Determinants of Credit Union and Commercial Bank Failures: Similarities and Differences, Universit y of California, Berkeley

Transcription

1 Determinants of Credit Union and Commercial Bank Failures: Similarities and Differences, James A. Wilcox, PhD Kruttschnitt Family Professor of Financial Institutions Haas School of Business Universit y of California, Berkeley

2

3 Determinants of Credit Union and Commercial Bank Failures: Similarities and Differences, JAMES A. WILCOX, PhD KRUTTSCHNITT FAMILY PROFESSOR OF FINANCIAL INSTITUTIONS HAAS SCHOOL OF BUSINESS UNIVERSITY OF CALIFORNIA, BERKELEY

4 Copyright 2007 Filene Research Institute. ISBN All rights reserved. Printed in U.S.A.

5 Filene Research Institute The Filene Research Institute is a 501(c)(3), nonprofit organization dedicated to scientific and thoughtful analysis about issues affecting the future of consumer finance and credit unions. We support research efforts that will ultimately enhance the well-being of consumers and assist credit unions in adapting to rapidly changing economic, legal, and social environments. Deeply embedded in the credit union tradition is an ongoing search for better ways to understand and serve credit union members and the general public. Credit unions, like other democratic institutions, make great progress when they welcome and carefully consider high-quality research, new perspectives, and innovative, sometimes controversial, proposals. Open inquiry, the free flow of ideas, and debate are essential parts of the true democratic process. In this spirit, the Filene Research Institute grants researchers considerable latitude in their studies of high-priority consumer finance issues and encourages them to communicate their findings and recommendations. The Filene Research Institute is governed by an administrative board made up of the credit union industry s top leaders. Research topics and priorities are set by a select group of credit union CEOs called the Research Council. Additional research input is furnished by the Filene Research Fellows, a blue ribbon panel of academic and industry experts. The name of the institute honors Edward A. Filene, the father of the U.S. credit union movement. Filene was an innovative leader who relied on insightful research and analysis when encouraging credit union development. Since its founding in 1989, the Filene Research Institute has worked with over one hundred academic institutions and published over 150 research studies. Please visit our Web site at to peruse our research library and learn more about the Filene Research Institute s past, present, and future. Progress is the constant replacing of the best there is with something still better! Edward A. Filene iii

6 iv

7 Acknowledgments I would like to thank several people for their valuable comments and suggestions: Luis Dopico, Macrometrix; William Jackson, the University of Alabama; Bob Hoel and George Hofheimer, Filene Research Institute; and the NCUA and CUNA for their assistance in our data collecting. v

8 vi

9 Table of Contents Executive Summary and Commentary About the Author Chapter 1: Introduction Chapter 2: Failure Rates of Credit Unions and Commercial Banks Chapter 3: Key Findings Chapter 4: Summary and Implications Appendix 1: Literature Review Appendix 2: Methodology Appendix 3: Abbreviations References vii

10 List of Figures Figure 1 Failure Rates of Credit Unions and of Commercial Banks: Figure 2 Failure Rates of Credit Unions and of Commercial Banks: Figure 3 Failure Rates, Weighted by Assets, of Credit Unions and of Commercial Banks: Figure 4 Insurance Loss Rates for the NCUSIF and the BIF (FDIC): Figure 5 A Basic Model of Failure for an Extended Period: Figure 6 An Extended Model of Failure for a Shorter Period: Figure 7 An Extended Model of Failure by Subperiod Figure 8 An Extended Model of Failure by Asset Size: Figure 9 Average Characteristics of Credit Unions and Commercial Banks: Figure 10 Distributions of Estimated Probabilities of Failure (EPFs) of Small and Medium- Sized Credit Unions and Commercial Banks in 1990 and Figure 11 Distributions of Estimated Probabilities of Failure (EPFs) of Medium- Sized Credit Unions and Commercial Banks in Figure 12 Distributions of Estimated Probabilities of Failure (EPFs) of Small Credit Unions and Commercial Banks in Figure 13 Distributions of Estimated Probabilities of Failure (EPFs) of Small Credit Unions in 1990 and Figure 14 Distributions of Estimated Probabilities of Failure (EPFs) of Medium- Sized Credit Unions in 1990 and viii

11 Executive Summary and Commentary by George A. Hofheimer, Chief Research Officer You have become accustomed to reading Filene Research Institute studies on the future of the credit union charter, examining exciting innovations, new strategic frameworks, and emerging consumer behaviors. Why, then, you may be asking yourself, do we present this seemingly downbeat study that investigates past credit union and commercial bank failures? Good question. We believe studying poor financial institutional performance is a superb way to enlighten readers about the qualities of safe and sound credit unions and, for that matter, commercial banks. In this report, James A. Wilcox, Filene Research Fellow and Professor of Financial Institutions at the University of California, Berkeley, deftly presents the first large- scale, long- term ( ) econometric analysis of individual commercial bank and credit union failures. WHAT DID THE RESEARCHER DISCOVER? Using a newly constructed database of credit union and commercial bank failures, Professor Wilcox estimates and compares statistical models of failure of credit unions and commercial banks and makes the following key discoveries: The likelihood of failure rose at credit unions and commercial banks that had: Fewer assets. Higher ratios (to their assets) of net loans, commercial and industrial loans, delinquent loans, or noninterest expenses. Higher unemployment rates in their states. Lower returns on assets (ROA) or capital ratios. The estimated sizes and the significance of the effects of many of the failure factors differed: Between credit unions and commercial banks. Across asset size ranges. Over certain time periods. For instance, having more residential mortgages tended to increase the likelihood of credit union failures but reduce the likelihood of commercial bank failures. 1

