eThesis - School of Business electronic theses

» hse_ethesis_12401.pdf   size:831 KB (850403)

Since 2007 several banks have fallen into bankruptcy in the U.S. What is historically notable in this situation is the amount of assets lost in bankruptcies that are already measured in hundreds of billions of dollars. Since financial institutions magnitude to the current economic system is crucial, bank failures have a dramatic impact also on real economy. Therefore, feasible bank failure prediction models can also diminish the real economy problems.

The purpose of this thesis is to study how accurately recent U.S. commercial bank failures can be predicted with logistic regression model utilizing financial statement variables. With the overall predictability of the model, also the statistical significance of the independent variables is studied. In addition, it is tested how the prediction accuracy reduces as the timeframe from the bankruptcy prolongs from one quarter to three years.

The evolution and history of bankruptcy prediction models and banking crises is also studied in order to develop the base for the bank failure model. It is also noted that several bankruptcy prediction models contain financial statement variables. In addition, by studying the history of the banking crises it is noticed that traditional banking factors such as liquidity, credit risk, and profitability have a substantial impact on failures of individual financial institutions. It is also argued that banks’ exposure to the subprime related securities might deteriorate the solvency of the financial institutions.

The data of the analysis in gathered from the FDIC database. It contains bank-specific variables from 2004 to 2009 including 124 commercial banks with total assets worth more than 500 million dollars. In the empirical part of the thesis it is noted that 25 independent variables are statistically significant for the bank failure prediction. On the other hand, several of these explanatory variables correlate significantly with each other, erasing the possibility to include all the statistically significant variables into the same model. Therefore, 72 potential models are constructed, which are then studied with the help of logistic regression. After short and long-term analysis it can be noticed that the most accurate failure prediction is provided by the model having nonaccrual rate, loan diversification, return on equity, capital growth, tax exposure, CMO ratio, uninsured deposits, risk free securities, dividend rate, loan growth, assets variation, and liquid assets as the explanatory variables. The accuracy of the model is measured by correctly classified (CC) percentage figure. The model in question can predict 95.16% of failures correctly one quarter prior the bankruptcies. CC figure is 82.26% one year, 72.58% two years, and 71.77% three years prior the bank failures.

The analysis recovers, however, that only two of the explanatory variables, nonaccrual rate and risk free loans, are both statistically significant in the long-term and consistent with time. Although the model with only two explanatory variables tends to lose some of its predicting accuracy, it permits a precise analysis of the independent coefficients. It can be confirmed that only one percentage point increase in nonaccrual rate at least doubles the bankruptcy probability. On the other hand, the more the bank ties its investments to risk free securities the greater the probability that the bank survives. Empirical analysis proves also that the logit model is slightly more suitable for bank failure prediction than the corresponding probit model.