On Small Sample Prediction of Financial Crisis

Abstract

Prediction of financial crisis is a challenging problem in financial research. On the basis of the information provided by financial statements, companies are usually classified into two groups, e.g., the groups of solvent and insolvent companies. Linear discriminant analysis (LDA), logistic regression and artificial neural network (ANN) are the most common statistical tools used for this classification. LDA and logistic regression separate the two groups using a hyperplane, and they provide good lower dimensional view of class separability. However, these methods are not robust against outliers and they also get affected by deviations from underlying model assumptions. Moreover, if the number of observations is small compared to the dimension of the measurement vector, these classical methods may lead to poor classification. On the contrary, ANN is more flexible and does not make any assumption about the population structure. But, it separates the competing populations using a complex surface. So, we sacrifice the lower dimensional view and the interpretability of the result, which are often the major concern in financial analysis. In this article, we propose to use a semiparametric method which preserves the interpretability and the lower dimensional view of class separability, but at the same time it is robust against outliers and capable to work well in high dimension and low sample size set up. We use two real life financial data sets to show the utility of this semiparametric method.