Actuarial Science and Financial Mathematics seminar series
Anas Abdallah
McMaster University
Room: M3 3127
From Neural Networks to SUR Copula Mixed Models: New Directions in Multivariate Loss Reserving and Risk Capital Analysis
In the property and casualty (P&C) insurance industry, reserves represent the largest component of liabilities, and accounting for dependencies across multiple lines of business (LOBs) is critical for accurate reserve estimation and risk capital determination. This presentation highlights recent developments in dependence modeling of loss reserves, integrating machine learning and modern statistical parametric approaches.
First, I introduce the Extended Deep Triangle (EDT), a recurrent neural network framework coupled with generative adversarial networks, which leverages multicompany data to predict multivariate reserves and generate predictive reserve distributions. The EDT achieves superior predictive accuracy and highlights the potential diversification benefits in risk capital.
Second, I present Seemingly Unrelated Regression (SUR) copula mixed models, which extend copula regression by incorporating random effects to capture company-specific heterogeneity, adopting flexible marginal distributions across LOBs, and using shrinkage techniques to improve stability. These models preserve the interpretability of the dependence structure while enhancing both robustness and predictive performance.
Finally, I outline ongoing work on hybrid frameworks that combine the predictive power of neural networks with the structural interpretability of SUR copula mixed models. Through real data applications and simulation studies, I show how these approaches provide new tools to improve reserve accuracy, quantify diversification, and support capital management.