Explainable Auto Insurance Risk Prediction via Exposure-Aware Feature Interaction and Cross-Fitted Residual Learning

Accurate prediction of motor insurance claim frequency must jointly address nonlinear interactions, heterogeneous exposure periods and transparent risk attribution. Building on the Transformer–SHAP framework of Sun et al. for interpretable auto-insurance risk evaluation, this paper develops exposure-aware interaction and cross-fitted residual learning (EICRL). A regularized LightGBM model estimates annual claim frequency under exposure weights, and an independently fitted Poisson offset model corrects out-of-fold log-frequency errors with sparse, interpretable interaction terms; validation-controlled shrinkage and a positive exponential link yield stable, nonnegative expected claim counts. On a stochastic motor-portfolio simulator with exposure heterogeneity, rare counts and nonlinear actuarial interactions, EICRL lowers held-out Poisson deviance relative to LightGBM in all 15 independent portfolios of 48,000 policies (0.39247 versus 0.39461; paired 95% interval of the reduction [0.00150, 0.00278]). Ablations separate the contributions of interaction modeling and cross-fitting, and whole-pipeline permutation analysis identifies bonus–malus and driver age as the dominant predictors. EICRL thus provides an exposure-consistent and explainable frequency model ready for validation on observed portfolios such as freMTPL2freq.