A1134
Title: The value-at-risk of a large insurance portfolio: Accounting for model risk
Authors: Hanieh Amjadian - Concordia University (Canada) [presenting]
Yang Lu - Concordia University (Canada)
Yunran Wei - Carleton University - School of Mathematics and Statistics (Canada)
Abstract: Recently, quantifying model risk on risk measures such as value-at-risk (VaR) for portfolios has been extensively studied in the context of credit risk, as well as in market risk, particularly within a time series framework. However, relatively few frameworks address model risk in a cross-sectional setting. One notable exception is a recent study, which explicitly considers model risk concerning the aggregation of individual risks. The purpose is to adopt their methodology in quantifying model risk when calculating risk measures, such as VaR or ES of a large insurance portfolio, and address their limitations by introducing an additional calibration dataset. Simulation studies are conducted using heavy-tailed data to evaluate the performance of our proposed method. The results show that relying solely on training data can lead to significant underestimation of risk, particularly when models are overfitted or misspecified. By incorporating a calibration dataset and using a bootstrap procedure, the approach produces more accurate and robust VaR estimates, especially in the tails of the distribution. It is also investigated how model complexity, the heaviness of the tails, and the number of bootstrap replicates affect the quality of the estimates, offering practical guidance for risk assessment in insurance portfolios.
