Predicting Exchange Rate Value-at-Risk and Expected Shortfall: A Neural Network Approach

Abstract: On the basis of the recommendation of the Basel Committee on Banking Supervision to transition from Value-at-Risk (VaR) to Expected Shortfall (ES) in determining market risk capital, this paper attempts to investigate whether a Recurrent Neural Network provides more accurate VaR and ES predictions of the EUR/USD exchange rate compared to the conventional GARCH(1,1) model. A number of previous studies has confirmed the forecasting ability of a plain vanilla Feedforward Neural Network over traditional statistical models. However, standard neural networks have limitations. Most notably, they rely on the assumption of independency among data observations, which presents a problem when data points are related in time. To circumvent this restriction, this study employs a Gated Recurrent Unit type of neural network to produce one-step-ahead volatility forecasts of the EUR/USD exchange rate, which are then used to compute VaR and ES predictions. The VaR and ES forecasts for both models are obtained through a Volatility Weighted Historical Simulation, and evaluated with backtesting procedures. The empirical results indicate that the GARCH(1,1) model outperforms the Gated Recurrent Unit neural network for VaR95%, while the Gated Recurrent Unit neural network appears more adequate in forecasting ES at a 95% confidence level.