Poisson Mixture Deep Learning Neural Network Models for the Prediction of Drivers’ Claims with Excessive Zero Claims Using Telematics Data

This article explores the role of neural networks in insurance claim prediction using Poisson mixture (PM) deep learning neural networks. The model is designed to handle drivers’ insurance claims with excessive zero occurrences by setting a prior probability of a safe group (low claim). The article evaluates PM networks using the negative log-likelihood (NLL) loss function instead of the common choice of mean square error loss function applicable for symmetric data. The NLL loss function captures the asymmetric distribution of claim counts with excessive zeros. The meticulous search for network architecture, employing both manual and Bayesian optimization search techniques, further improves the prediction accuracy of PM density networks. The commitment of this research to pushing the boundaries of predictive analytics is evident throughout the modeling, architecture selection, and evaluation process, positioning the PM deep learning neural network as a noteworthy advancement in insurance claim prediction.