Does Higher-Frequency Data Always Help to Predict Longer-Horizon Volatility?

When it comes to forecasting long-horizon volatility, multistep-ahead iterated forecasts using higher-frequency data can be more efficient than one-step-ahead direct forecasts using lower-frequency data. However, small violations of model specification in either the volatility or expected return models are compounded in the forward iteration and temporal aggregation for the higher-frequency model. In this paper, we show that realized conditional autocorrelation in return residuals is a strong predictor of the relative performance of different frequency models of volatility. When the conditional autocorrelation is high, the higher-frequency model performs markedly worse than its lower-frequency counterpart. Empirically, we show that residual autocorrelation exists in the broad cross-section of stocks at any given point in time, and that this misspecification can substantially decrease the prediction performance of higher-frequency models. Comparing the monthly volatility predictions using daily and monthly data, we show a trade-off between the gains from higher-frequency data and the susceptibility of its multistep-ahead iterated forecasts to model misspecification.