12 Although the overall failure rates of credit unions were generally higher than those of commercial banks, credit unions failure rates were typically lower than those of commercial banks in the same asset size range. Over the past two decades, financial strength has improved much more at small banks (those with less than $25M in assets) than at small credit unions. Using an estimated probability of failure within one year (EPF) for individual institutions, Professor Wilcox finds that among small institutions (those with less than $25M in assets), the percentage of credit unions exceeding the 0.1% Basel ceiling (44%) was more than twice that of banks (21%). 1 By contrast, among medium- sized institutions (those with $25M $250M in assets), fewer credit unions (3%) than banks (9%) had estimated EPFs of more than 0.1%. PRACTICAL IMPLICATIONS One of the most interesting and practical conclusions of this study is that the behaviors and operating procedures that foretell credit union failures differ from those that foretell bank failures. While I doubt this is what industry watchers have in mind when they talk about the credit union difference, the implications are potentially profound. Wilcox promotes the idea of regulators (and policymakers by proxy) using this research to assess how altering balance sheet and income statement variables might reduce failure risk for individual credit unions and banks. Put another way, the figures and statistics presented in this study are potential raw materials for regulatory relief. In addition to utilizing this study as a tool to promote a more nuanced regulatory environment, small institutions (those with less than $25M in assets) and medium- sized institutions (those with $25M $250M) can delve into the contributing factors leading to failures in banks and credit unions to better understand (and manage) the early warning signs of institutional failure. While larger credit unions (those with more than $250M in assets) experienced no failures over the studied time period, they too can utilize this study to become aware of the factors that contribute to institutional failure. Finally, this seemingly downbeat subject has a silver lining beyond the two practical suggestions listed above. Over the entire study period, assets in failed credit unions and losses to the insurance fund comprised a small percentage of credit union assets and, most recently, failure rates at both credit unions and banks have declined significantly. 1 An EPF of 0.1% means a financial institution has a 0.1% chance of failure within the next year. This rate is the ceiling established by the Basel International Standard as acceptable. This topic is discussed in greater detail in Chapter 3. 2

13 About the Author James A. Wilcox is the Kruttschnitt Family Professor of Financial Institutions at the University of California, Berkeley s Haas School of Business. He is an inaugural member of the Filene Fellows. From 1999 to 2001 he served as chief economist at the Office of the Comptroller of the Currency. He has also served as senior economist for the President s Council of Economic Advisors, as an economist for the Board of Governors of the Federal Reserve System, and as Chair of the Finance Group at the Haas School. He received his PhD in economics from Northwestern University. At the Haas School, Professor Wilcox teaches courses on risk management at financial institutions, on financial markets and institutions, and on business conditions analysis. He has written widely on bank lending, credit markets, real estate markets, monetary policy, and business conditions. His research has addressed reform of deposit insurance, the causes and consequences of the Gramm- Leach-Bliley Act, the effects of mergers on bank executives, the ability of banks to reduce costs following mergers, the differences in bank supervision and regulation around the world, the effects of bank loan losses and capital pressure on lending and small businesses, demographic effects on residential real estate prices, and the efficiencies and credit effects of electronic payments. His articles have been published in leading academic journals, including the American Economic Review; the Journal of Finance; the Journal of Economic Perspectives; the Journal of Money, Credit and Banking; the Journal of Banking & Finance; the Journal of Housing Economics; and the Review of Economics and Statistics. 3

14 4

15 Chapter 1: Introduction Over the last few decades, the credit union has become a more prominent player in the U.S. financial industry. As credit unions share of assets among depository institutions has grown, the structure of the credit union industry has shifted from one dominated by thousands of very small institutions to one dominated by hundreds of much larger and faster- growing institutions. Despite years of fast growth among larger credit unions, the performance of many other credit unions has been less satisfactory, in some cases leading to their closure or failure. Credit unions growing market share and the diverging performance of small and large credit unions make studying credit union failures and comparing them with commercial bank failures increasingly important to a large audience, including: Credit union executives. Insured and uninsured creditors (depositors and holders of subordinated debt and other securities). Rating agencies. Insurers such as the National Credit Union Administration (NCUA) and the Federal Deposit Insurance Corporation (FDIC) the penultimate uninsured risk holders. State and federal regulators and legislators. Taxpayers the ultimate risk holders. Academics. Many policymakers are already very familiar with commercial banks but may welcome some of the updated information on bank failures included here. However, they are likely less familiar with credit unions, their growing importance, and their different failure patterns. Failures of commercial banks, thrifts, and credit unions have not been studied to an equal extent over the last several decades. Failures of commercial banks were the first to be studied using statistically sophisticated methods, and they continue to attract the most scholarly interest. Government supervisors of commercial banks have for years formally used models akin to those used in the academic literature. Interest in thrift failures exploded in the 1980s and early 1990s as the thrift crisis evolved. Many studies of thrifts in general, and in particular of their failures, consider the implications of agency theory 2 and test and compare the impact of mutual and stock ownership; these studies serve to highlight some of the possible differences between commercial banks and credit unions. Compared with 2 Agency theories posit that the management of a firm sometimes has interests that conflict with the interests of the owners. Some forms of organization (e.g., mutual vs. stock) are likely to be better at resolving these conflicts. 5

