| Popis: |
Bankruptcy is a costly event. Holders of publicly traded securities can rely on security prices to reflect their risk. Other stakeholders have no such mechanism. Hence, methods for accurately forecasting bankruptcy would be valuable to them. A large body of literature has arisen on bankruptcy forecasting with statistical classification since Beaver (1967) and Altman (1968). Reported total error rates typically are 10%-20%, suggesting that these models reveal information which otherwise is unavailable and has value after financial data is released. This conflicts with evidence on market efficiency which indicates that securities markets adjust rapidly and actually anticipate announcements of financial data. Efforts to resolve this conflict with event study methodology have run afoul of market model specification difficulties. A different approach is taken here. Most extant criticism of research design in this literature concerns inferential techniques but not sampling design. This paper attempts to resolve major sampling design issues. The most important conclusion concerns the usual choice of the individual firm as the sampling unit. While this choice is logically inconsistent with how a forecaster observes financial data over time, no evidence of bias could be found. In this paper, prediction performance is evaluated in terms of expected loss. Most authors calculate total error rates, which fail to reflect documented asymmetries in misclassification costs and prior probabilities. Expected loss overcomes this weakness and also offers a formal means to evaluate forecasts from the perspective of stakeholders other than investors. This study shows that cost of misclassifying bankruptcy must be at least an order of magnitude greater than cost of misclassifying nonbankruptcy before discriminant analysis methods have value. This conclusion follows from both sampling experiments on historical financial data and Monte Carlo experiments on simulated data. However, the Monte Carlo experiments reveal that as the cost ratio increases, robustness of linear discriminant rules improves; performance appears to depend more on the cost ratio than form of the distributions. |
