Detection of Interacting Variables for Generalized Linear Models Using Neural Networks

The quality of generalized linear models (GLMs), which are frequently used by insurance companies, depends on the choice of interacting variables. The search for interactions is time-consuming, especially for data sets with a large number of variables, depends much on the expert judgement of actuaries, and often relies on performance indicators that require visual evaluation. Therefore, we present an algorithm that detects and recommends the next-best interaction that is missing in a GLM. This recommended interaction has a high chance of improving the predictive power of a GLM. Our approach relies on neural networks and a model-specific interaction detection method, which is computationally faster than traditionally used methods, such as Friedman’s H-Statistic or SHAP values. In our talk, we describe three main parts of the proposed interaction-detection approach, show its performance on an artificial data set as well as an open-source motor third-party liability insurance data set, and briefly comment on its performance on a big proprietary data set with over 10 million observations.

Our presentation is based on the following paper: Havrylenko, Y., and Heger, J. Detection of interacting variables for generalized linear models via neural networks. European Actuarial Journal (2023). https://doi.org/10.1007/s13385-023-00362-4