Elsevier

Journal of Economics and Business

Volume 54, Issue 4, July–August 2002, Pages 361-387
Journal of Economics and Business

Predicting large US commercial bank failures

https://doi.org/10.1016/S0148-6195(02)00089-9Get rights and content

Abstract

The present study applies empirical methods to the problem of predicting large US commercial bank failures. Due to sampling limitations, scant research has examined the feasibility of using computer-based early warning systems (EWSs) to identify pending large bank failures. In the late 1980s and early 1990s numerous large banks failed in the US enabling us to collect adequate samples of large banks with more than $250 million in assets for empirical analyses. Both the parametric method of logit analysis and the nonparametric approach of trait recognition are employed to (1) develop classification EWS models based on original samples and (2) test the efficacy of these models based on their prediction accuracy using holdout samples. Both logit and trait recognition performed well in terms of classification results. However, with regard to the prediction results using holdout samples, trait recognition outperformed logit in most tests in terms of minimizing Type I and II errors. Other results from the trait recognition models reveal that complex two- and three-variable interactions between financial and accounting variables contain additional information about bank risk not found in the individual variables themselves. We conclude that computer-based EWSs can provide valuable information about the future viability of large banks.

Introduction

Seminal work by Beaver (1966) and Altman (1968) introduced computer-based models using accounting information to predict firm failure and sparked a continuing stream of research in the corporate financial literature (e.g., see Beaver, 1968, Edmister, 1972, Blum, 1974; Altman, Haldeman, & Narayanan, 1977; Martin, 1977, Ohlson, 1980; Zavgren, 1985, Jones, 1987; Keasey & McGuinnes, 1990; Platt & Platt, 1990; Altman, 1993; Coats & Fant, 1993; Altman, Marco, & Varetto, 1994; Altman & Narayanan, 1997, and others).2 One of the most important extensions of this literature is in the area of banking. Bank regulators are keenly interested in developing early warning systems (EWSs) to supplement information obtained from on-site examinations and, in turn, help predict impending bank failures. By doing so, regulatory intervention may prevent a bank failure or reduce the costs of failure. Extensive research on failed banks has confirmed that computer-based models perform well as EWSs (e.g., see Meyer & Pifer, 1970; Stuhr & Van Wicklen, 1974; Sinkey, 1975; Santomero & Vinso, 1977; Bovenzi, Marino, & McFadden, 1983; Korobrow & Stuhr, 1985; West, 1985, Maddala, 1986; Lane, Looney, & Wansley, 1986; Whalen & Thomson, 1988; Espahbodi, 1991, Thomson, 1991; Kolari, Caputo, & Wagner, 1996, and others).3

A major issue confronting bank regulators, analysts, and others is the prediction of large bank failure.4 In this regard, the on-going consolidation movement in the banking industry is creating an increasing number of large banks (see Berger, Kashyap, & Scalise, 1995; Boyd & Graham, 1996). The growing numbers and size of large institutions raise new policy challenges for regulatory and government entities charged with the responsibility of ensuring the safety and soundness and smooth functioning of the banking system. One policy response to the potential for too-big-to-fail (TBTF) dangers (including competitive inequalities, moral hazard problems, and inefficiency) is to increase bank regulation of large institutions (see Hoenig, 1999).5 While the development of computer-based EWSs for large banks is consistent with this regulatory policy, scant research exists on this topic due to inadequate sample sizes for research purposes. Previous work on anticipating large bank failures has focused on the usefulness of stock price data as a bank-specific EWS (e.g., see Pettway, 1976, Pettway, 1980; Peavy & Hempel, 1988), in addition to financial ratio profiles of individual case studies of large banks prior to failure (e.g., see Sinkey, 1985, Federal Deposit Insurance Corporation, 1997). To our knowledge no previous studies in banking examine EWSs for large bank failures developed from publicly available accounting and financial data, despite their obvious importance to bank regulatory practice and market participants.

