Predicting Bank Failures: Evidence from 2007 to 2010

Transcription

1 Predicting Bank Failures: Evidence from 2007 to 2010 Dr. Dan J. Jordan, DBA, CPA/ABV, CVA, CFF Dominican University of California Dr. Douglas Rice, DBA Golden Gate University Dr. Jacques Sanchez, DBA Bank of the West Dr. Christopher Walker, DBA Northeastern University Dr. Donald H. Wort, PhD California State University, East Bay Classification Codes: G34, G21, G12 Key Words: Studies, Bank, Failure ABSTRACT This paper examines the 225 banks that failed between February 2, 2007, and April 23, 2010, comparing them to a random sample of banks that had not failed as of April 23, We performed regression and discriminant analysis on quarterly call report data for one year, two years, three years, and four years prior to bank failure to determine whether the failure could have been predicted. Our model is statistically significant at the 1% level and predicts bank failures with 88.2% accuracy one year prior to failure, 78.6% two years prior to failure, 71.4% three years prior to failure, and 66.0% four years prior to failure.

2 I. INTRODUCTION The economic crisis that began in 2007 put great pressure on the US banking system. According to the Federal Deposit Insurance Corporation (FDIC) Failed Bank List, from February 2, 2007 through April 23, 2010, 225 US banks failed. ( Many of the failed banks were caught in the real estate market collapse and because they did not have sufficient capital to ride out the cycle, were forced by the FDIC to merge with or be sold to other institutions. The FDIC Banking Review (2003) points to the Prompt Corrective Action (PCA) provisions in the Federal Deposit Insurance Corporation Act of 1991 (FDICA) as the mandate for banking regulators to promptly close critically undercapitalized banks. Capital amounting to two percent or less of tangible assets has been set as the threshold for classification as a critically undercapitalized bank. By the time a bank s tangible capital ratio falls to the two percent threshold, it is often too late to save the bank; particularly as asset quality deteriorates - forcing banks to write down asset values, severely weakening the bank s capital position and leaving the FDIC to pick up the pieces. This is a major reason why the Board of Governors of the Federal Reserve devised the CAMELS rating system which evaluates bank Capital, Asset quality, Management, Earnings, Liquidity, and Sensitivity to market risk to create a watch list of troubled or risky banks to be monitored between on-site examinations. 1

3 CAMELS ratings are kept strictly confidential by the Federal Reserve and are not available to the public. So, how can an investor or a lender identify and avoid weak banks early enough to prevent loss of value or liquidity associated with potential bank failure? Because depositors are covered by FDIC insurance and do not concern themselves with bank failure, the moral hazard associated with failing banks offering high interest rates to depositors makes it imperative that such banks be identified early by regulators to safeguard taxpayer funds. Altman s Z and Zeta score models are oriented toward industrial companies, as are most of the eighty-nine other models described in Aziz and Dar s (2006) literature review. Several recent articles use neural network methodologies to predict bank failures. These neural network methodologies are quite complicated and require a mapping of inputs to outputs using layers and neurons to create a complex learned algorithm (Muller, Steyn- Bruwer and Hamman, 2009). This paper employs a multiple discriminant analysis methodology similar to Altman s to devise a bank failure prediction formula and uses variables available or calculable from quarterly bank call reports. We feel that developing a model with publicly available data will provide significant value to not only the FDIC, but also investment firms, financial analysts, and the banks themselves. 2

4 This paper is organized in four parts: First is a review of the relevant academic literature. Second is hypothesis development. Third is a discussion of data and methods used in the study and the results from the study. Finally, the paper concludes with a discussion of the implications of the study s findings for theory, practice, and future research. II. LITERATURE REVIEW Over the years, numerous authors have attempted to predict corporate failure using various methodologies. The most well known is Altman s (1968) multiple discriminant analysis of thirty-three bankrupt and thirty-three non-bankrupt manufacturers. The variables used in his seminal study are: (1) working capital/total assets, (2) retained earnings/total assets, (3) earnings before interest and taxes/total assets, (4) market value equity/book value of total liabilities, and (5) sales/total assets. This model is shown in Altman s (1968) study to be effective in predicting bankruptcy up to two years prior to distress and that accuracy diminishes substantially as the lead time increases (Altman, 2000). Altman, Haldeman and Narayanan created the Zeta Credit Risk Model (1977) as a second generation discriminant model which appeared to be quite accurate for up to five years prior to failure (Altman, 2000). The Zeta model consists of seven variables: (1) return on assets, (2) stability of earnings, (3) debt service, (4) cumulative profitability, (5) liquidity, (6) capitalization, and (7) asset size. 3

