2023 Australasian Actuarial Education and Research Symposium

Eric Dong

University of New South Wales

Distributional forecasting via an interpretable actuarial neural network

This is joint work with Benjamin Avanzi, Patrick Laub, Bernard Wong

Neural networks have recently seen extensive developments in the actuarial field (Richman, 2022). Adapting neural network technology to actuarial science can be difficult as we place high importance on interpretability and distributional forecasting. Existing approaches include the Combined Actuarial Neural Network (Schelldorfer and Wuthrich, 2019) and the Mixture Density Network (Delong et al., 2021, Al-Mudafer et al., 2022). However, these architectures rely on prespecified distributional assumptions. These assumptions may not be appropriate and can impact the ability to quantify the variability of outcomes accurately and undermine the reliability of the distributional forecasts. This paper presents a new neural network architecture that allows for distributional flexibility and a level of interpretability; these are typically two conflicting objectives. The predictive performance and forecasting interpretability of our architecture are demonstrated on both synthetic and real actuarial datasets.