16 commercial bank and thrift failures, credit union failures have been studied only sporadically, and these studies have rarely used statistically sophisticated methods. Very few studies have addressed whether the factors that drive credit union failures differ from the factors that drive commercial bank failures. 3 Thus, this report presents the first large- scale, long- term ( ) study of credit union failures using statistically sophisticated methods (e.g., regressions using logit and ordinary least squares 4 ) that are commonly used in studies of commercial banks and thrifts. We used earlier studies as a springboard to compare the effects of mutual and stock ownership on the efficiency, portfolios of activities, and failure rates of financial institutions. We used data for individual institutions to perform the first in- depth comparison of failures and of the determinants of failures of credit unions and commercial banks. We found that the variables that have long been used to analyze commercial bank failures are also useful in analyzing credit union failures. The following variables were found to be statistically associated with failures of credit unions and commercial banks alike: Smaller asset size. Higher ratios (to assets) of net loans, commercial and industrial loans, provisions for loan losses, delinquent loans, and noninterest expenses. Higher state unemployment rates. Lower ratios of capital and returns on assets (ROA). However, we also found that while estimated models of institutional failure for credit unions are broadly similar to those for banks, there are substantial differences. Thus, most individual variables have coefficients 5 that are statistically different for samples of credit unions and samples of commercial banks. In particular, more residential mortgages appear to be associated with more credit union failures but fewer commercial bank failures. These differences require that the findings and conclusions drawn from one type of depository be applied to the other type only with caution. We also found that, based on credit union only or bank- only samples, the estimated sizes and significance of coefficients can vary across types of institutions, asset size ranges, and 3 For a detailed discussion of the literature on bank, thrift, and credit union failures, see Appendix 1. 4 Logit and ordinary least squares (OLS) are mathematical optimization techniques that attempt to discern the relationships between series of data. OLS is the most commonly used workhorse in econometrics. Logit is commonly used when a variable of interest takes a limited number of values (e.g., legally classified as a failed institution or not). 5 In statistical models, coefficients provide information about the relationship between two variables. For instance, a positive (negative) coefficient for residential mortgages in a regression for failures would imply that institutions with more residential mortgages have more (fewer) failures. 6

17 time periods. For instance, more unsecured loans are associated with more failures in small credit unions, but not in other samples of credit unions or banks. We also found that failures at medium- sized credit unions are not associated with several variables such as residential mortgages, noninterest expenses, and the state unemployment rate that affect the likelihood of failures in other samples. The declining frequency of credit union and bank failures in recent years may hamper our ability to predict the numbers or causes of such failures in the future. Among our estimated models, those that use more recent data (data that includes fewer failures) have lower R-squares and exhibit waning significance 6 for several variables such as asset size, residential mortgages, and the state unemployment rate that are significant in samples from earlier periods. Throughout this report, the size of both credit unions and commercial banks is defined as follows: Small institutions have less than $25 million (M) in assets. Medium-sized institutions have $25M $250M in assets. Large institutions have more than $250M in assets. All boundaries are in 2005 dollars. Please note that billions is abbreviated with B throughout this report. We found that failure rates and the percentage of institutions at a high risk of future failure are generally lower among credit unions, larger institutions and during the most recent periods. While credit unions in general have lower failure rates than commercial banks of the same size, in the last few years small credit unions have experienced higher failure rates than small commercial banks. We generated EPFs for individual institutions and totaled the institutions in excess of 0.1%, cited as the highest acceptable EPF in Basel standards. Despite much recent improvement, EPFs in 2006 were unacceptably high for twice as many small credit unions (44%) as small commercial banks (21%). In contrast, among mediumsized institutions, EPFs were unacceptably high for only one- third as many credit unions (3%) as commercial banks (9%). Our findings provide guidance on how individual institutions might reduce their risk of failure. They do not definitively answer whether credit unions or banks are inherently 6 In statistical models, significance is a measure of the reliability of a relationship between variables. R-squared is a commonly used measure of the overall reliability of a model. Ranging from 0 to 1, it states the fraction of the variation in the dependent variable (i.e., the one that we seek to explain) that is explained by the variation in the independent variables (i.e., the ones that we have posited might explain the dependent variable). 7

18 less prone to failure. Our findings imply that by altering their size, portfolio composition, expenses, capital, and other variables, both credit unions and banks can and do effectively choose their EPFs. Chapter 2 compares failure rates of credit unions and commercial banks. Chapter 3 presents the key findings of our analysis by reporting on: Estimated logit models of credit union and commercial bank failures. Data for the determinants of failures of credit unions and commercial banks. Estimated probabilities of failure of credit unions and commercial banks for different time periods and asset size ranges. Chapter 4 briefly summarizes the report. Finally, the appendices review existing literature on bank, thrift, and credit union failures and explain the methodology we used in our research. 8

