Blurring the boundaries: loss cost model test

Since its early days, insurance has operated on a foundation of risk prediction. Personal computers and the data age have made it crucial to use statistical modelling techniques to understand the complex relationship between risk factors and the potential loss cost associated with an insurance policy. In this context, generalised linear models (GLMs) have become indispensable, capable of predicting a variety of outputs, such as claims frequency and severity.

GLMs are still relevant in insurance but the new risk factors brought about by automation and new technologies add complexity to their construction. This is particularly relevant in motor insurance, where advanced driver-assistance systems and self-driving features are on the rise.

Machine learning (ML) models, on the other hand, have proven useful in capturing intricate relationships between large numbers of features and dependent variables. Gradient boosting models (GBMs) are gaining prominence in pricing, and there have been initiatives to explore ML techniques further.

Here, we explore four modelling techniques used to predict the loss cost, comparing GLMs’ predictive accuracy with that of GBMs, artificial neural networks (ANNs), and a hybrid model combining a GLM with an ANN.