Boya Chen

Title: Case Study: Interpretable Hybrid Models for Pure Premium Ratemaking in Automobile Insurance
Date: Thursday, August 6th, 2026
Time: 3:00pm
Location: Zoom
Supervised by: Himchan Jeong

Abstract:: Using French automobile insurance data, this case study examines a validation-gated hybrid workflow for pure premium ratemaking that extends a Tweedie generalized linear model (GLM) through a double generalized linear model (DGLM), neural-network-based reduction of high-cardinality categorical variables, and a combined actuarial neural network (CANN) correction. Three placements of the reduction step are compared using cross-validation. Approach 2, which performs reduction after the baseline DGLM, is retained as a competitive parent model because it balances validation performance, DGLM consistency, and complexity. On the holdout set, CANN reduces weighted Tweedie deviance from 31.992 to 31.484 and increases weighted Gini from 0.226 to 0.354, with smaller improvements in error metrics. Bootstrap out-of-bag analysis nevertheless shows greater variability for CANN. The findings support neural components as controlled, validation-gated extensions to an interpretable DGLM rather than automatic replacements for actuarial tariff models.