5 While the Altman models have been shown to be useful for manufacturing firms, they have not been shown to work well for financial companies, such as banks. In 1983, Richard J. Taffler created a UK-based z-score model that has been shown in a recent study by Agarwal and Taffler (2007) to have good failure prediction ability. Since 1968, numerous models, methodologies and theories have been put forth to improve upon Altman s 1968 and 1977 models and Taffler s 1983 model. Aziz and Dar (2006) list forty-six articles using the following eighty-nine methodologies: Multiple Discriminant Analysis 27 Neural Networks 8 Logit 19 Balance Sheet Decomposition Measure (entropy theory) 4 Genetic Algorithms 4 Recursive Partitioning (decision tree) Analysis 5 Rough Sets Model 3 Credit Risk Theories (including option pricing and macro-economic theories) 2 Univariate 3 Cash Management Theory 3 Case-Based Reasoning 2 Cumulative Sums Model (time series) 2 Partial Adjustment Model (time series) 1 Linear Probability Model 3 4

6 Probit 2 Gamblers Ruin Theory 1 Total 89 These models are applicable to companies in the following industries: Manufacturing Industries 11 Manufacturing and Retail 10 Industrial 9 Mixed Industries 39 Telecom 1 Retail Firms 3 Not Available 5 Banks 1 Motor Components 1 Construction Industries 3 Savings and Loan Associations 2 Mining and Manufacturing 1 Non-Financial Firms 1 Oil and Gas 2 Total 89 Of the above listed articles, only one specifically examines banks and two examine savings and loan associations; the latest covering the 1986 to 1987 time period. A search for recent articles covering banks identifies the following articles and methodologies: 5

7 (1) Schaeck (2008) uses a quantile regression approach to compare high-cost to low-cost bank failures. (2) Ozkan-Gunay and Ozkan (2007) use a non-linear artificial neural network approach to analyze 59 Turkish banks (36 successful and 23 failed banks). They find that seventysix percent of the failed banks are correctly indicated and ninety percent of the nonfailed banks are correctly indicated. (3) Ercan and Evirgen (2009) investigate the factors that were important in the failure of Turkish banks using a principal component analysis methodology. (4) Yim (2007) uses a hybrid artificial neural network methodology to predict failure of firms from Australia s financial services sector. She is successful in predicting 100% of failed firms one year before failure, but is only successful in predicting 33.3% of failed firms two years before failure. (5) Jesswein (2009) compares the Texas Ratio for a sample of 37 failed banks from 2008 and 2009 compared to 7,075 non-failed banks noting that such a measure offers important insights but may not be sufficient as a general, all-purpose tool. According to Jesswein, the ratio is calculated by dividing the bank's non-performing assets (non-performing loans plus other real estate owned) by the sum of its tangible equity capital and loan loss reserves. (6) Platt and Platt (1991) find that use of industry-relative ratios appear to add incremental information in bankruptcy prediction. 6

8 Recently, the authors of this paper have been analyzing the market-to-book ratios of a sizable sample of publicly-traded banks from 2006 to the present, using the following explanatory variables, among others (Jordan, et. al., 2009): (1) Ratio of non-interest income to interest income. (2) Ratio of non-accrual assets plus owned real estate ORE to total assets. (3) Ratio of interest income to earning assets. (4) Ratio of Tier One capital to total assets. (5) Bank Holding Company dummy variable representing whether or not the bank is part of a holding company. (6) Savings Bank dummy variable representing whether or not the bank is a savings bank. (7) MSA dummy variable representing whether or not the bank is located in a Metropolitan Statistical Area. Because we find that these variables have significant explanatory power on the market-tobook ratio of our prior sample of publicly traded banks, we investigate in this paper the extent of their power in predicting the surge of bank failures since Our ratios are somewhat different than ratios covered by the following studies: Study Pantelone and Platt (1987) Pettway and Sinkey (1980) Sinkey (1975) Variables Leverage, liquidity, profitability, management efficiency, diversification and risk Operating expenses as a percentage of operating income and investments as a percentage of total assets Cash and U. S. Treasury securities as a percentage of total assets, loans as a percentage of total assets, provision for loan losses as a percentage of operating 7

