Verbesserung der Vergleichbarkeit von Schätzgüteergebnissen von Insolvenzprognosestudien (German version of 'Improving the comparability of insolvency predictions')

This working paper aims at improving the comparability of forecast quality measures of insolvency prediction studies. For this purpose, in a first step commonly used accuracy measures for categorial, ordinal and cardinal insolvency predictions are presented. It will be argued, that ordinal measures are the most suitable measures for sample spanning comparisons concerning predictive power of rating models, as they are not affected by sample default rates. A method for transforming cardinal into ordinal accuracy measures is presented, by which comparisons of insolvency prediction results of older and present-day studies are enabled. In the second part of the working paper an overview of influencing variables – aside from the quality of the insolvency prediction methods – is given, which affect the accuracy measures presented in the first part of the paper and thus impair sample spanning comparison of empirically obtained forecast quality results. In this context, methods for evaluating information losses that are attributable to the discretization of continuous rating scales or preselection of portfolios are developed. Measure results of various insolvency prognosis studies are envisaged and compared with three benchmarks. First benchmark is the accuracy that can be achieved solely by taking into account legal status and industry classification of corporations. The second benchmark is the univariate prognosis accuracy of single financial ratios. As third benchmark, ALTMAN’s Zscore model is examined, a multivariate insolvency prediction model, that is currently used as reference rating model in many empirical studies. It turns out, however, that the Z-score’s forecast quality is so discontenting, that its application is not recommendable. Instead it is suggested to use those rating models that are cited in this discussion paper, which are fully documented and which therefore can be rebuilt and directly applied to any desired data sample. If applied to the respective target groups, their performance matches with the performance of commercial rating systems, like bureau and business scores for rather small companies, middle market rating models for SMB, or agency ratings for large public companies.