19 Chapter 2: Failure Rates of Credit Unions and Commercial Banks We obtained aggregate and individual data for different types of failures of federally insured credit unions (FICUs) and federally insured commercial banks (FICBs) for from the NCUA and the FDIC (2006). 7 We obtained call report data for individual FICUs and FICBs for from the NCUA (2006), the Credit Union National Association (CUNA), and the Federal Reserve Bank of Chicago (FRBC 2006). Figure 1 presents failure rates and the number of failures of FICUs and FICBs across different time periods and asset size ranges. 8 We include an earlier subperiod of during which failure rates were higher, a later subperiod of during which failure rates were lower, and the full period of To calculate period failure rates, we first compute annual failure rates (i.e., the number of failures during one year relative to the number of institutions on December 31 of the previous year) and then we average those annual failure rates across the years included in each period. Figure 1 also includes the number of institutions at the beginning and end of the two subperiods. For each variable in the figure, we present values for all FICUs and all FICBs, and for institutions with less than $25M (small), $25M $250M (medium-sized), and more than $250M (large) in assets (all boundaries are in 2005 dollars). 9 Figure 1 shows that failure rates are broadly lower for larger institutions, for both credit unions and commercial banks. However, the link between larger size and fewer failures seems more pronounced among credit unions than commercial banks. Failure rates in among small, medium- sized, and large institutions averaged 0.56, 0.12, and 0% for credit unions and 0.64, 0.38, and 0.36% for commercial banks. Comparing across institutions in the same asset size range, credit unions broadly display lower failure rates than commercial banks (0.56% vs. 0.64% among small institutions, 0.12% vs. 0.38% among medium- sized institutions, and 0% vs. 0.36% among large institutions). However, since smaller institutions account for the majority of credit union charters and a small fraction of commercial 7 The NCUA database on FICU failures identifies the name and date for individual involuntary liquidations (and purchase and assumption transactions) since 1971, some assisted mergers in , and all assisted mergers since 1984, but not cases of assistance to avoid liquidation. Wilcox (2005) explains in detail some of the complexities and data issues involved in deciding what to count as a failure and in combining credit union and commercial bank failure data consistently. 8 Since we report failure rates across asset size ranges, in Figure 1 we use financial data for individual institutions experiencing failure and do not include data prior to Also, we do not include cases of government assistance to avoid liquidation for FICUs (for which data was not available for individual institutions) or commercial banks (for consistency). 9 These boundaries reflect (1) the absence of failures among larger credit unions, (2) the relative rarity of failures among medium- sized credit unions, and (3) the concentration of credit union failures among their smallest institutions. Wilcox (2005) presents results for credit unions across smaller asset size ranges, including less than $1M and $1M $10M in assets. 9

20 Figure 1: Failure Rates of Credit Unions and of Commercial Banks: FICUs FICBs All Small (less than $25M) Mediumsized ($25M $250M) Large (more than $250M) All Small (less than $25M) Mediumsized ($25M $250M) Large (more than $250M) Failure rate (%) Number of failures ,478 1, , ,709 1, , Number of institutions (December 31) ,324 16,274 1, ,329 3,043 9,796 1, ,317 9,709 2, ,960 1,355 7,938 1, ,014 5,915 2, , ,170 1,875 Sources: NCUA, CUNA, FDIC, and FRBC bank charters, the measure of the failure rate for all FICUs (i.e., unweighted by assets) is higher than that for all FICBs (0.49% vs. 0.42%) in Comparing the earlier high- failure and later low- failure subperiods, most of the patterns discussed above apply. In each subperiod, larger institutions fail less often, and credit unions typically have fewer failures than do commercial banks in the same asset size range. There are, however, some differences. The reduction in failure rates between and is far more pronounced in the sample of all FICBs (from 0.79% to 0.05%) than in the sample of all FICUs (from 0.77% to 0.18%). Falling from 1.19% to 0.10%, the failure rate for small commercial banks becomes lower than that for small credit unions, which falls from 0.85% to 0.23%. Figures 2 4 further elaborate on the differences between credit union and commercial bank failures and insurance losses (see also Wilcox 2005). Figure 2 presents the annual evolution of the failure rate among FICUs and FICBs in Annual failure rates are typically higher among FICUs than among FICBs during this period. Annual FICU failure 10 Since this figure does not break out failure rates across asset size ranges, we report it since 1971 instead of Since the NCUA releases annually the total number of cases of government assistance to avoid liquidation, in this figure we include these cases in the computation of failure rates for both FICUs and FICBs. 10

21 Figure 2: Failure Rates of Credit Unions and of Commercial Banks: Percent of institutions FICU failure rate (%) FICB failure rate (%) Sources: NCUA, FDIC rates averaged 0.84% and peaked at 2.67% in Annual FICB failure rates averaged 0.36% and peaked at 2.04% in Like Figure 1, Figure 2 shows a marked reduction in failure rates among both FICUs and FICBs from to As discussed above, comparing failure rates computed across all FICUs with those of all FICBs disguises the fact that failure rates are typically lower among FICUs than they are among FICBs in similar asset size ranges. While small institutions dominate the number of credit union charters (94% in 1980 and 66% in 2005), they account for a far smaller proportion of credit union assets (38% in 1980 and 6% in 2005). 11 Thus, Figure 3 presents the annual evolution in of a version of the failure rate weighted by assets the assets of institutions failing during one year divided by the assets of all institutions on December 31 of the previous year for both FICUs and FICBs. 12 Correcting for the large number of small credit unions that hold a small proportion of assets, Figure 3 reveals failure rates for credit unions that are either similar to or far lower than those for commercial banks in most years during this period. The annual FICU failure rate weighted by assets averaged 0.12% and peaked at 0.46% in The annual FICB failure rate weighted by assets averaged 0.22% and peaked at 1.30% in Figure 4 presents the annual evolution in of insurance losses per insured savings for both the federal insurer for credit unions the National Credit Union Share Insurance Fund (NCUSIF) and the federal insurer for commercial banks (and savings banks) the 11 Small credit unions account for similar fractions of credit union savings (deposits) and members (depositors). 12 Since this version of the failure rate uses data for individual institutions, the series begins in 1981 instead of 1971 and excludes cases of government assistance to avoid liquidation. 11

