Forecasting Corporate Failure with Neural Network Approach: The Greek Case

Introduction A number of studies have been conducted in the past on the subject of predicting corporate bankruptcy. Amongst the oldest, are the ones of Beaver' and Altman2, who concentrated on predicting business failure of US companies, using classical statistical techniques. Beaver was the first to use a univariate discrimination test. He found that various financial ratios could be useful predictors of failure even five years before the event. Altman, using multiple discriminant analysis (MDA) reported a very high accuracy rate in predicting failure one year in advance. From then on, a large number of studies have tested bankruptcy prediction models, based either on MDA or multiple linear regression (e.g., Eisenbeis3, Altman4, Taffler5, Izan6, etc.) which provided evidence in favour of the usefulness of such models. However to date, very few studies were based on non-parametric approaches. Yang et al7, is one recent example of employing a non-parametric approach, and in particular neural networks, for bankruptcy prediction among firms, in the US oil and gas industry. In their study, they contrasted neural network approaches with linear discriminant analysis for 122 companies for the period 1984-1989. Their findings revealed that neural network approaches achieved higher overall classification accuracy. In this paper, we also contrasted the neural networks approach with multivariate discriminate analysis for 40 Greek companies in various industries for the period 1981-1985. With the exception of Grammatikos and Gloubos8, there are no other studies in the literature regarding bankruptcy prediction for Greek companies (using either classical statistical or non-parametric techniques). In 1983, amidst a very bad economic climate, the Greek Government introduced the 1386/83 Law, which set several criteria according to which a firm could be characterised as "problematic", and as such qualify for certain financial benefits. These were, basically, firms of the state or private sector in financial distress, heading for almost immediate bankruptcy. The Greek State attempted to salvage at least some of them, in order to avoid a further worsening of the economic climate. A group of 20 "problematic" companies, matched with an equal number of solvent companies, with similar characteristics, comprised the "target group" of our analysis. Non-Parametric Vs Linear Classification Neural networks are a consistent paradigm of the non-parametric approach in financial modelling. Their major strength lies in the fact that they do not require any a priori assumptions regarding the underlying structure of the relationship they are estimating. In contrast, the parametric approach does make restrictive assumptions regarding the nature of this relationship. For example, multivariate discriminant analysis is based on the assumption that the relationship between the explanatory variables (e.g., accounting ratios) and the dependent variable (e.g., a solvent or an insolvent firm) is linear. Whether, or not, this is an accurate description of reality depends on the actual problem. In the general case, however, apart from the ease of calculations, there is no other reason why we should make such a restrictive assumption and expect reality to conform to it. We demonstrate this fact with a simplistic twodimensional example. In Figure 1 we depict 40 points (as triangles and circles), corresponding to 40 pairs of a continuous dependent variable (y-axis) and a single explanatory variable (x-axis). Depending on the exact location of each point in the chart, it can be classified as belonging to class A (triangle) or to class B (circle). The underlying (and in practise unknown) classification rule corresponds to the broken curve that divides the triangles above from the dots below. As we can see, if we try to apply a linear classification approach (e.g., pass a linear regression through the dots and triangles), the end result will be far from satisfying since many members of class A will be wrongly classified as members of class B. …