Abstract

This paper conducts a comprehensive horse race between several popular machine learning models and compares their performances with a traditional logistic regression model at predicting US recessions. Each model is tested at three-, six-, nine-, 12-, and 18-month horizons. A gradient boosting, random forest, support vector classifier, feedforward neural network, and custom consensus model are all compared against the logistic regression model. A dataset spanning 1962-2023 using six leading indicators is used for training and testing the models. The predictive ability of each model is evaluated using two scoring methodologies. The first scoring methodology is a novel ranking system using accuracy, precision, recall, F1, and AUROC scores. The second scoring methodology uses bootstrapped AUROC scores to perform a one-sided test on every model pair across each horizon. The results strongly suggest the gradient boosting, random forest and consensus models are generally superior at recession classification across horizons in comparison to the logistic regression model. The evidence for outperforming the logistic regression model is significantly weaker for the neural network and support vector classifier models. Finally, based on the findings of the consensus model, the results indicate combining the predictions of multiple models can produce better recession predictions than relying on the outputs of a single model.