Predicting corporate failure: empirical evidence for the UK

Abstract

The main purpose of this study is to examine the incremental information content of operating cash flows in predicting financial distress and thus develop reliable failure prediction models for UK public industrial firms. Neural networks and logit methodology were employed to a dataset of fifty-one matched pairs of failed and non-failed UK public industrial firms over the period 1988–97. The final models are validated using an out-of-sample-period ex-ante test and the Lachenbruch jackknife procedure. The results indicate that a parsimonious model that includes three financial variables, a cash flow, a profitability and a financial leverage variable, yielded an overall correct classification accuracy of 83% one year prior to the failure. In summary, our models can be used to assist investors, creditors, managers, auditors and regulatory agencies in the UK to predict the probability of business failure.

Acknowledgements

We gratefully acknowledge the helpful comments and suggestions of P. Hadjicostas, G. Hadjinicolas, G. Liapis, C. Mar Molinero, A. Soteriou, L. Trigeorgis, N. Vafeas, workshop participants at the 1998 European Accounting Association conference, workshop participants at the 2002 Real Options Conference, as well as workshop participants at the University of Cyprus and University of Toronto. The project was partly financially supported by the University of Cyprus and by the Institute of Certified Public Accountants of Cyprus (PriceWaterhousecoopers, Arthur Andersen Chrysanthou and Christoforou; KPMG; Ernst & Young; Deloitte & Touche). Remaining errors are the responsibility of the authors.

Notes

For an in-depth review of the corporate failure-related literature see Zavgren (1983), Taffler (1984), Jones (1987), Keasey and Watson (1991), Allen and Chung (1998) and Laitinen and Kankaanpaa (1999).

The multiple discriminant approach (MDA) is based on the following main assumptions: (a) the independent variables are multivariate normal, and (b) the covariance matrices of the two groups (failed and non-failed) are equivalent.

Logistic regression has the following advantages over MDA models (Ohlson, 1980; Mensah, 1984): (a) no assumptions need to be made regarding prior probabilities of failure and the distribution of predictor variables, (b) the use of such models permits an assessment of the significance of the individual independent variables included in the model, and (c) the models calculate the weight which each coefficient contributes to the overall prediction of failure or non-failure and produce a probability score, which makes the results more accurate.

Taffler is one of the most prominent insolvency researchers in the UK. He developed a number of failure prediction models for the UK corporate sector, which have performed well in terms of classification accuracy and they have become widely accepted tools for practical financial analysis in the UK.

For a review of the failure-related literature undertaken prior to the mid-1980s, see Taffler (1984).

The cash flow from the operations variable was defined as operating earnings plus non-cash expenses/revenues (non-current accruals) plus changes in working capital except for changes in cash and cash equivalents (current accruals). This definition of cash flows differs from the traditional one, which approximates cash flows by adding only depreciation to earnings (Laitinen, 1994). The traditional measure was shown in prior studies to be a measure of profitability (Ali and Pope, 1995).

Jones (1987) states the advantages of matching: ‘Bankrupt firms are often disproportionately small and concentrated in certain failing industries. If non-bankrupt firms were drawn at random, there would probably be substantial differences between the two groups in terms of size and industry. The result is that the model attempting to discriminate between failing and healthy firms may actually be distinguishing between large and small firms, or between railroads and other industrials’.

Statistical analysis was performed with the SPSS (version 8.0) statistical package.

The neural network analysis was performed using the Matlab programming language.

Type I error is the misclassification of a failed firm as non-failed and the type II error is the misclassification of a non-failed firm as failed.

Nagelkerke's adjusted R ² for each model is also shown in Table 6. The measure is an improvement of Cox–Snell's R ² = 1 − exp(−model L ²/N), where L ² is the model χ², which is the direct counterpart to the global F-test in linear regression analysis. N is the number of observations.

As Jones (1987) comments, if a holdout sample is obtained from a later period, one can test for both overfitting and a violation of the stationarity assumption. The stationarity assumption implies that the relationship between the independent variables and the dependent variable will hold over time. However, only a small number of researchers tested their models on a sample obtained from a later period (e.g. Peel and Peel, 1988; Platt and Platt, 1990; Charalambous et al., 2000).

Panel C of Table 6 presents the estimation results of the logit model developed using the entire sample of companies (1988–97) and which, for convenience, is referred to as Logit Model II. To achieve consistency with the first logit model (referred to as Logit Model I), the same explanatory variables were used. The overall estimation results of Logit Model II for the first two years before the failure are much lower than the respective results of Logit Model I (86% and 73% vs 94% and 84%). Both models provide equal classification rates for the third year.

Cross-validation means verification using a time-coincident holdout sample.