Expected shortfall model based on a neural network

Abstract

Considering both the limitations of traditional models of value-at-risk and expected shortfall (ES) for risk estimation in the context of the Basel standards and the possibilities of applying neural network models for risk estimation purposes, our paper presents a new ES model or, more specifically, an ES-extreme-value-theory (ES-EVT) model improvement, as it is a combination of the standard multilayer perceptron model and the ES model based on EVT. This model exploits the advantages of both approaches in estimating financial risk. The model was tested on 15 example indexes of emerging European capital markets. The model quality assessment against the ES-EVT model used mean squared error, while model validation in the context of the Basel III standards was done using Berkowitz’s ES backtesting, based on bootstrap simulation, and Acerbi and Szekely’s first method. The results obtained imply that our neural network application improves ES-EVT model performance.