High-Dimensional Recession Prediction: A Comparative Analysis of Machine Learning Classifiers and Bayesian Logistic Regression with Alternative Priors

Abstract

It is important that economic recessions are predicted in time to enable policymakers and financial decision-makers but the official recession classifications are normally announced late. This paper suggests a data-driven model to forecast U.S. recessions using a wide range of macroeconomic variables that span the 1960-2026 timeframe. Having done the preprocessing and feature selection, a narrow group of predictors is then refined to enhance model efficiency and interpretability. The paper compares the performance of various machine learning methods, such as Logistic Regression, Random Forest, Support Vector Machine, and Gaussian Naive Bayes, as well as Bayesian Logistic Regression models with various prior assumptions including Normal, Laplace, Cauchy and Spikeand-Slab. Besides this, time-series methods are used to forecast macroeconomic variables six months in advance to aid forward-looking predictions. The results show that the Logistic Regression method has a good balance between predicting recession periods and false signals. Bayesian Logistic Regression is quite stable regardless of the choice of prior, although the Cauchy prior is the most effective in terms of detection and false alarms. These findings indicate that straightforward and understandable probabilistic models are still very useful to predict recessions and track the economy.