The information flow in modern financial markets is continuous, but major stock exchanges are open for trading for only a limited number of hours. No consensus has yet emerged on how to deal with overnight returns when calculating and forecasting realized volatility in markets where trading does not take place 24 hours a day. Based on a recently introduced formal testing procedure, we find that for the S&P 500 index, a realized volatility estimator that optimally incorporates overnight information is more accurate in-sample. In contrast, estimators that do not incorporate overnight information are more accurate for individual stocks. We also show that accounting for overnight returns may affect the conclusions drawn in an out-of-sample horserace of forecasting models. Finally, there is considerably less variation in the selection of the best out-of-sample forecasting model when only the most accurate in-sample RV estimators are considered.
