Eugene Kaciak, Brock University, Canada

Poor loan quality, mis-management and fraud have, among other things, led to an increased number of bank failures in newly emerging Central and Eastern European free market economies in the early 1990s. In this paper we propose a computer model for predicting bank failures based on a multivariate discriminant analysis algorithm. Because of the impact that banks have on the economy as a whole the early identification of potential bank failures is of utmost importance for bank regulatory agencies. A warning system could provide regulators with early identification of problem banks.

Many studies in the past have investigated the predictability of bank (or company) failure [see e.g. Altman 1968, Meyer and Pifer 1970, Deakin 1972, Blum 1974, Libby 1975, Sinkey 1975, Eisenbeis 1977, Altman 1977, Martin 1977, Altman and Eisenbeis 1978, Ohlson 1980, Bovenzi, Marino and McFadden 1983, Korobow and Stuhr 1985, Lane, Looney and Wansley 1986, Palepu 1986, Casey, McGee and Stickney 1986, Dopuch, Holthausen and Leftwich 1987, Espahbodi, 1991].

In each study a number of measures describing banks' (or companies') performance in the past were analyzed. Based on the results of the analysis potential bank (or company) failures were predicted. Among the most popular multivariate statistical methods used in the studies were: multiple discriminant analysis (both in linear and quadratic forms), the probit and the logit models.

The above multivariate statistical techniques allow bank regulators to:

The data used in bank failure studies usually consist of two samples: a sample of failed banks and a sample of non-failed banks in a given year. Because of usually low numbers of bank failures in a given country in a given year researchers put into the first (failed) sample as many banks as they can get data on. Hence this is not a randomly selected sample. On the other hand, the second (non-failed) sample is randomly selected from a population of non-failed banks in such a way that the non-failed banks match the failed banks as to their size, location, category, etc. The sizes of the two samples are very similar. Usually one failed bank is matched with a corresponding non-failed bank randomly selected from a given subpopulation of similar banks.

One can find in the literature the following financial ratios that have been used as measures describing banks' performance:

- cash and treasury securities/total assets

- total loans/total assets

- reserve for possible loan losses/total loans

- total loans/total equity capital

- total operating expense/total operating income

- total loan revenue/total operating income

- interest income on treasury securities/total operating income

- interest income on state and local government obligations/total operating income

- interest paid on deposits/total operating income

- other expenses/total operating income

- total time and savings deposits/total demand deposits

- total operating income/total assets

- real estate loans/total income

- total liabilities/total assets

- inventory/total assets

- book value of total assets

In our study, the analysis was conducted in the following manner:

One could use in this case a well known statistical package, such as SPSS for Windows or SAS. However, these and other packages are not very user-friendly, and therefore they are used mainly by researchers. Bank regulators need a rather user friendly computer program in order to perform required analyses. Such a multivariate discriminant analysis computer program has been developed by the author of this study and applied to real life data.

Now, we will describe the above steps in detail.

Steps 1 and 2.

Due to bank secrecy laws we can not reveal the number of failed banks in the country. We can only say the total number of all banks investigated was 84.

Step 3.

In this study we had to take into account the latest laws and regulations, accounting systems and procedures, etc. governing the country's banking system. Based on the detailed information we have prepared a preliminary list of 87 financial ratios. Out of these 87 ratios the following 85 have been retained for further analysis (2 ratios have been rejected due to lack of data):

1. Net Income/Total Expenses

2. Gross Income/Total Expenses

1. (Rejected, due to lack of data)

2. [17. - 28.] The same as in [3. - 14.]. OF substituted for TA.

Let: Total Operating Income = (Loan Income + Other than Loan Income) = TOI

1. Loan Expenses/Working Assets

2. Other than Loan Expenses (net)/Working Assets

3. (Rejected, due to lack of data)

4. Total Operating Income/TA

Let: Manufacturing and Services Sectors = MSS

[It is not our intention to describe each of the above ratios with greater detail and precision. The above list should serve only as an indication of the complexity of financial ratios investigated rather than precise description of accounting terms currently used in the analyzed banking system.]

Step 4.

A multidimensional scaling map of the 85 measures is presented in Figure 1 (Available from Author). Two black areas on the left and on the right depict those measures that have similar values across all 84 banks, and therefore they can be removed from further analysis as irrelevant. We kept only two measures as their "representatives": #66 (black area on the left) and #28 (black area on the right). The remaining measures retained for further analysis are clearly visible on the map. Going from the left to the right these are the measures: 66, 30, 70, 29, 3, 82, . . . , 49, 42, 47, and 28. One can see that multidimensional scaling allowed us to remove 35 redundant variables: only 50 were kept for further analysis.

