Dispersion modeling in Tweedie compound Poisson with combined actuarial neural networks

Accurate premium calculations in actuarial modeling are critical to the sustainability of insurance companies. Insurance portfolios are characterized by a nonnegative continuous distribution with a point mass at zero and considerable heterogeneity across policyholders. In this context, the Tweedie’s compound Poisson (CP) distribution is widely used in insurance data. However, classical statistical approaches often struggle to capture heterogeneity and complex non-linear relationships in insurance portfolio. This study proposes a new approach called “double combined actuarial neural networks (DCANN)” to address this problem. This model extends the combined actuarial neural network (CANN) approach in the literature by applying the double GLM framework. It simultaneously models both mean and dispersion parameters, combining classical regression with neural networks. In this study, pure premium estimates are obtained using auto insurance data under homogeneous and heterogeneous dispersion assumption. Under the homogeneous dispersion assumption, classical generalized linear model (GLM), neural network models, and CANN are applied. For the heterogeneous dispersion case, estimates are produced using DCANN—an extended version of the CANN model—along with Double GLM and double neural networks. As a result, evaluation metrics showed that the DCANN model outperformed the alternative approaches in data structures with high zero clustering.