9 Sinkey (1978) expense, operation expense as a percentage of operating income, loan revenues as a percentage of total revenue, U. S. Treasury security income as a percentage of total revenue, state and local obligation revenue as a percentage of total revenue, interest paid on deposits as a percentage of total revenues, and other expenses as a percentage of total revenue Net capital ratio Ratios explored by Cole and Gunther (1998) come the closest to our variable set. As a percentage of gross assets, they explore the following variables: equity capital, past due loans, nonaccrual loans, other real estate owned, net income, investment securities, and large certificates of deposit. III. HYPOTHESIS DEVELOPMENT It appears from the above research that bank failures can be reasonably predicted up to two years prior to failure using various methodologies. It also appears from the authors previous research that banks market-to-book ratios can be explained with high statistical significance by using the seven variables listed in the literature review. Our hypotheses are: Hypothesis 1: The seven variables identified can be used to predict bank failure up to four years prior to the failure date. Hypothesis 2: The ratio of the expense provision for bad debts as a percentage of total gross loans is a predictor of bank failure. 8

10 Hypothesis 2 is based on the reasoning that one of the seven variables explored in Hypothesis 1, the ratio of non-accrual assets plus ORE to total assets is a balance sheet measure of bad assets on the books. An income statement measure, provision for uncollectible accounts, represents the current period increase in bad debts on the books. It is logical that both measures, current and future bad debts, would concern bank regulators. Due to the housing crisis, mortgage loan delinquencies have increased significantly within the 2007 to 2009 time period. According to the Mortgage Bankers Association quarterly National Delinquency Surveys, the seasonally adjusted delinquency rate on one-to-fourunit residential properties was 5.82 percent of all loans outstanding in the fourth quarter of However, it increased to 7.88 percent in the fourth quarter of 2008 and stood at 9.47 percent in the fourth quarter of 2009; an increase of 206 basis points over the two-year period. Hypotheses 1 and 2 include variables about current and future bad debts. These raise questions about current real estate loans that might go bad and whether regulators should be concerned if a bank has a high percentage of its assets invested in real estate loans. This leads to our third hypothesis. Hypothesis 3: The ratio of real estate loans as a percentage of total assets is a predictor of bank failure. 9

11 IV. DATA AND STUDY DESIGN Sample Data We initially selected all the banks that failed from January 1, 2007 through April 23, 2010 from the FDIC web site: The banks failed in these years as follows: (Through 4/23/2010) 57 Total 225 Subsequently, we selected a random sample of 225 non-failed banks from the fourth quarter 2009 call reports from the FDIC web site: and then randomly matched the dates of the quarterly call report data of each of the non-failed banks to each of the failed banks. The failed and non-failed banks were located in the following states: All Banks Failed Banks Non-Failed Banks Georgia Illinois California Florida Minnesota Texas Washington

12 Nevada Other States Totals The failed and non-failed banks size characteristics are as follows: Total Assets (000) All Banks Failed Banks Non-Failed Banks Over 10,000, ,000,000 to 9,999, ,000,000 to 4,999, ,000,000 to 1,999, ,000 to 999, ,000 to 499, ,000 to 199, Under 100, Totals The failed and non-failed banks bank class characteristics are as follows: Bank Class All Banks Failed Banks Non-Failed Banks National Bank, Federal Reserve Member State Bank, Nonmember Savings Association Savings Bank State Bank Totals

