Machine Learning Prediction of Recessions: An Imbalanced Classification Approach

Abstract

The author examines the problem of predicting recessions from a machine learning perspective and employs a number of machine learning algorithms to predict the likelihood of recession in a given month using historical data from a set of macroeconomic time-series predictors. The author argues that, owing to the low frequency of historical recessions, this problem is better dealt with using an imbalanced classification approach. The author applies measures to compensate for class imbalance and uses various performance metrics to evaluate and compare models. With these measures in place, ensemble machine learning models predict recessions with high accuracy and great reliability. In particular, a random forest model achieves a near perfect true-positive rate within the historical training sample, generalizes extremely well to a test period containing the 2008–2009 financial crisis, and shows elevated recession probabilities during the last few months of 2019, associated with the tightened macroeconomic environment and worsened by the COVID-19 pandemic.

TOPICS: Big data/machine learning, performance measurement, simulations

Key Findings

• • We examine prediction of US recessions from a machine learning (classification) perspective.

• • We employ class imbalance adjustments and carefully analyze its impacts on predictions.

• • Ensemble models accurately predict recessions during and preceding major financial crises.