Predicting Corporate Failure: The GRASP-LOGIT Model

Abstract

Predicting corporate failure is an important problem in management science. This study tests a new method for predicting corporate failure on a sample of Spanish firms. A GRASP (Greedy Randomized Adaptive Search Procedure) strategy is proposed to use a feature selection algorithm to select a subset of available financial ratios, as a preliminary step in estimating a model of logistic regression for predicting corporate failure. Selecting only a subset of variables (financial ratios) reduces the costs of data acquisition, increases prediction accuracy by excluding irrelevant variables, and provides insight into the nature of the prediction problem allowing a better understanding of the final classification model. The proposed algorithm, that it is named GRASP-LOGIT algorithm, performs better than a simple logistic regression in that it reaches the same level of forecasting ability with fewer accounting ratios, leading to a better interpretation of the model and therefore to a better understanding of the failure process.