Data Aggregation in MIDAS Models: Improving Forecasting through Optimal Data Piling

The MIxed DAta Sampling (MIDAS) model has proven to be a valuable tool in the modeling of data sampled at different frequencies. With this new possibility arises the option to aggregate explanatory data into intermediate frequencies before regressing upon them. As such the room for noise can be decreased. This paper studies the added value of data piling in the use of MIDAS models. It finds that oftentimes there is an intermediate frequency of data aggregation that produces better forecasts than either the raw data or the fully aggregated data. Furthermore it appears that R2 and Akaike Information Criteria are in some cases accurate predictors of model optimality. However, the number of cases in which they are wrong remains too large. Still data piling should be regarded as a valuable addition to today’s econometric Time Series toolbox.