This paper explores the application of advanced machine learning (ML) techniques, particularly gradient-boosted trees (GBTs), multilayer perceptron deep learning (MLP DL), and transformer-based deep learning (TRF DL), for the pricing of multiperil insurance policies—aiming to address the potential dependencies among different covers and between frequency and severity of claims. The predictive performance of these techniques is compared on three distinct actuarial datasets. While all methods provide reasonable predictive performance, the TRF consistently outperforms the others, demonstrating superior ability to capture interdependencies among perils and risk components, although it is computationally more complex. The paper also investigates model interpretation and variable importance, emphasizing practical considerations for actuarial applications. This paper was supported by a 2023 individual grant from the CAS and the Society of Actuaries.
Spedicato et al. (Mon,) studied this question.