22 Figure 3: Failure Rates, Weighted by Assets, of Credit Unions and of Commercial Banks: Assets in failing FICUs per FICU assets (%) Percent of assets Assets in failing FICBs per FICB assets (%) Sources: NCUA, CUNA, FDIC, and FRBC Figure 4: Insurance Loss Rates for the NCUSIF and the BIF (FDIC): Percent of insured savings or deposits NCUSIF insurance losses (%) BIF (FDIC) insurance losses (%) Sources: NCUA, FDIC FDIC through 1989 and the Bank Insurance Fund (BIF) of the FDIC in Consistent with the findings of Figures 1 and 3, insurance losses were much larger at the BIF (FDIC) than at the NCUSIF both in absolute terms and per insured savings (deposits). 14 From 1971 to 2005, the BIF (FDIC) reported total insurance losses of $38B ($61B in 13 The FDIC ( ) and the BIF ( ) insure some savings banks, but not the thrifts (e.g., savings and loans) insured by the Federal Savings and Loan Insurance Corporation (FSLIC, ) and the Savings Association Insurance Fund (SAIF) of the FDIC ( ). 14 Wilcox (2005) also compares the annual evolution for both credit unions and commercial banks of different measures of insurance losses per assets in failing institutions during this period and finds them to be roughly similar. 12

23 2005 dollars). During this period, BIF (FDIC) annual insurance losses per insured deposits averaged 0.07% and peaked at 0.40% in In contrast, NCUSIF insurance losses totaled $1B ($1.5B in 2005 dollars). NCUSIF annual insurance losses per insured savings averaged 0.02% and peaked at 0.08% in The peak for NCUSIF insurance losses is only slightly higher than the mean for BIF (FDIC) insurance losses. 13

24 14

25 Chapter 3: Key Findings The following sections present (1) statistical models of failure in credit unions and commercial banks (i.e., what characteristics tend to be statistically associated with failure) and (2) the average characteristics of credit unions and commercial banks both across periods of time and across asset sizes. Next we combine those statistical models and the characteristics of individual institutions to generate estimates of the probabilities of failure for depository institutions with different charters and asset sizes and in various time periods. 15 STATISTICAL MODELS OF FAILURE IN CREDIT UNIONS AND COMMERCIAL BANKS ( ) Figures 5 8 present the results for logit regressions of failure on measures of the financial conditions for individual institutions and, in some cases, macroeconomic conditions. In each figure, independent variables are listed in the left- hand column and the results for different regressions are shown in the numbered columns. Figure 5 presents results for basic regressions with the few variables for which there is complete data for an extended period of time ( ). Figure 5 compares regressions for samples with credit unions only (column 1); commercial banks only (column 2); credit unions and commercial banks together, with a single set of coefficients for both types of institutions (column 3); and credit unions and commercial banks with interaction terms that reflect the difference in coefficients across the two types of institutions (column 4). Many of the results in Figure 5 are broadly consistent across credit unions and commercial banks. For instance, smaller asset size, more net loans, more provisions for loan losses, lower capital, and lower ROAs are statistically linked with higher likelihoods of failure for both types of institutions (both at the 1% level of significance). 16 However, not all results are consistent. For instance, more residential mortgages are associated with more failures among credit unions, but that variable has the opposite effect among commercial banks (each at the 1% level of significance). This divergence could mean that residential mortgages are among the riskiest assets that credit unions typically hold, compared with their other mainstays such as auto lending and government securities. In contrast, it appears that residential mortgages have historically been among the least risky assets held by commercial banks, compared with, for instance, commercial and industrial (C&I) loans. 15 For a complete discussion on our methodology, please see Appendix In statistical models, significance at the 1% level typically implies a very strong, or very reliable, relationship between two variables. Significance at the 5% level typically implies a strong, or reliable, relationship. Significance at the 10% level typically implies a weak, or weakly reliable, relationship. Failing to find significance at least at the 10% level typically implies that the relationship between two variables is not statistically strong and thus is not reliable. 15

