Neural Networks in Developing and Validating ES-Estimation Models

The new Basel III standards provide a revised framework for determining the capital charge for market risk in internal models, with a shift from Value-at-Risk (VaR) to Expected Shortfall (ES). The latter one is a risk measure for better capturing tail risk which has some more favourable properties such as coherence. There are plenty of models proposed by the literature for ES estimation from historical simulation or parametric statistical models to quantile regression or GARCH methodology. Our aim is to conduct a large-scale empirical research in order to compare and contrast the back-testing performance of about 50 models. Building on the joint elicitability property of ES with VaR, we can rely on the so-called ridge back-test proposed by Acerbi and Székely recently. This valuable contribution can be utilised in model selection (relative validation), namely in ranking competitive models based on their forecasting performance. It is intended to find out whether some less well-known statistical models can outperform their classical counterparts. As a main contribution, we propose to develop a multi-level adaptive neural network which is capable to give significantly better ES-estimates than other models.

Find the Q&A here: Q&A on 'Data Driven ERM'