13 For each selected bank, the quarterly call reports are downloaded from the FDIC web site: The quarterly call reports downloaded cover the period from the third quarter of 2002 through the fourth quarter of This provides a full four years of data prior to the failure quarter for our analysis of each failed bank and each randomly selected matching non-failed bank. We then used raw data for each of the failed and the non-failed banks for each selected quarter and calculated the following variables for each bank for each quarter: (1) Ratio of non-interest income to interest income. (2) Ratio of non-accrual assets plus ORE to total assets. (3) Ratio of interest income to earning assets. (4) Ratio of Tier One capital to total assets. (5) Bank Holding Company dummy variable. (6) Savings Bank dummy variable. (7) MSA dummy variable. (8) Ratio of bad debt expense provision to total gross loans. (9) Ratio of real estate loans to total assets. To test our hypotheses, we performed two types of statistical tests ANOVA regression analysis and MDA. For our total sample of failed and non-failed banks, we performed regression analysis to determine the relationship and strength of each variable to the failure status of each bank. These tests were performed using failed bank and matching non-failed bank data from all 12

14 quarters beginning one quarter prior to failure and extending back four years prior to failure. We also analyzed data from each of the quarters: (1) one year prior to failure, (2) two years prior to failure, (3) three years before failure, and (4) four years prior to failure. We randomly split the failed and non-failed bank data into two groups: (1) a training group, and (2) a testing group. We conducted MDA on the training group data from all quarters beginning one quarter prior to failure and extending back four years prior to failure to produce a discriminant formula for predicting bank failure. This is the same time period used above in our summary regression analysis. Finally, we tested the discriminant formula using the testing group data. Regression Results: V. RESULTS The results of the regressions of the failure status (1 = failed and 0 = not failed) against the nine dependent variables noted above for the four years prior to failure and the summary data is as follows: Yr -1 Yr -2 Yr -3 Yr -4 ANOVA Regression Characteristics Failure Status and Various Financial Metrics 1 Top number is the coefficient and bottom number is the t-value. * Indicates statistical significance at the 0.10 level. ** Indicates statistical significance at the 0.05 level. *** Indicates statistical significance at the 0.01 level. Independent Variables Summary Model 1 Model 1 Model 1 Model 1 Model 1 Intercept (-4.292)*** (-1.107) (2.743)*** (-1.542) (-1.280) Non-Interest Income to Interest Income Ratio.002 (.596).015 (2.297)** (-.046) (-.450) (-1.240) Ratio of Non-Accrual Assets ORE to Total Assets Ratio of Interest Income to Earning Assets (16.602)***.692 (2.703)*** (8.464)***.321 (.339) (1.035).462 (.442) (-.214).265 (.194) (.290) (-.032) 13

15 Tier One Capital to Total Assets Ratio (-3.993)*** (-4.743)***.141 (.405).051 (.141).643 (1.780)* Ratio of Bad Debt Expense Provision to Total Gross (.155).097 (1.888)*.023 (.775).031 (.605).655 (1.405) Loans Ratio of Real Estate Loans to Total Assets.703 (22.742)***.692 (5.113)***.976 (7.233)***.779 (6.010)***.681 (5.062)*** Bank Holding Co. Dummy Variable.021 (1.749)*.006 (.135).055 (1.115).052 (1.008).026 (.469) Savings Bank Dummy Variable (-5.692)*** (-.960) (-1.498) (-1.687)** (-1.349) MSA Dummy Variable )***.213 (4.936)***.265 (5.516)***.309 (6.349)***.331 (6.420)*** N 6, R F-Value *** *** *** *** *** Significance.000***.000***.000***.000***.000*** Regression Discussion: As reported in the above table, each model is statistically significant at the.000 level. The dependent variable in our regression is the failure status of a bank. Either the bank failed (1 is entered as the dependent variable) or it did not fail (0 is entered as the dependent variable). Accordingly, we can interpret the coefficients as follows: A. Coefficients: Non-Accrual plus ORE to Total Assets Ratio This is a ratio of bank assets in a non-accrual status plus owned real estate repossessed to total assets and represents a balance sheet measure of bad assets on the books. The positive coefficient means that higher non-accrual assets and owned real estate as a percentage of total assets result in a higher probability of bank failure. This is not at all surprising. In the current crisis, the root cause of bank failure has been bad loans and mortgage backed securities. We expected a strong positive relationship between bank failure and bad loans and foreclosed real estate; which is what we see for the summary and the year -1 model. Both years have large positive coefficients that are statistically significant at the 1% level. 14

