Can Text-Based Statistical Models Reveal Impending Banking Crises?

Abstract

This paper introduces statistical models Wordscores and Wordfish to study and predict banking crises. While Wordscores is akin to supervised learning, Wordfish is analogous to unsupervised learning. Both methods estimate the position of banking distress on a tranquil-to-crisis spectrum. Findings suggest that the two statistical methods signal banking crisis up to two-years in advance, with robust results from AUROC, Granger causality and VAR impulse responses. Both methods outperform random forests in predicting crises using textual data. The Wordscores index highlights increased usage of banking sector nomenclature two years preceding a crisis, and Granger causes a crisis series with one and two lag lengths. Results from the Wordfish technique, a statistical model with Poisson distribution, show the index spikes before and during the Global Financial Crisis, when a large share of the countries in the world encountered banking crises. This paper contributes to literature on text-based models of banking crises by bolstering the preemptive policy responses available to policy makers. Given their early warning signals, both Wordscores and Wordfish can be considered a part of the toolset to monitor the stability and resilience of the banking sector.