Machine Learning and Deep Sequence Models for US Recession Prediction: A Redux

This paper revisits and extends the machine learning framework for U.S. recession prediction introduced by Yazdani2020 by incorporating post-pandemic macroeconomic dynamics, an expanded predictor set and machine learning models. Using monthly data from January 1959 through December 2024, recession forecasting is formulated as an imbalanced binary classification problem. We use downsampling for static models and class-weighted loss functions for neural networks and evaluate model performance using classification metrics robust to rare events. We further examine structural stability across four economic regimes and assess economic value through a dynamic stock–bond allocation strategy. We observe that ensemble tree methods, particularly gradient boosting (XGBoost, LightGBM) and random forests, consistently deliver the strongest discrimination, with out-of-sample AUC above 0.99 and PR-AUC above 0.96. The Transformer achieves probability calibration, and Deep sequence models exhibit high discrimination, while performance deteriorates across model classes in the 2020–2024 regime, especially for linear specifications. We also examine risk-adjusted returns of models. Overall, ensemble trees and Transformers show high predictive power and emerge as complementary tools in macroeconomic recession forecasting.