Scaling Laws, Tabular Data and Actuarial Ratemaking Models

Battle Count: 60

The paper is primarily focused on actuarial insurance pricing (motor insurance claim frequency), not quantitative trading. However, the scaling-law methodology, power-law fitting, compute-optimal allocation, and tabular deep learning techniques (Transformers, MoE, ensembles) have methodological transferability to financial prediction tasks. The Poisson deviance framework and credibility theory connections are actuarial-specific. The general insight that model family matters as much as resources being scaled is broadly applicable to any tabular prediction problem in finance.

The paper involves training 10+ model families with multiple size configurations across 6 training fractions with 5 seeds each, requiring significant compute (A100, GH200, B200 GPUs). Novel Transformer modifications (MultiCLS, TokenMoE routing, layer-wise value embeddings, swap-style SSL, head-fix mechanisms) add architectural complexity. The scaling-law fitting pipeline (Pareto frontier extraction, power-law regression in log space, bootstrap uncertainty) requires careful implementation. However, the core methodology (train models, measure deviance, fit power laws) is conceptually straightforward. The supplementary GitHub repo provides a simpler reproduction workflow.

The main experiment uses proprietary motor insurance data (~4.5M rows) that cannot be shared. However, a supplementary open-source implementation on public French MTPL (FMTPL) frequency data is provided via GitHub (https://github.com/RonRichman/frmtpl-scaling-laws). The experimental protocol, hyperparameters, and model configurations are fully documented in appendices. The scaling-law fitting methodology is standard and well-described. Reproduction of the main results requires access to the contributing insurer's data.