Is TabPFN the Silver Bullet for Insurance Pricing?

Bruno Deprez; Tim Verdonck; Wouter Verbeke

arxiv: 2605.22892 · v2 · pith:MPNCMKI6new · submitted 2026-05-21 · 💱 q-fin.RM · cs.LG

Is TabPFN the Silver Bullet for Insurance Pricing?

Bruno Deprez , Wouter Verbeke , Tim Verdonck This is my paper

Pith reviewed 2026-06-30 16:25 UTC · model grok-4.3

classification 💱 q-fin.RM cs.LG

keywords TabPFNtabular foundation modelsinsurance pricingGLMXGBoostmotor insuranceclaim frequencyclaim severity

0 comments

The pith

TabPFN does not consistently outperform GLM or XGBoost for motor insurance pricing.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper evaluates whether TabPFN, a tabular foundation model that uses in-context learning from pre-training, can replace conventional methods for pricing non-life insurance. It compares TabPFN directly against generalized linear models and gradient-boosted machines on two public motor third-party liability datasets for both claim frequency and severity modeling. The tests find no reliable improvement in accuracy, much slower inference, and strong dependence on the amount of in-context data supplied. These outcomes indicate that current tabular foundation models do not yet provide a practical substitute for established actuarial tools in typical insurance settings.

Core claim

On two publicly available MTPL datasets, TabPFN does not consistently outperform GLM and XGBoost in modeling claim frequency and severity. It also shows substantially longer inference times and sensitivity to the size of the in-context training set. Tabular foundation models are promising especially in data-scarce settings but their current performance does not make them a viable replacement for established actuarial methods.

What carries the argument

In-context learning from a pre-trained tabular foundation model applied to new insurance data without any dataset-specific fitting or hyperparameter tuning.

If this is right

GLM and XGBoost remain preferable for standard motor insurance pricing workflows.
Tabular foundation models may hold more value in settings with very limited training data.
Further scaling or architectural changes to these models would be needed before they compete in data-rich actuarial tasks.
Actuaries should continue using conventional methods until tabular foundation model performance improves on representative insurance tables.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same benchmarking approach could be applied to other insurance lines such as property or liability to test whether the pattern holds.
Optimizations that reduce TabPFN inference time could change its practicality even without accuracy gains.
Hybrid systems that use TabPFN only on small subsets of data might combine its strengths with the speed of traditional models.

Load-bearing premise

The two public MTPL datasets are representative enough of real-world non-life insurance pricing problems for the negative results to generalize.

What would settle it

A new insurance dataset on which TabPFN produces lower error than both GLM and XGBoost at inference speeds comparable to XGBoost.

Figures

Figures reproduced from arXiv: 2605.22892 by Bruno Deprez, Tim Verdonck, Wouter Verbeke.

**Figure 2.** Figure 2: The RMSE and gamma deviance across the five folds for the two severity datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 2.** Figure 2: The RMSE and gamma deviance across the five folds for the two severity datasets. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Training and inference time of the models. 5k and 50k for readability reasons, but this does not change the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

read the original abstract

Modelling claim frequency and severity for non-life insurance pricing predominantly relies on generalised linear models, with gradient-boosted machines as the leading machine learning alternative. Tabular foundation models (TFMs) present a fundamentally different inference paradigm. By pre-training on large collections of synthetic datasets, TFMs enable inference on new data through in-context learning, without any dataset-specific fitting or hyperparameter tuning. This paper presents a first empirical evaluation of TabPFN for motor insurance pricing, benchmarking it against GLM and XGBoost on two publicly available MTPL datasets. Our results show that TabPFN does not consistently outperform established baselines, exhibits substantially longer inference times, and is sensitive to the size of the in-context training set. While tabular foundation models represent a promising direction, particularly in data-scarce settings, their current performance does not offer a viable replacement for established actuarial methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TabPFN does not beat GLM or XGBoost on the two MTPL tables and the paper shows the practical gaps in speed and context sensitivity.

read the letter

TabPFN does not consistently beat GLM or XGBoost on these two motor insurance tables, and the paper makes that the main result.

The work is new as the first reported head-to-head of TabPFN against the standard actuarial baselines on public MTPL data. It does a clean job of measuring what matters in practice: out-of-sample performance on frequency and severity, inference time, and how results change with the size of the in-context training set. Those details are useful for anyone who has to decide whether to try these models in a pricing pipeline.

The paper is straightforward about the downsides. Longer inference times and clear sensitivity to context size are real constraints, and the authors do not claim the method is ready to replace existing tools. They leave room for future improvement in data-scarce regimes without overstating the current evidence.

The soft spot is the data scope. Both datasets are MTPL, so the claim that tabular foundation models are not yet a viable replacement rests on how representative these tables are of other non-life lines. The stress-test note is right: different claim frequencies, feature distributions, or scarcity levels could shift the picture, and the paper gives no additional checks on that point.

A reader working on insurance pricing models would get value from the benchmark. The thinking is direct and the comparison is honest, so the paper deserves a serious referee even with the negative finding.