26 Figure 5: A Basic Model of Failure for an Extended Period: Credit unions and Credit unions and Credit unions only Commercial banks only commercial banks commercial banks (1) (2) (3) (4) Constant 1.69*** 4.39*** 1.72*** 4.39*** (7.81) (7.66) ( 9.85) (7.66) Log real assets 0.49*** 0.37*** 0.16*** 0.37*** ( 31.15) ( 11.51) ( 16.09) ( 11.51) Net loans (%) 0.02*** 0.03*** 0.01*** 0.03*** (13.14) (10.82) (8.38) (10.82) Residential mortgages (%) Provisions for loan losses (%) 0.03*** 0.05*** 0.01*** 0.05*** (9.38) ( 10.08) (3.90) ( 10.08) 0.18*** 0.21*** 0.33*** 0.21*** (10.20) (11.36) (25.37) (11.36) Capital (%) 0.16*** 0.69*** 0.25*** 0.69*** ( 27.20) ( 46.84) ( 45.52) ( 46.84) ROA (%) 0.15*** 0.07*** 0.15*** 0.07*** ( 10.30) ( 5.48) ( 13.34) ( 5.48) CU Constant 2.69*** ( 4.40) CU Log real assets 0.12*** ( 3.46) CU Net loans (%) 0.01*** ( 4.25) CU Residential mortgages (%) CU Provisions for loan losses (%) 0.08*** (13.58) 0.03 ( 1.20) CU Capital (%) 0.54*** (33.87) CU ROA (%) 0.08*** ( 4.18) Number of observations 318, , , ,328 R-square For each variable, we present its logit coefficient and (in parenthesis) its t-statistic. *** represents significance at the 1% level. 16

27 Thus, while models of institutional failure in credit unions and commercial banks may be broadly similar, there are substantial differences that require that findings and conclusions regarding one type of institution be applied to the other type with caution. We performed a logit regression across both credit unions and commercial banks, forcing a single set of coefficients for both types of institutions (column 3), and we performed a Chow test 17 to confirm formally that the coefficients for one type of institution did not apply to the other. Since the test rejected the hypothesis that the same coefficients applied to both samples, we performed a regression (column 4) with interaction terms (variables obtained by multiplying each original variable by a dummy variable containing a value of 1 for credit unions and a value of 0 for commercial banks). 18 For instance, the product of the credit union (CU) dummy variable and ROA is entered as CU ROA. The t-statistics 19 for the interaction terms in column 4 tell us whether each variable has a significantly different effect on failure for credit unions and commercial banks. Our results imply that equal increases in asset size and ROA have a larger (positive) impact on the safety of credit unions than on commercial banks. Equal increases in net loans and capital have a larger (negative and positive) impact on the safety of commercial banks. Provisions for loan losses (one of our measures of credit risk) do not affect the failure rates of credit unions and commercial banks differently. Comparing the R-squares for credit unions and commercial banks, we find that our basic model explains a far larger proportion of the variation for commercial banks than for credit unions. This is consistent with the findings of Gordon et al. (1987) and Shafroth (1997) that many credit union failures occur for largely idiosyncratic reasons (such as sponsor failures and poor record- keeping) that are unrelated to the historical financial condition of the individual institution. Figure 6 presents the results for extended models using more variables (e.g., those that began to be reported at a later date, such as C&I loans and noninterest expenses), but for a shorter time period ( ). Thus, Figure 6 largely mimics Figure 5, except for the addition (or substitution 20 ) of variables. Below, we largely concentrate on analyzing the 17 Chow tests are a standard technique used to determine whether two samples of data may be combined into a single sample without losing important information. Chow tests are commonly used to determine the contours of a sample (or market). For instance, in studying the price of fruit, it might not be a good idea to lump together information for apples and oranges. However, combining information for apples sold in North Dakota and apples sold in South Dakota might be appropriate. 18 Thus, coefficients and t-statistics for the original variables in column 4 refer to commercial banks (and are identical to those in column 2), and the coefficients and t-statistics for the interaction terms reflect the difference between the coefficients for credit unions and commercial banks and whether that difference is statistically significant. 19 t-statistics are a measure of significance. For regressions with large samples, t-statistics above 2.57 imply significance at the 1% level, t-statistics above 1.96 imply significance at the 5% level, and t-statistics above 1.65 imply significance at the 10% level. 20 We performed regressions including, alternatively, provisions for loan losses and delinquent loans, and we found the results to be broadly similar. 17

28 variables that were not included in Figure 5. Again, many of the results are broadly consistent across credit unions and commercial banks (columns 5 and 6). More C&I loans, higher noninterest expenses, and a higher state unemployment rate are found to be statistically linked with a higher likelihood of failure for both types of institutions (in each case at the 1% level of significance). However, we also found some differences across the two types of institutions. More unsecured loans appear to be associated with more failures among credit unions (at the 5% level of significance), but the same variable has an insignificant effect among commercial banks. This divergence seems to indicate that unsecured loans are riskier than average assets at credit unions but not at commercial banks. Since a Chow test rejected the hypothesis that the same coefficients applied to both samples (column 7), we performed a logit regression with interaction terms (column 8). The difference between the coefficients for credit unions and commercial banks is statistically significant for most variables. Figure 7 compares models of failure for an earlier time period ( ) and a later time period ( ). 21 Some results are broadly consistent across time periods and types of institutions. More delinquent loans, less capital, lower ROA, and higher noninterest expenses are found to be statistically associated with a higher likelihood of failure in both periods and for both types of institutions (in each case at the 1% level of significance). However, we did find some differences across the two time periods. In general, more variables are found to be statistically significant and R-squares are substantially higher in the earlier period. Thus, we found residential mortgages and unsecured loans to be significant predictors of credit union failure in the earlier period, but not in the later period. We also found asset size, net loans, residential mortgages, and the state unemployment rate to be significant predictors of commercial bank failure in the earlier period, but not in the later period. The size of the impact of different variables on failure also changes significantly across the two periods. For instance, the coefficient for asset size for credit unions nearly doubled. While the coefficients for capital, ROA, and noninterest expenses are roughly similar for credit unions across time periods, these coefficients change dramatically for commercial banks (with that for capital falling and those for ROA and noninterest expenses increasing). These results back the argument that regulators should periodically revisit the variables included in their failure models and, unlike the practice today, update the coefficients in their models. However, these results also imply that despite changes and the need for further ongoing research, there has been a good deal of stability in the likely causes of historical failures. 21 We performed regressions with the basic model for a longer earlier period ( instead of ) and found similar results. We also performed but do not report regressions without and with interaction terms and Chow tests for each group of regressions that we compare informally in Figure 7. 18

