Comparing the estimation of Value at Risk and Expected Shortfall with LSTM and EGARCH family members

This paper aims to compare the performance of traditional GARCH-type models and an LSTM-based approach for forecasting Value at Risk (VaR) and Expected Shortfall (ES) under different symmetric and skewed distributions. To assess model performance, eight stock indices from diverse international markets are analyzed. The models are evaluated using three backtesting approaches and a model selection criterion, the Weighted Absolute Deviation (WAD). The results indicate that the selected indices exhibit heavy tails and asymmetry. In general, the results obtained under skewed distributions generally outperform those obtained under symmetric distributions. In most cases, the LSTM model is selected as the top performing model. However, some models from the EGARCH family remain strong competitors, especially under the asymmetry distributions, and might be preferred for certain indices.