Referee Report

1 major / 0 minor

Summary. The manuscript presents an empirical benchmark of TabPFN, a tabular foundation model using in-context learning, for non-life insurance pricing of claim frequency and severity. It compares TabPFN against GLM and XGBoost on two publicly available MTPL datasets, finding that TabPFN does not consistently outperform the baselines, exhibits substantially longer inference times, and is sensitive to in-context training set size. The authors conclude that while TFMs are promising in data-scarce settings, they do not currently offer a viable replacement for established actuarial methods.

Significance. If the results hold, this provides a useful first empirical evaluation of tabular foundation models in insurance pricing, documenting practical limitations (inference time, data sensitivity) that temper enthusiasm for immediate adoption. The work credits the in-context learning paradigm as distinct from fitted models like GLM/XGBoost and focuses on real insurance data, which strengthens its relevance for the actuarial community.

major comments (1)

[Abstract and §4 (Results)] The central claim that TabPFN 'does not offer a viable replacement for established actuarial methods' is load-bearing on the representativeness of the two MTPL datasets. The abstract and results provide no justification, additional datasets, or sensitivity analysis showing that findings extend to other non-life lines, claim frequency regimes, or feature distributions; this directly affects the scope of the conclusion.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the single major comment below.

read point-by-point responses

Referee: [Abstract and §4 (Results)] The central claim that TabPFN 'does not offer a viable replacement for established actuarial methods' is load-bearing on the representativeness of the two MTPL datasets. The abstract and results provide no justification, additional datasets, or sensitivity analysis showing that findings extend to other non-life lines, claim frequency regimes, or feature distributions; this directly affects the scope of the conclusion.

Authors: We agree that the manuscript's conclusions should be more narrowly scoped. The study is explicitly positioned as a first empirical evaluation on two publicly available MTPL datasets, which are standard benchmarks in the actuarial literature. MTPL constitutes a major share of non-life business, but we do not claim or demonstrate representativeness across all non-life lines, frequency regimes, or feature distributions. We will revise the abstract and the discussion in §4 to state that the observed limitations of TabPFN apply to the MTPL setting studied here, and that broader applicability to other lines remains an open question requiring further research. We will also add a short paragraph contextualizing the two datasets' characteristics (e.g., claim frequency levels and feature types) without adding new data or performing cross-line sensitivity analyses, which are outside the scope of the current work. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark with no derivation chain

full rationale

The paper is a straightforward empirical comparison of TabPFN against GLM and XGBoost on two fixed public MTPL datasets. It reports observed performance metrics, inference times, and sensitivity to in-context size without any claimed first-principles derivation, fitted-parameter prediction, uniqueness theorem, or self-citation that reduces the central result to its own inputs. No equations or ansatzes are invoked that could create self-definitional or load-bearing circularity. The findings are falsifiable against the stated tables and do not rely on renaming known results or smuggling assumptions via prior work by the same authors. This is the normal case of an empirical study whose conclusions rest on external data rather than internal construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated premise that the chosen public datasets and evaluation protocol are adequate proxies for production insurance pricing.

pith-pipeline@v0.9.1-grok · 5677 in / 1102 out tokens · 33426 ms · 2026-06-30T16:25:09.034194+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Foundation Models for Credit Risk Prediction: A Game Changer?

Baesens, B., Goethals, A., Lessmann, S., V os, S. D., Bravo, C., Martens, D., Medina-Olivares, V ., Mues, C., Os- karsd´ottir, M., vanden Broucke, S., Verdonck, T. & Verbeke, W. (2026), ‘Foundation models for credit risk predic- tion: A game changer?’,arXiv preprint arXiv:2605.18147. URL:https://arxiv.org/abs/2605.18147 Chu, J. Z. K., Than, J. C. M. & Jo,...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1080/10920277.2020.1745656 2026
[2]

P., Grabocka, J

URL:https://doi.org/10.1080/10920277.2025.2451860 M¨uller, S., Hollmann, N., Arango, S. P., Grabocka, J. & Hutter, F. (2022), Transformers can do bayesian inference,in ‘International Conference on Learning Representations’. URL:https://openreview.net/forum?id=KSugKcbNf9 Padayachy, K., Richman, R., Scognamiglio, S. & W ¨uthrich, M. V . (2026), ‘In-context ...

work page doi:10.1080/10920277.2025.2451860 2025

[1] [1]

Foundation Models for Credit Risk Prediction: A Game Changer?

Baesens, B., Goethals, A., Lessmann, S., V os, S. D., Bravo, C., Martens, D., Medina-Olivares, V ., Mues, C., Os- karsd´ottir, M., vanden Broucke, S., Verdonck, T. & Verbeke, W. (2026), ‘Foundation models for credit risk predic- tion: A game changer?’,arXiv preprint arXiv:2605.18147. URL:https://arxiv.org/abs/2605.18147 Chu, J. Z. K., Than, J. C. M. & Jo,...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1080/10920277.2020.1745656 2026

[2] [2]

P., Grabocka, J

URL:https://doi.org/10.1080/10920277.2025.2451860 M¨uller, S., Hollmann, N., Arango, S. P., Grabocka, J. & Hutter, F. (2022), Transformers can do bayesian inference,in ‘International Conference on Learning Representations’. URL:https://openreview.net/forum?id=KSugKcbNf9 Padayachy, K., Richman, R., Scognamiglio, S. & W ¨uthrich, M. V . (2026), ‘In-context ...

work page doi:10.1080/10920277.2025.2451860 2025