Measuring Intra-Daily Market Risk: A Neural Network Approach

The value at risk (VaR) measure often relies on an assumption about the return (or price) distribution of the underlying risky assets. Different distributional assumptions may produce widely different computed VaR values. When estimating VaR using intra-daily equity returns, the question arises as to what assumption should be made about the return distribution. Because of the difficulty of decomposing trading noise, it is very hard to identify the return distribution at the tick-by-tick level. In this paper, we circumvent the difficulty of making a distributional assumption of intra-daily market fluctuations by specifying a neural network approach. With this approach, no distributional assumption regarding the return distribution is required for estimating and forecasting the VaR using intra-daily data. Using this approach, we forecast VaR using high-frequency data for the German equity market index. Our neural network forecasts, evaluated on the basis of several statistical performance measures and compared with alternative time-series models, suggest that the performance of the neural network approach in VaR computation dominates that of the commonly used time-series models.