Are Failure Prediction Models Widely Usable? An Empirical Study Using a Belgian Dataset

Abstract

Faced with the question as to whether failure prediction models can easily be transferred and applied to a new data setting, this study examines the performance of seven models on a dataset of Belgian company failures after re-estimation of the coefficients. The validation results indicate that some models are widely usable: they are strongly predictive when applied to the new data set. The Gloubos-Grammatikos models and Keasey-McGuinness appear among the best performing models, and also Ooghe-Joos-De Vos and Zavgren seem to be widely usable, respectively for failure prediction 1 and 3 years prior to failure. At the same time, the Altman and Bilderbeek models show very poor results when applied to the Belgian dataset. The best performing models seem to combine the right variables in an intuitively right sense and it appears that the combination of some types of variables generally leads to good predictive results. On the contrary, the estimation technique, complexity and number of variables do not explain the predictive performances (JEL: G33,M49).