Scaling Laws, Tabular Data and Actuarial Ratemaking Models
cs.LG · q-fin.RM
Quick verdict
Claims tabular scaling laws hold in actuarial ratemaking (TabM α≈0.41); credible single-dataset study, but tiny gains and no GBDT baseline limit impact.
Summary
This paper applies the scaling-law framework from large language models to actuarial ratemaking in what the author presents as the first systematic such study. On a proprietary motor portfolio of ~4.5 million policy-period records with 60–70 rating factors, GLMs, FFNs, TabM/TabM-mini, and multiple Transformer variants (MultiCLS, head-fix enhancements, TokenMoE, swap-style self-supervision, Transformer+TabM hybrids) are trained across data fractions of 5%–100% with five seeds each, evaluated by held-out Poisson deviance and summarized via power laws. TabM-style models show the strongest data scaling (α≈0.309 on the main sweep; ≈0.409 including full-data TabM-mini extensions) and parameter scaling (β≈0.148), while purely supervised Transformers scale weakly with parameters (β≈0.025–0.037), improving to β≈0.055 with the self-supervised objective. At full data the best model beats the GLM by only 0.00315 mean deviance (~1.1% relative). Small-data regimes favor GLMs and Transformer+TabM hybrids; the large-data crossover to TabM is supported at 75–100%. Main limitations: one proprietary dataset with an IID split, no GBDT baseline sweep, heuristically fixed hyperparameters, and uncertainty quantified only at seed level; a public French MTPL supplement with code partially mitigates reproducibility.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Held-out Poisson deviance follows smooth power-law improvement with training-data size for all tested model families on a real ~4.5M-row motor portfolio.
- TabM/TabM-mini exhibit the strongest data scaling (α≈0.309 main sweep, ≈0.409 with full-data extensions) and parameter scaling (β≈0.148), unlike supervised Transformers (β≈0.025–0.037).
- Purely supervised Transformers scale weakly with parameters unless augmented with TabM-style adaptation or swap-style self-supervision (TokenMoE+SSL β≈0.055).
- Small-data regimes favor GLMs and Transformer+TabM hybrids, with a stable crossover to TabM only at 75–100% of the ~4M-row training set.
- Under C∝NP, compute-optimal allocation favors growing parameters faster than data for TabM (β/α≈0.36–0.40) versus ≈0.12 for Transformers and FFNs.
Strengths
Fills a real benchmark-gap: million-row tabular scaling
Table 1 contrasts OpenML-CC18 (≤10^5 rows) and TabArena (~10^3–1.5×10^5) with this study's 2×10^5–4×10^6 policy-period rows, enabling scaling questions small benchmarks cannot answer.
Unusually honest effect-size reporting
Section 7.3 states the full-data GLM→best-model gain is only 0.00315 mean deviance (~1.1% relative), and Section 7.3.2/Table 6 explicitly declare the 50% crossover (ensemble gap 0.00004) unstable since paired single-seed differences favor the runner-up.
Mechanistic diagnosis of Transformer scaling failure
Credibility-theory bridge gives exponents actuarial meaning
Fit robustness and regime diagnostics
Weaknesses
No GBDT baseline
Envelope-based exponent estimation is selection-sensitive
All exponents conditioned on one shared fitted floor
Heuristic, asymmetric tuning across families
IID split despite acknowledged non-stationarity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.03106, cs.LG/q-fin.RM) asks whether LLM-style scaling laws — L(N) ≈ L∞ + AN·N^−α — hold for actuarial claim-frequency models. On one anonymized multinational motor portfolio (~4.5M policy-period records, 60–70 rating factors, pre-COVID), the author trains GLMs, FFNs, TabM/TabM-mini, and a progression of Transformer variants (vanilla, MultiCLS, enhanced/head-fix, TokenMoE, TokenMoE+SSL, Transformer+TabM-mini) at fractions f ∈ {0.05…1.00}, five seeds each, scored by held-out Poisson deviance. Four claims: (1) smooth power-law data scaling exists in this noisy tabular setting; (2) TabM-style models scale best (α≈0.309 main sweep; ≈0.409 with full-data TabM-mini extensions); (3) supervised Transformer parameter scaling is weak (β≈0.025–0.037) and is rescued only by TabM-style adaptation or swap-style SSL (β≈0.055; TabM β≈0.148); (4) exponents yield compute-optimal guidance (β/α≈0.36 for TabM vs ≈0.12 for Transformers/FFNs). Section 1.1 recasts Bühlmann–Straub credibility, Z(n)=n/(n+k), as a scaling law with the same distance-to-asymptote algebra — a genuinely useful framing device.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the paper find about scaling laws in insurance ratemaking?
On a ~4.5M-row motor portfolio, held-out Poisson deviance improves smoothly as a power law with training-data size for every tested model family. TabM/TabM-mini scale fastest (data exponent α≈0.309, or ≈0.409 with full-data extensions), while supervised Transformers improve with data but barely benefit from more parameters (β≈0.025) unless self-supervision (β≈0.055) or TabM-style ensembling is added.
Do neural networks beat GLMs for insurance pricing in this study?
Only modestly, and only with a lot of data. At the full ~4M training rows, the best TabM-mini variant beats the GLM by 0.00315 mean Poisson deviance (~1.1% relative, about 0.0016 nats of log-likelihood per policy). At the smallest fractions (5–25%), GLMs and Transformer+TabM hybrids lead, and the paper advises simple, stable models when labelled claims data are scarce.
Is the data or code for this paper public?
The main motor portfolio is proprietary; only aggregated summaries are reported. The author released a supplementary open-source workflow on the public French MTPL (freMTPL2) dataset at github.com/RonRichman/frmtpl-scaling-laws, which reproduces the pipeline but not the proprietary headline numbers.
Why do Transformers underperform on tabular insurance data in this paper?
The author argues the supervised signal is too weak (sparse claim counts, low signal-to-noise) to convert added Transformer capacity into accuracy: the parameter exponent is β≈0.025–0.037 vanilla versus 0.148 for TabM. What helped in this experiment: multiple CLS pooling, value-pathway capacity (head-fix), token-level mixture-of-experts routing, and a swap-style self-supervised auxiliary task.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with marked [...sampling gap...] omissions (e.g., parts of the Section 4 data description, portions of Sections 5.1 and 7.4–7.5, Appendix configuration tables, and references); conclusions above rest on the included sections and in-text summaries of the omitted ones.
- Figures appear only as extracted axis labels and captions; visual curve shapes (e.g., Figures 4, 7–11) could not be inspected, so curve-level judgments rely on the reported tables and text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.