Machine Learning Models for Early Warning of Financial Crises in the U.S. Economy Using Macro-Financial Indicators

Early identification of U.S. recessions is critically important for policymakers, financial institutions, and investors; however, accurate prediction remains challenging. The data is sparse, things in the economy shift in ways that aren’t always predictable, and macro-financial indicators don’t always move in straight lines. This study looks at a set of machine learning models, logistic regression, random forest, and ‎XGBoost, to see how well they can flag recessions six months ahead. We use a monthly dataset stretching from 1970 to 2025, and define ‎recession periods using the National Bureau of Economic Research dates. The problem is set up as a binary classification: will a recession ‎happen or not? To ensure methodological rigor, the models are trained using strict time-series cross-validation, and we evaluate their performance using ROC-AUC, precision-recall curves, Brier scores, calibration curves, and the lead time to issue a warning. We find that non-‎linear models like XGBoost and random forests tend to beat logistic regression by 5–10% on ROC-AUC. These models capture nonlinear ‎interactions between indicators that simpler linear models may fail to detect. The yield curve spread comes out as the most reliable signal, ‎followed by unemployment and the VIX volatility index. Credit indicators add a little extra, but not much. The models can give useful warning signals about 5–6 months before a recession starts, though accuracy drops when you look further ahead or during times when the economy goes through big changes, like after 2008 or post-COVID-19. Calibration tests show that predicted probabilities aren’t perfect; turning ‎them into reliable risk estimates needs some care. The study also points out gaps in previous ML-based early warning work: people often ‎rely on random cross-validation, don’t test models over different forecast horizons, and don’t focus enough on interpretability. By tackling ‎these issues, our approach is more robust and respects the time dimension, making it more suitable for real-world policy use. Overall, the ‎results suggest that well-calibrated, interpretable ML models can work alongside traditional econometrics to help policymakers act before a ‎recession hits.