Bankruptcy forecasting: An empirical comparison of AdaBoost and neural networks

Abstract

The goal of this study is to show an alternative method to corporate failure prediction. In the last decades Artificial Neural Networks have been widely used for this task. These models have the advantage of being able to detect non-linear relationships and show a good performance in presence of noisy information, as it usually happens, in corporate failure prediction problems. AdaBoost is a novel ensemble learning algorithm that constructs its base classifiers in sequence using different versions of the training data set. In this paper, we compare the prediction accuracy of both techniques on a set of European firms, considering the usual predicting variables such as financial ratios, as well as qualitative variables, such as firm size, activity and legal structure. We show that our approach decreases the generalization error by about thirty percent with respect to the error produced with a neural network.

Introduction

Predicting corporate failure is a hot topic in management science due to its importance for making correct business decisions. The accuracy of the forecasting model is clearly of crucial importance in failure prediction because many economic agents not only enterprises but financial institutions, auditors, consultants, policy makers or clients are affected by the bankrupt of a firm. In classification terms, the type I error is especially important, i.e. when a firm which will fail in the future is classified as healthy. Owing to this fact many researchers have focused their effort on finding the most efficient classifier. In the last decades artificial neural networks have received special attention and several studies have dealt with failure forecasting using this technique. Here we present some of them only as examples. Wilson and Sharda [55] used a sample of 129 firms, 65 of which went bankrupt between 1975 and 1982 and 64 non bankrupt firms matched on industry and year. They applied resampling techniques to generate the training and test data sets and reached very satisfactory results with only five accounting ratios.

Serrano–Cinca [48] provided a data set made up of 66 Spanish banks, 29 of them in bankruptcy and the rest solvent. He used nine ratios chosen from amongst those most commonly employed in accounting empirical research. The author proved the superiority of the neural network model against linear discriminant analysis using the leaving one out estimation of the error. Charalambous et al [16] applied several neural networks methods to a dataset of 139 matched-pairs of bankrupt and non-bankrupt U.S. firms for the period 1983–1994. The authors compared the predictive performance of five methods, namely Learning Vector Quantization, Radial Basis Function, Feedforward networks that use the conjugate gradient optimization algorithm, the back-propagation algorithm and the logistic regression.

In this research, the neural network approach is compared to AdaBoosted [19] classification trees for predicting corporate failure. As far as we are aware, this is the first study to compare AdaBoost and Neural Networks capabilities for corporate failure prediction. To illustrate its usefulness, we will apply AdaBoost on a selection of Spanish companies, and in order to ensure that these results are general and can be projected to other European countries and to the United States, we will use financial ratios that have proved significant for predicting business failure in previous studies (e.g. Frydman [23]).

The lack of a unified theory on corporate failure has meant that most studies dealing with distress prediction have focused on increasing the accuracy of the model and have not always paid enough attention to the model interpretation. This is clearly important in failure prediction as the firm must make appropriate decisions. Ensemble methods like AdaBoost do not improve model interpretation by themselves. Even more, they break the model interpretation conveyed by a decision tree. But, on the other hand, attribute importance methods can be devised to provide useful information for problem understanding. We will also calculate a novel measure for the importance of variables to facilitate model interpretation. This measure takes into account how often variables are actually used in the individual trees and, on the basis of this measure, the variables can be ranked in terms of importance.

The following factors should be taken into account within the empirical application. We use the legal definition of corporate failure which only includes bankrupt and temporary receivership firms. This is the most common definition in corporate failure prediction literature. One numerical (the firm size) and two categorical variables (activity sector and legal structure) are included as descriptors in addition to the usual financial ratios. The AdaBoost method is applied to the failure prediction, analyzing the extent to which this methodology is suitable for the subject.

In Section 2 of this paper, we present the AdaBoost method included in the study with a discussion of how it works in practice and we describe the algorithm used. The following sections introduce the failure prediction problem and the data used in the analysis. The classification results are then presented and the well-known neural network model is compared with the novel AdaBoost classifier. Finally, following on from the empirical analysis, we present our conclusions.