Machine Learning Models for Early Warning of Financial Crises in the U.S. Economy Using Macro-Financial Indicators
International Journal of Accounting and Economics Studies · 16.03.2026 · 13 (2) · 513-525
DOI: 10.14419/wkr3pj23 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Accounting and Economics Studies, online erschienen am 16.03.2026. Verfasst von Md Khalilor Rahman und 7 weiteren Autor:innen.
Abstract (Originalsprache)
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.
Zitieren
APA
Rahman, M. K., Hossain, M. S., Haque, S. U., Jahed, K. A., Robbani, M. S., Shati, M. A., Tuly, T., & Pramanik, M. T. (2026). Machine Learning Models for Early Warning of Financial Crises in the U.S. Economy Using Macro-Financial Indicators. International Journal of Accounting and Economics Studies, 13(2), 513-525. https://doi.org/10.14419/wkr3pj23
Vancouver
Rahman MK, Hossain MS, Haque SU, Jahed KA, Robbani MS, Shati MA, et al. Machine Learning Models for Early Warning of Financial Crises in the U.S. Economy Using Macro-Financial Indicators. International Journal of Accounting and Economics Studies. 2026;13(2):513-525. doi:10.14419/wkr3pj23
BibTeX
@article{Rahman2026,
author = {Rahman, Md Khalilor and Hossain, Md Sazzad and Haque, Sayem Ul and Jahed, Kazi Abu and Robbani, Md Sharif and Shati, Mahamuda Akter and Tuly, Tanjina and Pramanik, Md Toushif},
title = {Machine Learning Models for Early Warning of Financial Crises in the U.S. Economy Using Macro-Financial Indicators},
journal = {International Journal of Accounting and Economics Studies},
year = {2026},
volume = {13},
number = {2},
pages = {513--525},
doi = {10.14419/wkr3pj23}
}
EUVIMED