Combining nearest neighbor predictions and model-based predictions of realized variance: Does it pay?

The increasing availability of intraday financial data has led to improvements in daily volatility forecasting through the use of long-memory models of realized volatility. This paper demonstrates the merit of the non-parametric nearest neighbor (NN) approach for S&P 100 realized variance forecasting. The NN approach is appealing a priori because, unlike model-based methods, it can reproduce complex dynamic dependencies, while largely avoiding misspecification and parameter estimation uncertainty. We evaluate the forecasts through straddle trading profitability metrics and using conventional statistical accuracy criteria. The ranking of individual forecasts confirms that there is not a one-to-one mapping between statistical accuracy and profitability. In turbulent markets, the NN forecasts lead to higher risk-adjusted profitability levels, even though the model-based forecasts are superior statistically. A directional combination of NN and model-based forecasts is more profitable than any of the individual forecasts, in both calm and turbulent market conditions.