Date/Time
Date(s) - 03/10/2023
3:30 pm - 4:30 pm

Location: University Hall (UH) 112

Speaker(s): Pengfei Cai and Pratheepa Jeganathan  

Title: Recurrent Neural Networks Methods for Modeling Dependence between Loss Triangles

Abstract: Reserves are the actuary’s best estimate for future unpaid claims in a P&C company. The reserves for different lines of business (LOB) are related, and there are only a few tools developed for multivariate loss reserving using recurrent neural networks (RNN). The goal is to develop a predictive distribution for reserves using machine learning. The solution is the extended Deep Triangle (EDT) model, which uses incremental paid loss sequence from two LOBs as the input/output and the symmetric squared loss of two LOBs as the loss function and Generative Adversarial Network (GAN). With the obtained predictive distribution of loss reserve, we found that risk capitals calculated from EDT combined with copula GAN are smaller than copula regression models, implying a larger diversification benefit. To illustrate our method, we apply and calibrate EDT-GAN on personal and commercial automobile lines from a major US P&C insurance company. Finally, these findings are also confirmed in a simulation study.  

(refreshments at 3 in HH 216 – networking opportunity for the seminar attendees)