Developing an early warning system with machine learning and post-crisis information

This study develops financial crisis prediction models using machine learning algorithms applied to the Jord & agrave;-Schularick-Taylor Macrohistory Database. We construct an early warning system (EWS) that integrates post-crisis information in two ways. First, we use a three-outcome discrete dependent variable (normal, pre-crisis, and post-crisis) instead of a binary indicator and apply machine learning classification methods. Second, we introduce a predictor indicating whether other countries are in a post-crisis regime. Our results are mixed, suggesting that including post-crisis observations does not necessarily improve EWS performance.