16 For years -2, -3, and -4, and we see non-significant relationships between bank failure and the non-accrual plus ORE to total assets ratio. Ratio of Bad Debt Expense Provision to Total Gross Loans This is the ratio of bad debt expenses for the quarter to total gross loans. This is an income statement measure of the deterioration of loan quality for the quarter. The positive coefficient means that higher bad debt expenses as a percentage of total loans result in a higher probability of bank failure. This also seems obvious. However, we find that, although the coefficient is positive, it not statistically significant except for the year -1 model, where it is significant at the 10% level. Ratio of Real Estate Loans to Total Assets This is the percentage of bank assets tied up in real estate loans. The positive coefficient means that higher real estate loans as a percentage of total assets result in a higher probability of bank failure. As noted above, the root cause of bank failure during our test period was generally the deteriorating quality of real estate loans. We would, therefore, expect a strong positive relationship between bank failure and real estate loans; which is reflected in all of the periods analyzed. The coefficients are positive and statistically significant at the 1% level in all of the models. Interest Income to Earning Assets Ratio This is the ratio of total interest income to earning assets. The positive coefficient means that higher interest income as a percentage of earning assets results in a higher probability 15

17 of bank failure. This is puzzling. One would expect that higher interest income would result in a lower probability of failure. It is possible that this is also a measure of risk and that banks making riskier loans would earn higher interest but not enough to make up for the higher risk of default. The higher interest income may also result from institutions increasing loans rates or implementing default interest rates for borrowers who are experiencing financial difficulties. This ratio is statistically significant at the.01 level for the summary model, but it is not significant for any of the yearly models. Non-Interest Income to Interest Income Ratio This is the ratio of non-interest income to interest income. A positive coefficient would mean that higher non-interest income relative to interest income results in higher probability of bank failure. This ratio is generally not statistically significant in our model results. Tier One Capital to Total Assets Ratio This is the ratio of Tier One capital available to support the assets of the bank. It is a cushion to cover possible future operating and investment losses that might be incurred by the bank. A negative coefficient means that more Tier One capital is associated with a lower probability of bank failure. This ratio is statistically significant at the.01 level for the summary and the year -1 models. It is generally not significant for the other models. MSA Dummy Variable This is 1 if the bank is located in an MSA or 0 if the bank is not located in an MSA. A 16

18 positive coefficient would indicate that being located in a metropolitan area is related to an increase in a bank s probability of failure. This variable is statistically significant at the.01 level for all periods examined. Bank Holding Company Dummy Variable This is 1 if the bank is a part of a holding company or 0 if the bank is not part of a holding company. A positive coefficient would indicate that doing business under a holding company umbrella is related to an increase in a bank s probability of failure. This variable is generally not statistically significant. Savings Bank Dummy Variable This is 1 if the bank is a savings bank or 0 if the bank is not a savings bank. A negative coefficient indicates that savings banks are related to a lower probability of failure. This variable is statistically significant at the.01 level for the summary model but is not generally significant for any of the other models. So, as expected, higher non-accrual and real estate owned assets and higher real estate loans are associated with more bank failures; while higher Tier One capital ratios are associated with fewer bank failures. This is generally as expected. The MSA dummy variable and the Interest Income to Earning Assets Ratio relationships, however, are unexpected. Nevertheless, the overall model produced statistically significant regression results. 17

19 Accordingly, we can use this information to devise a predictive model of bank failure similar to Altman s Z score and Zeta score models by performing multiple discriminant analysis of the sample data. As noted in the design section of this paper, we split the total sample of failed and nonfailed bank data into two groups: (1) a training group, and (2) a testing group to create and test a discriminant function. To create the function, we run multivariate discriminant analysis on the training group, using the same variables as in the regressions above, using data from all quarters beginning one quarter prior to failure and extending back four years prior to failure. The resultant discriminant function is as follows: BankZ Score TM =.022 Nonii% OreNaaset% IntIncEarnasst% Rbct1% Elnatr% Lnre% BHCDummy SBDummy MSADummy Where: Nonii% = Ratio of Non-Interest Income to Interest Income. OreNaaset% = Ratio of Non-Accrual Assets + ORE to Total Assets. IntIncEarnasst% = Ratio of Interest Income to Earning Assets. Rbct1% = Tier One Capital to Total Assets Ratio. Elnatr% = Ratio of Bad Debt Expense Provision to Total Gross Loans. Lnre% = Ratio of Real Estate Loans to Total Assets. 18

