or XGBoost, exhibits higher variance in predictions, and suffers from prohibitive inference times for real-time deployment.

The authors modestly suggest that foundation models may still find utility elsewhere in the insurance value chain, such as fraud detection (classification/anomaly detection), loss reserving forecasting, or mortality projections, but their current formulation does not solve the core problem of pricing. Future work is proposed to investigate fine-tuning these models with actuarial knowledge and analyzing their interpretability regarding risk drivers.