Temporal dynamics in tree-based models and applications to lapse behaviour in life insurance

This thesis, « Temporal dynamics in tree-based models and applications to lapse behaviour in life insurance », delves into the application of novel statistical models and strategies to address lapse behavior in life insurance. This work is an amalgamation of two introductive parts and three research articles, each offering a unique perspective and various contributions to the understanding of lapse management and the inclusion of a time dimension in tree-based Machine Learning models. The first article, « Including individual Customer Lifetime Value and competing risks in tree-based lapse management strategies », presents a retention strategy that goes beyond a mere prediction of lapse. This strategy is grounded on a lapse management framework that integrates Customer Lifetime Value and profitability, with a focus on the individualization of existing approaches. The study demonstrates how survival tree-based models outperform parametric approaches, thereby leading to more efficient and informed management of retention campaigns, for life insurers. Building on this work, the second article, « A longitudinal framework for lapse management in life insurance », emphasizes the importance of a time-informed approach in lapse management. The article proposes a longitudinal lapse management strategy framework that leverages the complete past trajectory of policyholders, a resource often overlooked yet abundantly available in insurers' information systems. This methodology further refines the targeting precision, thereby enhancing the insurers' understanding of their global portfolio risk. The final article, « Time penalized tree (TpT): a new tree-based data mining algorithm for time-varying covariates », introduces a novel decision tree algorithm that accounts for time-varying covariates within longitudinally structured datasets. The article challenges the traditional static assumption of covariates by proposing an algorithm that allows recursive partitioning of the covariates space together with time. This innovative method helps to capture relevant historical trends for analysis, enabling a more accurate and interpretable study of phenomena evolving in dynamic environments.

Overall, this thesis provides a global approach to lapse behavior analysis in life insurance by integrating advanced tree-based models and longitudinal analysis. It underscores the potential of these innovative strategies in informing commercial and strategic decisions in the insurance industry while ensuring their interpretability and explainability.