Generalizability of Scaling Laws Across Actuarial Tasks

Determine whether scaling behavior analogous to that observed for large-language-models generalizes to noisy, heterogeneous, regulated actuarial tasks beyond the single motor-insurance frequency dataset, including different lines of business, targets, feature sets, jurisdictions, and validation regimes.

Background

The paper establishes empirical power-law relationships for Poisson claim-frequency prediction using one large anonymized motor-insurance portfolio and an IID train/test split. The authors explicitly caution that the fitted exponents and model rankings are conditional on this experimental regime. They identify temporal nonstationarity, alternative lines of business, different feature-engineering choices, and other actuarial targets as settings in which the observed scaling behavior may not persist.

The unresolved issue is whether the apparent scaling regularities are portfolio-specific empirical phenomena or robust properties of actuarial tabular learning more broadly. Resolving it would require experiments across additional datasets, targets, validation designs, and model families.