A machine learning model for lapse prediction in life insurance contracts

We use the Random Forest methodology to predict the lapse decision of life insurance contracts by policyholders. The methodology outperforms the logistic model, even if features interactions are considered. We use global and local interpretability tools to investigate how the model works. We show that non-economic features (the time passed from the incipit of the contract and the time to expiry, as well as the insurance company and its commercial approach) play a significant effect in determining the lapse decision while economic/financial features (except the disposable income growth rate) play a limited effect. The analysis shows that linear models, such as the logistic model, are not adequate to capture the heterogeneity of financial decisions.