Keep it simple with realized volatility forecasts

 

One of the key risk management problems is forecasting volatility. A key issue for any option trader is forecasting volatility. You cannot do either without good volatility forecasts. So how do you get good forecasts? The Machine Learning (ML) crowd will say that you should use the latest non-linear techniques to improve forecasts. This view assumes that more complexity is better than a simple model, yet this is an assumption that should be testable. 

In the paper, "Forecasting realized volatility: Does anything beat linear models?", the authors compare different tests on the quality of linear and non-linear ML models for forecasting realized volatility. They conclude that heterogeneous autoregressive (HAR) models should remain as the workhorse for forecasting volatility.  The HAR models generate a volatility forecast using past volatility across different horizons. The ML techniques include neural networks as well as tree-based methods. 

They find that adding predictors will improve the out-of-sample forecasts for short-term forecasts, but there is no evidence that ML models can outperform the linear models. The models are tested against MSE, one-day ahead VaR, and realized utility. Simple works well and forecasting realized volatility does not need the added work from ML procedures.