Non-Classical Ratios and Lasso Selection for Bankruptcy Prediction

Abstract

This study pursues two issues in the context of reduced-type prediction of corporate bankruptcies. Firstly, we investigate the marginal usefulness of some non-classical ratios and indicators. We find that the performance of classical accounting and market ratios can be significantly improved by their squared terms and technical ratios built from itemized accounts. Also, missing values of many accounting items are proven to be informative.

Secondly, as the non-classical ratios dramatically increase the number of candidate variables, the variable selection becomes particularly important. We review various traditional selection methods and investigate the innovative lasso regularization technique. We show that lasso performs for these purposes better than the popular stepwise selection, and requires considerably less computational time. Furthermore, lasso proves to have a few specific advantages in the investigated context, e.g. being able to efficiently identify the non-classical ratios which complement (rather than replace) classical ratios.

Overall, the non-classical information used in this study increased the out-of-sample accuracy ratio by 1.9%. When applying the lasso instead of stepwise, this advantage increased to 2.3%.

Keywords: Default Prediction, Bankruptcy Prediction, Reduced Models, Non-Structural Models, Accounting Ratios, Market Ratios, Non-Classical Ratios, Missing Values, Logistic Regression, Variable Selection, Stepwise Selection, Lasso, L1 Regularization

JEL Classification: C25, C51, C52, G24, G32, G33