Improving Insurance Catastrophic Data with Resampling and GAN Methods
Norbert Dzadz, Maciej Romaniuk
TL;DR
This work addresses the scarcity and extreme-value nature of catastrophic insurance data by comparing bootstrap, bootknife, and GAN-based resampling to augment both the quantity and quality of observations. It combines a risk-process framework with empirical evaluation on MSE and MAE, and adds a fuzzy-number extension to provide an expert-like, imprecise assessment of the predicted claim process $S_T$, leveraging $oldsymbol{S}_{T, ext{alpha}}$ intervals. The study finds that GAN-based generation often yields the lowest predictive errors, especially when paired with a Power Law intensity function and Weibull claim values, and shows that bootstrap methods offer modest gains. Practically, the findings inform insurers about which data-enhancement technique best stabilizes risk estimates and how to incorporate uncertainty via fuzzy outputs, potentially affecting pricing, reserves, and risk management decisions.
Abstract
The precise and large dataset concerning catastrophic events is very important for insurers. To improve the quality of such data three methods based on the bootstrap, bootknife, and GAN algorithms are proposed. Using numerical experiments and real-life data, simulated outputs for these approaches are compared based on the mean squared (MSE) and mean absolute errors (MAE). Then, a direct algorithm to construct a fuzzy expert's opinion concerning such outputs is also considered.