Multilogit approach to predicting corporate failure--Further analysis and the issue of signal consistency

This paper has two purposes. The first purpose is to examine how far the results of Peel and Peel [4] are replicable on a different data set. For the bi-nomial models, we obtain approximately similar within sample results to those of Peel and Peel (approximately 90% predictive accuracy one year prior to failure, to 70% predictive accuracy three years prior to failure). For the multilogit models we also note similar results to those of Peel and Peel; high predictive accuracy for healthy firms and failing firms one year prior to failure, poor predictive accuracy for failing firms for data two or more years prior to failure. The second, and more important purpose of the paper is to examine the whole issue of dating failure in more depth. Although Peel and Peel admirably brought the technique of multilogit analysis to the problem of predicting failure, the issues involved with using it to date failure were given little discussion. We discuss the role of signal consistency across various years of data when dating failure and then examine, on a case by case basis, how well a decision-maker could date failure for the present data set. Whilst noting that the notion of signal consistency is far from straightforward, we conclude for the present data set that healthy firms seem to give consistent patterns of signals. We further conclude that whilst this is not generally true of failing firms, a decision-maker could adopt a rule of thumb that would allow the successful dating of failures that occur in the near future. For more distant failures, we find dating would be a far more haphazard process.