Akademik Veri Yönetim Sistemi
VII. International Applied Statistics Congress (UYIK-2026) , İstanbul, Türkiye, 11 - 13 Mayıs 2026, ss.1, (Özet Bildiri)
One of
the primary objectives of the insurance industry is to accurately estimate
premiums based on policyholders’ risk profiles. Generalized Linear Models
(GLMs) have long been the standard approach in actuarial pricing due to their
interpretability and solid statistical foundation. However, they may exhibit
limitations in capturing complex nonlinear relationships and interactions
present in large-scale insurance datasets. In this study, several advanced
machine learning and deep learning models are evaluated as potential
alternatives to GLMs using the French Motor Third Party Liability (MTPL)
dataset. The models considered include Gradient Boosting Machines (XGBoost),
Combined Actuarial Neural Networks (CANN), and Distributional Refinement Network
(DRN). To ensure robust model evaluation and mitigate overfitting, a 5-fold
cross-validation strategy is employed. In addition, hyperparameter tuning is
conducted using the Optuna optimization framework to obtain model-specific
optimal configurations. The premium estimation process is decomposed into claim
frequency and claim severity components. For frequency modeling, the CANN model
achieves the lowest deviance and highest Gini coefficient, indicating superior
predictive performance. For severity modeling, the DRN model outperforms other
approaches, particularly in capturing heavy-tailed distributions, which are
critical in insurance risk modeling. The findings suggest that hybrid approaches
combining statistical rigor with neural network flexibility can provide a
competitive and robust alternative to traditional GLM-based pricing models.