A Guided Neural Network Approach to Volatility Forecasting

We introduce a novel neural network approach to volatility forecasting using realised measures and daily returns, and possibly other measurements of volatility. Our method combines forecasts from benchmark parametric econometric models with a feedforward neural network. We find in an empirical forecasting exercise that for realised volatility measures, the benchmark parametric models are hard to beat, except at the longest horizon under consideration. For squared returns at moderate horizons, we find some gains in forecasting performance. We find in an economic application to Value-at-Risk and expected shortfall forecasting that the best parametric and neural network models can perform similarly, but that the quality of neural network forecasts depends heavily on tuning; “Bad" parametric forecasts can substantially impact the forecasts from our new approach.