20 BHCDummy = Dummy variable equal to 1 if bank is a member of a bank holding company, otherwise the value is equal to 0. SBDummy = Dummy variable equal to 1 if bank is a savings bank, otherwise the value is equal to 0. MSADummy = Dummy variable equal to 1 if bank is located in a metropolitan area, otherwise the value is equal to 0. Based on the above noted MDA analysis, if the resulting BankZ Score TM is greater than 2.866, a bank is considered a failure risk. Multiple Discriminant Analysis Results The following are the results of our Multiple Discriminant Analysis on the training group of our sample of failed and non-failed banks: Statistical Analysis: The assumption of equal covariance matrices is tested using Box s M test. The value of the statistic is and the p-value is zero, using the F approximation. Therefore, the assumption of constant variance is not satisfied. The adequacy of the model is tested as follows: (1) the canonical correlation is.498, (2) the Eigenvalue is.330, (3) Wilks Lambda is.752, and (4) the Chi-square test value is These values indicate that the above noted discriminant function is significant at the.000 level. 19

21 We then use the above noted discriminant function to classify failed and non-failed banks for the training group and the testing group (the holdout sample). The results of the Multiple Discriminant Analysis for the quarter before failure and the four prior years for the training sample, the holdout sample, and the total sample are as follows: Multiple Discriminant Analysis Results Summary Results Yr -1 Results Yr -2 Results Yr -3 Results Yr -4 Results Training Sample: Failed Banks Correctly Predicted 1, Type I Error Type II Error Non-Failed Banks Correctly 1, Predicted Total 3, Incorrectly Predicted in Total Correctly Predicted in Total 2, % of Failed Banks Correctly 75.6% 86.8% 78.9% 68.5% 63.3% Predicted % of Non-Failed Banks Correctly 68.5% 68.5% 66.7% 74.0% 70.2% Predicted % of Total Incorrectly Predicted 27.8% 22.2% 26.9% 28.8% 33.3% % of Total Correctly Predicted 72.2% 77.8% 73.1% 71.2% 66.7% Holdout Sample: Failed Banks Correctly Predicted 1, Type I Error Type II Error Non-Failed Banks Correctly 1, Predicted Total 3, Incorrectly Predicted in Total Correctly Predicted in Total 2, % of Failed Banks Correctly 78.1% 89.3% 78.3% 74.3% 68.8% Predicted % of Non-Failed Banks Correctly 71.6% 60.2% 67.3% 76.0% 71.7% Predicted % of Total Incorrectly Predicted 25.1% 22.4% 26.8% 24.9% 29.8% % of Total Correctly Predicted 74.9% 77.6% 73.2% 75.1% 70.2% Total Sample: Failed Banks Correctly Predicted 2, Type I Error Type II Error Non-Failed Banks Correctly 2, Predicted Total 6, Incorrectly Predicted in Total 1,

22 Correctly Predicted in Total 5, % of Failed Banks Correctly 76.8% 88.2% 78.6% 71.4% 66.0% Predicted % of Non-Failed Banks Correctly 71.3% 64.8% 67.0% 75.0% 71.0% Predicted % of Total Incorrectly Predicted 26.0% 22.3% 26.8% 26.9% 31.6% % of Total Correctly Predicted 74.0% 77.7% 73.2% 73.1% 68.4% Multiple Discriminant Analysis Discussion: From the above table, it can be seen that our model successfully predicts from 66.0% (4 years prior to failure) to 88.2% (1 year prior to failure) of failed banks, with an overall success rate of 76.8%. This compares favorably to the 80% success rate for the 1948 to 1965 time period bank study noted by Aziz and Dar (2006) and the 76% success rate reported by Ozkan-Gunay and Ozkan (2007). Our model also compares favorably to the Yim (2007) model that predicted 100% of the failed Australian financial firms one year in advance but only 33.3% failed firms two years in advance. VI. DISCUSSION The hypotheses tested in this paper are as follows: H1: The seven listed variables can predict bank failure up to four years prior to the failure date. H2: The ratio of the expense provision for bad debts as a percentage of total gross loans is a strong predictor of bank failure. H3: The ratio of real estate loans as a percentage of total assets is a strong predictor of bank failure. 21

