Mixing mixed frequency and diffusion indices in good times and in bad: an assessment based on historical data around the great recession of 2008

In this paper, we analyze the forecasting performance associated with using machine learning, shrinkage, and variable selection methods during a historical period that contains the Great Recession of 2008. We find that these methods are most useful during “low” GDP growth periods, while simple autoregressive models are adequate during “high growth” periods. This finding stems from the introduction of very simple “hybrid” models that employ dynamic recursive (rolling) thresholding in order to switch between benchmark linear models and more complex index-driven models, depending on GDP growth conditions. In the context of predicting both quarterly real GDP growth and CPI inflation, these hybrid models are found to be superior, for all forecast horizons. When comparing the hybrid models against a host of alternatives, mean square forecast error gains reach as high as 35%, during the Great Recession, and remain significant throughout our entire prediction period. Additionally, the very best short-term GDP forecasting models contain variants of the Aruoba et al. (2009) business conditions index, although these models are most useful when diffusion indices are also incorporated. Thus, mixing mixed frequency and diffusion indices matters. Finally, across all experiments, we find strong new evidence of the usefulness of survey predictions, including those from the Survey of Professional Forecasters, and those from the Livingston Survey. While we leave the examination of alternative datasets, such as those including other recessionary periods, episodes of war, and epidemics to future research, we hypothesize that the findings in this paper point to the potential usefulness of machine learning, shrinkage, and variable selection methods during recessions, as well as to the usefulness of the hybrid models that we introduce.