Accounting Data And The Prediction Of Business Failure - The Setting Of Priors And The Age Of Data

In a recent issue of this Journal, both Casey [1980] and Zimmer [1980] reported the results of almost identical studies. In extending Libby's [1975] study on bankers' ability to predict corporate failure (loan default), both gave participating bankers a three-year series of financial ratios of real (but disguised) companies, half of which had failed, half of which had not. In contrast, Libby's subjects used ratios of a single year. The results of the studies of Casey and Zimmer were inconsistent (see table 1) in that Zimmer's Australian bankers achieved a mean prediction accuracy of 77.1%, whereas Casey's U.S. bankers achieved only 56.7%. This difference occurred even though it does not appear that the diagnostic power of the data for the sets of companies used by the two researchers was markedly different. According to Libby [1981, p. 126] the most likely reason for these inconsistent results is the fact that since the actual ratio of failed to nonfailed companies (50/50) used in the experiment was much higher than what the participants were used to (e.g., 1/99) the specification of that ratio to participants as part of the background information (as in the Zimmer study) should have resulted in significantly more accurate predictions than if no mention was made of the ratio (as in Casey's