29 Figure 6: An Extended Model of Failure for a Shorter Period: Credit unions Commercial banks Credit unions and Credit unions and only only commercial banks commercial banks (5) (6) (7) (8) Constant 1.73*** 1.38* 3.46*** 1.38* ( 4.08) (1.92) ( 11.04) (1.92) Log real assets 0.32*** 0.27*** 0.15*** 0.27*** ( 12.91) ( 7.11) ( 9.11) ( 7.11) Net loans (%) 0.01*** ** 0.01 (6.21) (1.21) (2.51) (1.21) Residential mortgages (%) 0.02*** 0.02*** *** (4.62) ( 3.43) (0.63) ( 3.43) C&I loans (%) 0.04*** 0.02*** 0.05*** 0.02*** (6.47) (3.33) (17.53) (3.33) Unsecured loans (%) 0.004** *** (2.35) (0.67) (2.66) (0.67) Delinquent loans (%) 0.09*** 0.20*** 0.13*** 0.20*** (22.52) (16.97) (31.42) (16.97) Capital (%) 0.19*** 0.64*** 0.32*** 0.64*** ( 24.08) ( 37.29) ( 41.95) ( 37.29) ROA (%) 0.17*** 0.06*** 0.16*** 0.06*** ( 10.35) ( 4.32) ( 14.35) ( 4.32) Noninterest expenses (%) 0.21*** 0.08*** 0.19*** 0.08*** (12.59) (3.58) (15.76) (3.58) State unemployment rate (%) 0.06*** 0.19*** 0.17*** 0.19*** (2.60) (7.45) (11.67) (7.45) CU Constant 3.10*** ( 3.73) CU Log real assets 0.04 ( 0.98) CU Net loans (%) 0.01 (1.63) CU Residential mortgages (%) 0.04*** (5.45) CU C&I loans (%) 0.02*** (2.95) 19

30 Figure 6: An Extended Model of Failure for a Shorter Period: (continued) Credit unions only (5) Commercial banks only (6) Credit unions and commercial banks (7) Credit unions and commercial banks (8) CU Unsecured loans (%) (0.15) CU Delinquent loans (%) 0.10*** ( 8.37) CU Capital (%) 0.45*** (23.82) CU ROA (%) 0.11*** ( 4.94) CU Noninterest expenses (%) 0.13*** (4.49) CU State unemployment rate (%) 0.13*** ( 3.95) Number of observations 221, , , ,148 R-square For each variable, we present its logit coefficient and (in parenthesis) its t-statistic. * represents significance at the 10% level. ** represents significance at the 5% level. *** represents significance at the 1% level. We did not design our study to formally test whether the Federal Deposit Insurance Corporation Improvement Act (FDICIA) substantially changed the failure process of commercial banks. However, one could informally review Figures 1 and 7 and attempt to correlate the passage of FDICIA, which is binding on commercial banks but not on credit unions, 22 with changes in failure rates and models of the failure process before and after FDICIA. The results of our study, however, have mixed implications for the import of FDICIA. Whereas a summary review of Figure 7 might lead one to conclude that models of the failure process have changed more for commercial banks than for credit unions, both sets of models have changed substantially. A sharp drop in failure rates among commercial banks coincides with the passage and implementation of FDICIA, but failure rates fall in a similar fashion among credit unions, which are not subject to FDICIA. 22 Capital standards roughly similar to those for commercial banks became binding for credit unions only much later (with the passage of the Credit Union Member Access Act of 1998) and were not introduced in response to high failure rates but as part of legislative compromises following a Supreme Court decision regarding credit union fields of membership (see Wilcox 2002 and 2005). 20

31 Figure 7: An Extended Model of Failure by Subperiod Credit unions (9) (10) Constant 2.35*** 0.07 ( 4.74) ( 0.07) Commercial banks (11) (12) 1.40* 6.70*** (1.89) ( 2.66) Log real assets 0.24*** 0.44*** 0.27*** 0.06 ( 8.38) ( 8.34) ( 6.76) ( 0.45) Net loans (%) 0.01*** 0.02*** 0.01** 0.02 (3.37) (3.74) (2.22) (1.44) Residential mortgages (%) 0.02*** *** 0.01 (5.01) (1.47) ( 3.40) ( 0.32) C&I loans (%) 0.04*** 0.04*** ** (5.81) (2.62) (1.64) (2.46) Unsecured loans (%) 0.004** (1.96) (0.72) ( 0.59) (0.75) Delinquent loans (%) 0.10*** 0.09*** 0.21*** 0.16*** (20.00) (9.50) (15.85) (4.29) Capital (%) 0.17*** 0.18*** 0.62*** 0.32*** ( 17.34) ( 11.87) ( 33.94) ( 5.00) ROA (%) 0.20*** 0.14*** 0.05*** 0.40*** ( 10.37) ( 4.49) ( 3.74) ( 5.77) Noninterest expenses (%) 0.24*** 0.19*** 0.07*** 0.21*** (12.27) (5.67) (2.95) (3.73) State unemployment rate (%) *** 0.07 (0.52) ( 0.53) (6.19) ( 0.67) Number of observations 94, ,981 88,671 96,486 R-square For each variable, we present its logit coefficient and (in parenthesis) its t-statistic. * represents significance at the 10% level. ** represents significance at the 5% level. *** represents significance at the 1% level. 21

