Johdanto

In this work, the aim is to study selected machine learning methods and apply them in practice to a selected actuarial problem, namely the prediction of life insurance lapse risk. The selected machine learning methods were gradient boosting and neural networks.

First, a theoretical background for machine learning and the selected methods is presented. Second, the determinants of life insurance lapse risk and previous research on the topic are discussed. Finally, the process for applying the selected machine learning methods to life insurance lapse prediction is described and the modeling results are presented.

The results imply that the prediction of life insurance lapse can be improved by using machine learning. Both of the selected machine learning methods improved the lapse prediction results compared to more traditional statistical methods. Gradient boosting (XGBoost) performed slightly better than the neural network model and, especially taking into account the time efficiency of the fitting process and interpretability of the model, the XGBoost model was concluded to be better suited for lapse prediction.

The usefulness of the machine learning models in lapse prediction was also studied by using an economic metric for the lapse management. The result in terms of the economic metric was improved when the model was fit based on regression for the economic gain instead of using a pure classification. The actual economic gain, however, depends on how the parameters for the economic model are determined. Depending on the parameter selection, the resulting economic gain could remain limited or be quite significant. Therefore, the benefits of the lapse prediction depend on how efficient the insurance company is in their lapse management and on the overall profitability of the insurance product.

All in all, the machine learning methods are concluded to be useful in lapse prediction as well as other actuarial applications and their significance is also likely to increase in the future.

Share Share Share