What do banks tell us about financial stability? Predicting systemic crises using text-based machine learning

This paper extends the literature studying the prediction of financial crises in two ways, namely by: (i) developing a new text-based indicator measuring banks' sentiment tailored to the context of financial stability, and (ii) applying machine learning (ML) techniques to predict systemic crises in the euro area as defined by the European Systemic Risk Board (ESRB). In-sample analysis indicates that banks' financial stability sentiment (BFSS) is a highly statistically significant predictor of systemic crises, with a negative one standard deviation shock in the BFSS indicator corresponding to increases in the probability of a systemic crisis of 7 and 3 percentage points one-quarter and four-quarters ahead, respectively, while controlling for the credit cycle. Out-of-sample results show that, while the BFSS tends to improve the predictive performance of baseline logistic regression models, ML models grounded in financial stability dictionaries deliver substantially higher predictive accuracy in forecasting systemic crises. By improving the accuracy and timeliness of systemic crisis prediction, this novel application can be useful to complement conventional approaches for calibrating macroprudential policy tools and enhance crisis prevention frameworks.