Advancing loss reserving: A hybrid neural network approach for individual claim development prediction

Accurately estimating loss reserves is critical for the financial health of insurance companies and informs numerous operational decisions. We propose a novel neural network architecture that enhances the prediction of incurred loss amounts for reported but not settled claims. Moreover, differing from other studies, we test our model on proprietary datasets from a large industrial insurer. In addition, we use bootstrapping to evaluate the stability and reliability of the predictions and Shapley additive explanation values to provide transparency and explainability by quantifying the contribution of each feature to the predictions. Our model shows superiority in estimating reserves more accurately than benchmark models, like the chain ladder approach. Particularly, our model exhibits nuanced performance at the branch level, reflecting its capacity to effectively integrate individual claim characteristics. Our findings emphasize the potential of using machine learning in enhancing actuarial forecasting and suggest a shift towards more granular data applications.