In the late 1980s and early 1990s a surge of bank failures occurred in the United States due to regional economic difficulties. Focusing on the 1989–1992 period, and defining large banks to be greater than $250 million in total assets, we are able to collect a sample of 55 (60) large failed banks6 with data available up to (2) years prior to failure in this period as well as over 1,000 nonfailed large banks in each year. Although our large bank failure samples are substantial by historical standards, they are quite small in terms of minimum sampling requirements in most EWS models. It is common practice to split the sample of failed banks into (1) an original sample used to build a classification model and (2) a holdout sample reserved for prediction purposes to determine EWS model efficacy. In the present paper we use 18 large failed banks in 1989 and all other large banks in the US in that year to build the classification models, and holdout samples of large banks that failed in 1990, 1991, and 1992 are used to comparatively test the predictive power of both parametric and nonparametric EWS models. Due to its widespread application in previous finance and banking studies, the parametric approach of logit analysis was chosen. Also, we selected the nonparametric approach of trait recognition as applied to small bank failures in Kolari et al. (1996) due to its reported usefulness on small samples.7 Implicit in our analyses is the assumption that large banks are different than small banks and, consequently, should not be aggregated with smaller banks in developing EWS models.

In general, based on accounting data collected 1–2 years prior to the failure event, our results suggest that computer-based EWSs are a viable means of evaluating large bank failure risk. Both logit and trait recognition performed well in the classification results of original samples, with accuracy rates in the range of 95%–100%. However, with regard to the prediction results using holdout samples, trait recognition outperformed logit in terms of minimizing Type I errors (i.e., identifying a failed bank as nonfailing) and Type II errors (i.e., identifying a nonfailed bank as failing). In most cases the logit models did not perform better than chance; by contrast, trait recognition was able to clearly exceed chance in out-of-sample tests using both 1 and 2 years prior income and condition data of large banks. Other results from the trait recognition models reveal that complex two- and three-variable interactions between financial and accounting variables contain information about bank risk not found in individual variables. We conclude that computer-based EWSs can provide valuable information about the future viability of large banks.

The next section overviews logit and trait recognition EWS models, Section 3 describes our empirical methodology, Section 4 reports the empirical results, and the last section gives the summary and conclusions.

Section snippets

Parametric modeling approach—logit analysis

A logistic distribution is used in many limited dependent variable applications. The resulting model (i.e., a logit model) is common in the finance and banking EWS literature. The posterior probability of failure can be derived directly from the following logit specification:logPi1−Pi=a+b1Xj1+b2Xj2+⋯+bnXin,where Pi is the probability of bank i’s failure, and b=b1,…,bn is a vector of regression coefficients for predictor variables Xj (j=1,…,n). The logit model is preferred over the linear

Bank samples

Table 1 gives details of the original and holdout bank samples. Our analyses are restricted to the period 1989–1992 because sufficient numbers of large bank failures (i.e., banks closed by the FDIC) could be gathered for statistical purposes. At that time an economic downturn marked by falling energy prices and real estate prices caused numerous bank failures, especially in the southwest and northeast regions of the US Few bank failures occurred throughout the remainder of the 1990s due to the

Logit models

Table 3 reports the estimated logit models and classification results using 1–2 years prior to failure data for the 1989 original samples. Despite dropping the significance threshold from the default 0.10–0.30 in the stepwise logit procedure, only 4 out of 28 independent variables entered the model in the 1 year prior to failure run. However, these four variables were highly significant (at the 0.01 level) in the t-tests of mean differences between nonfailing and failing bank discussed above

Summary and conclusions

The present study empirically examined the efficacy of using computer-based EWS models to assess failure risk among large US commercial banks. Due to a lack of sufficient data points in the past, few studies have been published on this subject, with the exception of studies of stock prices and failure risk and case studies of individual large bank failures. In the late 1980s and early 1990s numerous large banks failed in the US which allowed us to gather samples of more than 50 large banks with

