Investigating Bank Failures Using Text Mining

We extend beyond healthiness assessment of banks using quantitative financial data by applying textual sentiment analysis. Looking at public annual reports for a large sample of U.S. banks in the 2000-2014 period, we identify 52 public bank holding companies that were associated with bank failures during the global financial crisis. Utilizing sentiment dictionaries designed for financial context, we find that negative and positive sentiments discriminate between failed and non-failed banks 88% and 79%, respectively, of the time. However, we find that positive sentiment contains stronger predictive power than negative sentiment; out of ten failed banks, on average positive sentiment can identify six true events, while negative sentiment identifies five failed banks at most. While one would link financial soundness with more positive sentiment, it appears that failed banks exhausted more positive sentiment than their non-failed peers, whether ex-ante in anticipation of good news or ex-post to conceal financial distress.