APPLICATIONS OF MACHINE LEARNING IN FORECASTING RECESSIONS: BOOSTING UNITED STATES AND JAPAN

Abstract

Does applying machine learning on large datasets yield accurate recession forecasts? This paper applies boosting, considered one the best off-the-shelf classifiers in machine learning, to forecasting recessions in the United States and Japan. Instead of forecasting recessions with one or a few predictors, we utilize large macroeconomic datasets and use boosting to select the most predictive variables and perform prediction. We investigate if a large predictor set, specifically the 132 monthly predictors from Stock and Watson (2005), combined with boosting can forecast recessions better than the best logit model in the United States. We then look outside of the United States to see if a similarly large predictor set in Japan predicts recessions better than the best logit model. We find that while boosting outperforms the best logit model in-sample, boosting actually performs worse than the best logit model in the United States and Japan out-of-sample. By carefully selecting a smaller dataset that consists of leading indicators, we are able to boost a small dataset that performs better than boosting the large dataset. Our general finding reiterates the parsimonious principle, that simpler models often outperform more complex models.