Risk management in life insurance requires proper modelling, measuring, and managing of the key risk drivers of life and pension business. This includes lapse risk since the termination of a life insurance contract prior to maturity has a significant impact on the cash flow profile and the profitability of life insurance business. To reflect this risk, market consistent valuations are based on best estimate future lapse rates and the Solvency II standard formula assesses lapse risk in a specific risk module.

For the derivation of best estimate lapse rates, the insurance portfolio is typically divided into sub-portfolios based on contract characteristics like type of contract, country, or distribution channel. Lapse rates are then derived for each sub-portfolio independently, considering the dependency on factors like time since inception, contract duration or policyholder age. However, ignoring dependencies between these sub-portfolios can lead to inaccurate cash flow projections.

To address this, multivariate lapse models have been developed to model lapse rates on the individual contract level using all available covariates simultaneously. If set up properly, these models can take dependencies between sub-portfolios into account and provide more reliable estimates. However, the specification of a sophisticated model is associated with a considerable effort. In particular, the increasing number of potential covariates requires a thorough analysis and models are still prone to over- or underfitting.

The application of data science methods can replace this largely manual process by a more automated process. In this paper, we use the LASSO variable selection method which is based on a multivariate model and can identify patterns within the data automatically. In order to identify hidden structures, we use recently developed extended versions of the LASSO algorithm that allow different sub-penalties for individual covariates. We show how these sub-penalties can be combined to satisfy the needs of a practical application, in particular with respect to goodness of fit and computation time.

In contrast to random forests or neural networks, the predictions of our lapse model remain fully interpretable and explainable. The advantages of the method are illustrated based on data from a European life insurer operating in four countries. We show how structures can be identified efficiently and fed into a highly competitive, automatically calibrated lapse model.