Variable Selection in Predictive MIDAS Models

Abstract

In short-term forecasting, it is essential to take into account all available information on the current state of the economic activity. Yet, the fact that various time series are sampled at different frequencies prevents an efficient use of available data. In this respect, the Mixed-Data Sampling (MIDAS) model has proved to outperform existing tools by combining data series of different frequencies. However, major issues remain regarding the choice of explanatory variables. The paper first addresses this point by developing MIDAS based dimension reduction techniques and by introducing two novel approaches based on either a method of penalized variable selection or Bayesian stochastic search variable selection. These features integrate a cross-validation procedure that allows automatic in-sample selection based on recent forecasting performances. Then the developed techniques are assessed with regards to their forecasting power of US economic growth during the period 2000-2013 using jointly daily and monthly data. Our model succeeds in identifying leading indicators and constructing an objective variable selection with broad applicability.

Keywords: Forecasting, Mixed frequency data, MIDAS, Variable selection, GDP

JEL Classification: C53, E37