32 Figure 8 compares models of failure for samples of small, medium- sized, and large credit unions and commercial banks. 23 Some results are broadly consistent across asset sizes and types of institutions. More delinquent loans and less capital are found to be statistically associated with a higher likelihood of failure for all asset sizes and both types of institutions (in each case at the 1% level of significance). However, we also found many differences across asset sizes. For instance, C&I loans have a significant effect on failures among all the institutions shown except small and large commercial banks. In general, we found many of our variables to be insignificant or less significant among larger institutions. Thus, residential mortgages and noninterest expenses are not significant predictors of failure among medium- sized credit unions and large commercial banks. Unsecured loans are a significant predictor of failure only among small credit unions (i.e., among the institutions that engage the most in unsecured lending; see Figure 9). The state unemployment rate is insignificant in predicting failures only among medium- sized credit unions. Overall, these results should encourage regulators to take different variables into account when considering the likelihood of failure for smaller and larger institutions. AVERAGE CHARACTERISTICS OF CREDIT UNIONS AND COMMERCIAL BANKS ( ) Figures 5 8 present the coefficients (the betas) for variables in regressions modeling failures of credit unions and commercial banks. Figure 9 presents the averages of the values in those variables (the Xs). Columns 1 and 2 list averages for both FICUs and FICBs. Columns 3 8 list averages for small, medium- sized, and large FICUs and FICBs. Row 1 presents average asset sizes in 2005 dollars. In rows 2 10, variables are presented as a percentage of assets. For each variable, we list averages for three periods: an earlier subperiod (beginning with the earliest year since 1980 for which data was available for both credit unions and commercial banks and ending in 1992), a later subperiod ( ), and the two subperiods together. Period averages were computed by first computing annual averages (weighted by assets for all variables except assets) and then computing and reporting simple arithmetic averages across the years included in each period. Presenting Figures 5 8 and Figure 9 together highlights the difficulties in trying to assess whether one type of institution is inherently riskier than the other. Our results point out what types of activities might make credit unions or commercial banks more prone to 23 We performed regressions with the basic model for a longer period ( instead of ) and found similar results. We performed but do not report regressions without and with interaction terms and Chow tests for each group of regressions that we compare informally in Figure 8. We also performed, but do not report, regressions for small credit unions and commercial banks separately and together without and with interaction terms in the earlier and later periods; regressions for medium-sized credit unions and commercial banks separately and together without and with interaction terms in the earlier and later periods; and regressions for large commercial banks in the earlier and later periods. 22

33 Figure 8: An Extended Model of Failure by Asset Size: Credit unions Commercial banks Small Medium-sized Small Medium-sized Large (less than $25M) ($25M $250M) (less than $25M) ($25M $250M) (more than $250M) (13) (14) (15) (16) (17) Constant 2.57*** * 5.13** ( 5.41) (0.64) (1.42) (1.74) (2.09) Log real assets 0.25*** 0.69** 0.63** 0.36*** 0.33*** ( 8.75) ( 1.96) ( 2.49) ( 3.88) ( 2.93) Net loans (%) 0.01*** *** 0.02*** 0.05*** (5.14) (1.30) (3.32) (2.68) ( 4.39) Residential 0.02*** ** 0.02** 0.02 mortgages (%) (4.54) (1.02) ( 2.19) ( 2.25) ( 0.98) C&I loans (%) 0.04*** 0.05*** *** 0.01 (4.05) (2.89) (0.34) (3.48) (0.79) Unsecured 0.01*** loans (%) (2.82) ( 0.10) ( 0.56) (0.17) (1.27) Delinquent 0.09*** 0.21*** 0.18*** 0.19*** 0.25*** loans (%) (22.15) (4.46) (7.60) (12.57) (7.62) Capital (%) 0.18*** 0.29*** 0.60*** 0.68*** 0.52*** ( 22.41) ( 6.85) ( 17.89) ( 30.05) ( 11.08) ROA (%) 0.15*** 0.46*** *** 0.12** ( 8.94) ( 5.67) (0.01) ( 4.19) ( 2.53) Noninterest 0.22*** ** 0.06** 0.04 expenses (%) (13.39) (0.82) (2.01) (2.04) (0.78) State 0.05** *** 0.18*** 0.19*** unemployment (2.09) (0.73) (4.25) (5.69) (3.10) rate (%) Number of observations 172,218 43,962 22, ,862 30,635 R-square Note: Since there were no failures of credit unions with more than $250M in assets in , we do not include a column for that asset size. For each variable, we present its logit coefficient and (in parenthesis) its t-statistic. * represents significance at the 10% level. ** represents significance at the 5% level. *** represents significance at the 1% level. 23