23 Based on the Summary Model ANOVA regression results, Variable (2) -- The Ratio of Non-Accrual Assets + ORE to Total Assets, Variable (3) -- The Ratio of Interest Income to Earning Assets, Variable (4) -- The Tier One Capital to Total Assets Ratio, Variable (9) -- The Ratio of Real Estate Loans to Total Assets, the Savings Bank and MSA Dummy variables all have a strong statistical relationship at the.01 level to bank failure status. These variables are somewhat strong in the year -1 model, but generally lose their potency as individual variables as the models move back to year -2, -3, and -4, with the exception of the Ratio of Real Estate Loans to Total Assets, which is strong in all of the models. The MSA Dummy variable is also strong in all of the models. The F-values of each of the models is significant at the.000 level which suggests that all of the variables have value in the regression equation. The Summary Model (including all of the variables), which is used in the Multiple Discriminant Analysis, is statistically significant at a very high level and the Discriminant Function is successful in correctly predicting from 66.0% (4 years prior to failure) to 88.2% (1 year prior to failure) of failed banks, with an overall 76.8% success rate in predicting bank failure. Consequently, H1 and H3 are accepted and H2 is not outright rejected due to its contribution to the overall model. 22

24 The results of this study suggest that the models used currently by bank regulators can be improved and that banks at risk of failure can be identified earlier (up to four years prior to failure). Considering the PCA provisions in the Federal Deposit Insurance Corporation Act of 1991 and the resultant mandate for banking regulators to promptly close critically undercapitalized banks, any improvement in early detection of at risk banks should allow those banks to improve operations and capital adequacy in sufficient time to avoid bank failure. This has the potential of saving countless millions in stockholder value and taxpayer funds. In addition, because the variables in our model are readily available to the public, investors and lenders can easily calculate the BankZ Score TM of each bank they are considering doing business with. This will allow them to modify the terms and monitoring activities involved in their relationship, better matching their risks with their rewards. VII. LIMITATIONS OF THIS STUDY As with any study, this study has limitations that might affect the generalization of the results. First, the results reported might be valid only for the period of time studied. Second, the economics associated with the recent rash of bank failures might not be repeated in the future or, if it is, the causality may differ from the time period examined by this study. Accordingly, extreme care should be taken when attempting to generalize the results of this study to any other time periods. 23

25 Future studies may wish to examine the outliers in our failed bank data, any management differences between failed and non-failed banks, any geographic differences between failed and non-failed banks, and whether size has a significant effect on bank failure. However, notwithstanding the limitations of this study, we believe the work makes a number of important contributions. First, it analyzes, in a timely manner, the 2007 to 2010 surge in bank failures. Second, the work provides a mechanism that uses readily available bank data to predict which banks to avoid as a lender, depositor, or an investor. Third, it offers insight and a basis for action to policy makers and bank decision-makers as a means to avoid such problems in the future. Further, it is hoped that this study will stimulate future research. VIII. CONCLUSION The goal of this study is to determine whether a formula, which includes certain open sourced variables readily available from bank quarterly call reports, has strong predictive power in identifying failed banks up to four years prior to failure. The results of this study strongly support the proposition that a formula that includes the seven listed variables can predict with 88.2%, 78.6%, 71.4%, and 66.0% accuracy whether a bank will fail within one year, two years, three years, or 4 years. The results also have important implications for potential bank investors and lenders who are considering putting money into a bank that has quantifiable risk of failure. In addition, due to the potential moral hazard issue, the FDIC has a strong incentive to identify early 24

