MACHINE LEARNING ALGORITHMS FOR EARLY WARNING SYSTEMS: PREDICTING SYSTEMIC FINANCIAL CRISES THROUGH NON-LINEAR ECONOMETRIC MODELS

Abstract

In view of the severe output and welfare losses caused by past systemic crises, including those of the Great Depression and the 2008 global financial collapse, there is a pressing need for early warning systems that effectively capture nonlinear risk dynamics. This study proposes a machine learning framework for predicting systemic financial crises, based on macroeconomic data from 35 countries spanning 1970 to 2022. Utilising ensemble methods—specifically random forest and XGBoost—the approach significantly enhances early warning capabilities compared to conventional linear models. Incorporating 78 economic and financial indicators, the methodology applies advanced feature engineering and selection to identify complex systemic risk patterns. The models demonstrate strong predictive performance, achieving an AUROC score of 0.97 and substantially outperforming logistic regression. Key predictors include GDP growth, trade volumes, investment trends, and demographic variables. The findings highlight critical nonlinear relationships and interactions among economic indicators, offering deeper insights into financial system vulnerabilities. This work underscores the potential of machine learning to support more intelligent, forward-looking financial risk assessment and policy design.