Step 5.

A custom-made and, in our opinion, a user-friendly multivariate discriminant analysis computer program may be made available upon request.

Step 6.

The results of the multivariate discriminant analysis are as follows. The model has found eight variables (out of the 50) that discriminate best between failed and non-failed banks. The variables are:

The multiple discriminant analysis model has therefore the following form:

Z = (1.8601)Var1 + (-2.008)Var6 + (4.7043)Var9 + (-1.263)Var11 + (1.4248)Var46 + (2.3574)Var56 + (-1.656)Var58 + (4.4307)Var62,

where Z = Discriminant Score.

The computer program calculates the value of Z for each bank. The value of Z < 0 indicates a possibility of the bank failure in the nearest future. The classification accuracy in this case was very high: 95.7%.

The interpretation of the above discriminant loadings is the following: the higher is the score a bank achieves for a variable with a positive (negative) discriminant loading the more unlikely (likely) is the bank to fail in the nearest future. Notice that negative discriminant loadings relate to variables that depend on the ("negative") measures such as: operating expenses, salaries and bad loans. High values of these three factors seem to be the most crucial in shaping the negative future of the bank. On the other hand, high values of such ("positive") factors as net income, loan revenue, commercial paper, non-banking operations, and own funds will give more confidence to this bank performance.

Finally, we would like to mention that the above methods could also be used in predicting failures of companies in other sectors of the economy. The importance of these methods to banks' credit departments, insurance companies, etc. could be enormous - they could assist loan officers in evaluating credit worthiness of potential corporate clients, or help insurance agents in determining insurance fees, etc.

Altman E. I. 1968, Financial ratios: Discriminant analysis and the prediction of corporate bankruptcy, The Journal of Finance, Sept., 589-609

Altman E. I. 1977, Predicting performance in the savings and loan association industry, Journal of Monetary Economics, 443-466

Altman E. I. and R. A. Eisenbeis 1978, Financial applications of discriminant analysis: A clarification, Journal of Finance and Quantitative Analysis, March, 185-195

Blum M. 1974, Failing company discriminant analysis, Journal of Accounting Research, Spring, 1-25

Bovenzi J. F., J. A. Marino and F.E. McFadden 1983, Commercial bank failure prediction models, Economic Review (Federal Reserve Bank of Atlanta), Nov., 14-26

Casey C. J., V. E. McGee and C. P. Stickney 1986, Discriminating between reorganized and liquidated firms in bankruptcy, The Accounting Review, April, 249-262

Deakin E. B. 1972, Discriminant analysis of predictors of business failure, Journal of Accounting Research, Spring, 167-179

Dopuch N., R. W. Holthausen and R. W. Leftwich 1987, Predicting audit qualifications with financial and market variables, The Accounting Review, July, 431-454

Eisenbeis R. A. 1977, Pitfalls in the application of discriminant analysis in business, finance, and economics, The Journal of Finance, June, 875-900

Espahbodi P., 1991, Identification of problem banks and binary choice models, Journal of Banking and Finance, 15, 53-71

Korobow, L. and D. Stuhr 1985, Performance measurement of early warning models, Journal of Banking and Finance, 9, 267-273

Lane, W. R., S.W. Looney and J.W. Wansley 1986, An application of the Cox proportional hazards model to bank failure, Journal of Banking and Finance, 10, 511-531

Libby R. 1975, Accounting ratios and prediction of failures: Some behavioral evidence, Journal of Accounting Research, Spring, 150-151

Martin D. 1977, Early warning of bank failure - A logit regression approach, Journal of Banking and Finance, 249-276

Meyer P. and H. Pifer 1970, Prediction of bank failures, The Journal of Finance, Sept., 853-868

Ohlson J. A. 1980, Financial ratios and the probabilistic prediction of bankruptcy, Journal of Accounting Research, Spring, 109-131

Palepu K. G. 1986, Predicting takeover targets: A methodological and empirical analysis, Journal of Accounting and Economics, March, 3-35

Sinkey J. F. Jr.1975, A multivariate statistical analysis of the characteristics of problem banks, The Journal of Finance, March, 21-36