26 those banks that offer high interest rates for deposits which will be covered by deposit insurance should the bank fail. If investors, lenders and regulators use our BankZ Score TM formula in a preemptive manner, they have the potential to identify banks with a high risk of failure and, in so doing, save much heartache and many dollars. References Agarwal, Vineet and Richard J. Taffler Twenty-five years of the Taffler Z-Score Model: Does It Really Have Predictive Ability? Accounting and Business Research. 37(4): Atlman, Edward I., Predicting Financial Distress of Companies: Revisiting The Z- Score and ZETA Models. July, 2000 Working Paper. Available at Aziz, M. Adnan and Humayon A. Dar Predicting Corporate Bankruptcy: Where We Stand? Corporate Governance. 2006, 6(1): Cole, Rebel A. and Jeffery W. Gunther Predicting Bank Failures: A Comparison of On- and Off-Site Monitoring Systems. Journal of Financial Services Research. 13:2: Ercan, Hakan and Ozgu Evirgen Predicting Bank Failures in Turkey by Discrete Choice Models. METU Studies in Development. 2009, 35(Special Issue): Jesswein, Kurt R An Examination of the Texas Ratio as a Bank Failure Model. Academy of Banking Studies Journal. 2009, 8(2): Jordan, Dan J, Douglas Rice, Jacques Sanchez and Donald H. Wort, Taking TARP Funds Can Be Hazardous To Your Bank s Wealth. (November 6, 2009). Available at SSRN: Mortgage Bankers Association, Delinquencies and Foreclosures Increase in Latest MBA National Delinquency Survey. March 6, Available at mortgagebankers.org/newsandmedia/presscenter/60619.htm 25

27 Mortgage Bankers Association, Delinquencies Continue to Climb in Latest MBA National Delinquency Survey. March 5, Available at mortgagebankers.org/newsandmedia/presscenter/68008.htm Mortgage Bankers Association, Delinquencies, Foreclosure Starts Fall in Latest MBA National Delinquency Survey. February 19, Available at mortgagebankers.org/newsandmedia/presscenter/71891.htm Ozkan-Gunay, E. Nur and Mehmed Ozkan Prediction of Bank Failures in Emerging Financial Markets: an ANN Approach. Journal of Risk Finance. 2007, 8(5): Pantalone, Coleen C. and Marjorie B. Platt Predicting Failure of Savings & Loan Associations. AREUEA Journal. Summer. 1987, 15(2): Pettway, Richard H. and Joseph F. Sinkey, Jr Establishing On-Site Bank Examination Priorities: An Early-Warning System Using Accounting and Market Information. The Journal of Finance. March. 1980, 35(1): Platt, Harlan D. and Marjorie B. Platt A Note on the Use of Industry-Relative Ratios in Bankruptcy Prediction. Journal of Banking & Finance. Dec. 1991, 15(6): Schaeck, Klaus Bank Liability Structure, FDIC Loss, and Time to Failure: A Quantile Regression Approach. Journal of Financial Services Research. 2008, 33: Shibut, Lynn, Tim Critchfield and Sarah Bohn Differentiating Among Critically Undercapitalized Banks and Thrifts. FDIC Banking Review. 2003, 15(2). Sinkey, Joseph F., Jr A Multivariate Statistical Analysis of the Characteristics of Problem Banks. The Journal of Finance. March, 1975, 30(1): Sinkey, Joseph F., Jr Identifying Problem Banks: How Do the Banking Authorities Measure A Bank s Risk Exposure? Journal of Money, Credit and Banking. May, 1978, 10(2): Spence, Michael Signaling in Retrospect and the Informational Structure of Markets. American Economic Review, 92 (3): doi: / Federal Reserve System Joint Troubled Company Subgroup, A Comparison of the Insurance and Banking Regulatory Frameworks for Identifying and Supervising Companies in Weakened Financial Condition. Working Paper, April 19, United States Treasury Department Capital Purchase Program Transactions, available at (last viewed May 25, 2009). 26

28 Federal Deposit Insurance Corporation Failed Bank List, available at (last viewed July 5, 2010). Yim, Juliana and Heather Mitchell Predicting Financial Distress in the Australian Financial Service Industry. Australian Economic Papers. Dec, 2007, 46(4):