Mining actuarial risk predictors in accident descriptions using recurrent neural networks

One crucial task of actuaries is to structure data so that observed events are explained by their inherent risk factors. They are proficient at generalizing important elements to obtain useful forecasts. Although this expertise is beneficial when paired with conventional statistical models, it becomes limited when faced with massive unstructured datasets. Moreover, it does not take profit from the representation capabilities of recent machine learning algorithms. In this paper, we present an approach to automatically extract textual features from a large corpus that departs from the traditional actuarial approach. We design a neural architecture that can be trained to predict a phenomenon using words represented as dense embeddings. We then extract features identified as important by the model to assess the relationship between the words and the phenomenon. The technique is illustrated through a case study that estimates the number of cars involved in an accident using the accident's description as input to a Poisson regression model. We show that our technique yields models that are more performing and interpretable than some usual actuarial data mining baseline.