Évf. 2026 szám 6: Neural networks in developing and validating Expected Shortfall estimation models

Máté Uzsoki, Gyöngyi Bugár and Balázs Kovács

The new Basel III standards establish a revised framework for determining the capital charge for market
risk in internal models, transitioning from Value-at-Risk (VaR) to Expected Shortfall (ES). The latter is
a risk measure that more effectively captures tail risk and exhibits several advantageous properties, such
as coherence. Numerous models have been proposed in the literature for estimating ES, ranging from
historical simulation and parametric statistical models to quantile regression and Generalized
Autoregressive Conditional Heteroscedasticity (GARCH) methodology. Our goal is to develop a multilevel
adaptive neural network which is capable to provide significantly improved ES estimates compared
to a benchmark model, selected for its previous superior prediction performance compared to other
models.

Keywords: Risk management, Expected shortfall, Modelling market risk, Artificial intelligence