Using Inductive Learning to Predict Bankruptcy

Abstract

An emerging trend in organizational computing is using information technology to learn decision knowledge from enterprise data. The primary contribution of this study is the presentation of a sound theory and a comprehensive technique for learning the decision model for predicting bankruptcy. The theory is based on the information contained in cash flow components, which is the foundation of valuation theory, and an analytical system that measures the amount of uncertainty in the cash flow information. The approach links a tree-based inductive learning system that relies on the concept of entropy, with an information system based on the cash flow of a firm. A test of the cash flow approach involves the cash flow components for a sample of 99 failed and 99 non-failed companies. The structural instability of cash flow components generated by an inductive learning system is a serious issue for financial analysts. However, this shortcoming is overcome by using a jackknife procedure to develop a global tree that identifies the most important cash flow components. The final global tree found only 3 cash flow components were needed to classify correctly 89% of the companies as either failed or non-failed. Only a few early studies achieved a higher level of predictive accuracy. The 3 significant cash flow components were dividends, net investment, and net operating cash flow. Using the same data, a probit statistical technique generated a 67.5% predictive accuracy. In summary, the inductive learning results indicate that cash flow components are not only a natural tool for explaining the bankruptcy process, but they provide a high level of predictive accuracy.