Actualités

Do We Need Intra-Daily Data to Forecast Daily Volatility? (version préliminaire)

Mercredi | 2012-05-16
B103

Denisa BANULESCU-RADU – Bertrand Candelon – Christophe HURLIN

Considering mixed data sampling (MIDAS) regressions, we analyze the inuenceof the sampling frequency of intra-daily predictors on the accuracy of the volatility forecasts. We propose various in-sample and out-of-sample comparisons of daily,weekly and bi-weekly volatility forecasts issued from MIDAS regressions based on intra-daily regressors sampled at di erent frequencies. First, we show that increasing the frequency of the regressors improves the forecasting abilities of the MIDAS model. In other words, using regressors sampled at 5 minutes gives more accurate forecasts than using regressors sampled at 10 minutes, etc. These results are robust to the choice of the loss function (MSE, Qlike, etc.) and to the choice of the forecasting horizon. Third, the MIDAS regressions with high-frequency regressors (sampled between 5 minutes and 30 minutes) provide more accurate in-sample forecasts than a GARCH model based on daily data. However, except the one-period-ahead forecasts of the calm period, the out-of-sample forecasts of MIDAS models are not signi cantly di erent from the GARCH forecasts, whatever the sampling frequency used, con rming that the direct use of high-frequency data does not necessarily improve volatility predictions (version préliminaire)