Acknowledgements

The opinions expressed here are those of the authors and do not necessarily represent those of the Office of the Comptroller of the Currency. The authors would like to thank participants in seminars at the Office of the Comptroller of the Currency, Federal Reserve Bank of Chicago, and Federal Reserve Bank of St. Louis for helpful comments, especially Larry Mote, Douglas Evanoff, Elijah Brewer, Catharine Lemieux, Julapa Jagtiani, James Moser, Mark Vaughn, Alton Gilbert, and Andrew Meyer. Also,

References (51)

  • Altman E. (1993). Corporate financial distress and bankruptcy (2nd ed.). New York:...
  • Altman, E. I., & Narayanan, P. (1997). Business failure classification models: An international survey. In F. Choi...
  • E.I. Altman et al.

    Credit risk measurement: Developments over the last 20 years

    Journal of Banking and Finance

    (1998)
  • Beaver, W. H. (1966). Financial ratios as predictors of failure. Empirical research in accounting: Selected studies,...
  • W.H. Beaver

    Market prices, financial ratios, and the prediction of failure

    Journal of Accounting Research

    (1968)
  • Benavidez, A., & Caputo. M. (1988). Pattern recognition of earthquake prone areas in the Andes. Revisita Geofisica,...
  • Berger, A. N., Kashyap, A. K., & Scalise, J. M. (1995). The transformation of the US banking industry: What a long,...
  • M. Blum

    Failing company discriminant analysis

    Journal of Accounting Research

    (1974)
  • M.M. Bongard et al.

    The use of self learning programs in the detection of oil containing layers

    Geology Geofiz

    (1966)
  • Bovenzi, J. F., Marino, J. A., & McFadden, F. E. (1983). Commercial bank failure prediction models. Federal Reserve...
  • Boyd, J. H., & Graham, S. L. (1996). Consolidation in US banking: Implications for efficiency and risk. Working paper...
  • P. Briggs et al.

    Pattern recognition applied to uranium prospecting

    Nature

    (1977)
  • P. Briggs et al.

    Pattern recognition applied to earthquake epicenters in California and Nevada

    Geological Society of America Bulletin

    (1977)
  • P. Coats et al.

    Recognizing financial distress patterns using a neural network tool

    Financial Management

    (1993)
  • R.A. Cole

    FIMS: A new monitoring system for banking institutions

    Federal Reserve Bulletin

    (1995)
  • Cited by (152)

    • Bank failure prediction models: Review and outlook

      2024, Socio-Economic Planning Sciences
    • EU-27 bank failure prediction with C5.0 decision trees and deep learning neural networks

      2022, Research in International Business and Finance
      Citation Excerpt :

      Tam and Kiang (1992) were the first authors to apply NN to bank failure prediction and found that NN outperformed other previously applied methods. Kolari et al. (2002) developed an early warning system based on logit and nonparametric trait recognition models for large US banks. Boyacioglu et al. (2009) examined NN, SVM, and multivariate statistical methods to predict the failure of Turkish banks.

    • Bootstrap-DEA management efficiency and early prediction of bank failure: Evidence from 2008-2009 U.S. bank failures

      2022, Central Bank Review
      Citation Excerpt :

      When compared, the predictions of trait recognition models are substantially accurate in one-year-ahead and two-years-ahead bank failure prediction. Kolari et al. (2002), apply logit analysis on a small sample of banks to predict large U.S. commercial bank failures and find the model allows bank failure prediction from one year to two years prior to the failure. Jordan et al. (2010) study bank failure risk using multiple discriminant analysis.

    View all citing articles on Scopus
    1

    The authors are Chase Professor of Finance, Texas A&M University, Senior Economist, Office of the Comptroller of the Currency, Assistant Professor, The University of Texas-Pan American, and Professor of Geophysics, Instituto di Fisica della Università di Roma, Rome, Italy, respectively.

    View full text