Deep learning in insurance: Accuracy and model interpretability using TabNet

Generalized Linear Models (GLMs) and XGBoost are widely used in insurance risk pricing and claims prediction, with GLMs dominant in the insurance industry. The increasing prevalence of connected car data usage in insurance requires highly accurate and interpretable models. Deep learning (DL) models have outperformed traditional Machine Learning (ML) models in multiple domains; despite this, they are underutilized in insurance risk pricing. This study introduces an alternative DL architecture, TabNet, suitable for insurance telematics datasets and claim prediction. This approach compares the TabNet DL model against XGBoost and Logistic Regression on the task of claim prediction on a synthetic telematics dataset. TabNet outperformed these models, providing highly interpretable results and capturing the sparsity of the claims data with high accuracy. However, TabNet requires considerable running time and effort in hyperparameter tuning to achieve these results. Despite these limitations, TabNet provides better pricing models for interpretable models in insurance when compared to XGBoost and Logistic Regression models.