Time-Varying Parameter MIDAS Models: Application to Nowcasting US Real GDP

We introduce a novel time-varying parameter mixed-data sampling (TVP-MIDAS) framework. Specifically, we allow both the MIDAS weights and the coefficients representing the overall impacts of the high-frequency variables to vary over time. This is done by introducing a class of linear parameterizations, which facilitate estimation in settings with a large number of high-frequency predictors. We demonstrate the usefulness of this framework via an application of nowcasting US GDP in real-time using monthly, weekly and daily predictors. The results show that the TVP-MIDAS models produce superior nowcasts, and are particularly effective in capturing the downside risk compared to their time-invariant counterparts.