Forecasting banks’ stock realized volatilities with mixed-frequency data

This research compares the Mixed Data Sampling (MIDAS) approach, where the regressand is measured on a lower frequency than the regressor, with an AR(1) model where both of the variables are measured at the same frequency. In the first half of the research some of the results from Clements and Galvão [2009] are replicated. In their paper the researchers forecast quarterly output growth using 10 or 5 macroeconomic leading indicators. For the second part the models are used to make 1- and 2-step ahead forecast of the weekly realized volatility of four banks’ stocks for two different time periods. Two variants of the Multivariate MIDAS (M-MIDAS) models are considered: with the Almon Lag polynomial or the Step function as the weighting scheme. Furthermore, a calm and a turmoil period for forecasting are considered. For the replicate part I found that the M-MIDAS is more accurate in all cases compared to the AR(1) model except for the replicate part when 5 leading indicators are used. In the second part the M-MIDAS model was more accurate in most of the cases for the calm period, for the turmoil period there was no difference in accuracy compared to the AR(1) model.