Random forest versus logit models: Which offers better early warning of fiscal stress?

This study seeks to answer whether it is possible to design an effective and useful machine learning‐based early warning system signaling a risk of fiscal stress in the near future. To do so, multiple models based on econometric logit and the random forest models are designed and compared. Using a dataset of 20 annual frequency variables pertaining to 43 advanced and emerging countries during 1992–2018, the results confirm the possibility of obtaining an effective model, which correctly predicts 70–80% of fiscal stress events and tranquil periods. The random forest‐based early warning model outperforms logit models both in terms of aggregate forecasting properties and when applied to an example of the euro area sovereign debt crisis. While the very effective random forest model is commonly understood to provide lower interpretability than logit models do, this study employs tools that can be used to provide useful information for understanding what is behind the black‐box. These tools can provide information on the most important explanatory variables and on the shape of the relationship between these variables and the outcome classification. Thus, the study contributes to the discussion on the usefulness of machine learning methods in economics.