Academic journal article Journal of Business and Entrepreneurship

On Altman's Failure/nonfailure Model: A Comparison of Discriminant, Logit, Nearest Neighbor and Neural Net Models*

Academic journal article Journal of Business and Entrepreneurship

On Altman's Failure/nonfailure Model: A Comparison of Discriminant, Logit, Nearest Neighbor and Neural Net Models*

Article excerpt


The underlying premise in most failure studies is that there exists a group of accounting and financial ratios of firms which distinguishes failed firms from nonfailed firms. Ex post facto classification analysis of OTC market data one year, two years, three years, and four years prior to failure was used. Models incorporating Altman's variables were developed using three statistical procedures-discriminant, logit, and nearest neighbor-and an artificial neural network. Results indicate that the statistical models have low predictive power as compared with the artificial neural network model.


In the present economy of the United States, small businesses play a very influential role. They greatly outnumber large business firms. Whereas a large number of small businesses enter the economy each year, many leave the market place either by way of failure or by mergers and acquisitions. The sheer number of small businesses functioning in the economy stimulates attention as to their functionality; and, more so, does their failure. Accordingly, in the last two decades, considerable interest has been shown by researchers trying to develop mathematical models for predicting business failure.

It has been indicated by Beaver, Kennelly, and Vass (1968) that accounting data have predictive ability which could be used for several purposes; for example, prediction of financial performance and accounting earnings (Kinney, 1971; Lawson, 1971; Patell, 1976; Chen & Shimerda, 1981), predicting security market behavior (Brown & Niederhoffer, 1968; Foster, 1973; Lev, 1979; Kross & Schroeder, 1984), and predicting business failure (Beaver, 1967,1968; Airman, 1968,1971,1973; Airman & Spivack, 1983). Most of this research has dealt with firms in general, without devoting specific attention to small business. Researchers have used financial and accounting ratios to develop models used for purposes of predicitng business failures. A more recent paper (Jones, Cronan, & Stettler, 1988) developed several classification models for data two years prior to failure for OTC firms. Their models exhibited only modest predictive ability; however, they did not investigate the ability of Altaian's widely cited variables (Airman, 1968) to predict failure/nonfailure.


Research Objective

The purpose of this study was to develop a model which incorporates Airman's variables (Airman, 1968) to discriminate failed and nonfailed firms in the OTC market.1 Directly related to this objective is the sensitivity of the classification rates to the procedure employed and, of course, the external validity of these variables to a population different from that considered by Airman2.

Sample Selection

A paired sample design matching each failed firm with a nonfailed firm was used in the study. Firms were matched according to Moody's Industrial Classification and total asset size with the additional requirements that the paired firms existed during the same time frame and published complete financial data. Lacking a comprehensive list of failed firms, identification of the sample of these firms was verified through several sources-articles in professional journals, news reporting magazines, and Moody's OTC Industrial Manual.

After the failed firms were selected and grouped according to Moody's Industrial Classification, the nonfailed firms (for which complete financial data were available) were selected that had the closest total asset sizes during the same time frame.3 Data up to four years prior to failure for 176 firms during the time period 1967-1983 were collected.

Variable Selection

Since a focus of the present study was to investigate the external validity of Airman's variables, a variable selection technique was not utilized to identify a set of "best" predictor variables. The variables used in the model, except for one surrogate, were those suggested by Airman (Airman, 1968). …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.