Failure pattern-based ensembles applied to bankruptcy forecasting

Bankruptcy prediction models that rely on ensemble techniques have been studied in depth over the last 20 years. Within most studies that have been performed on this topic, it appears that any ensemble-based model often achieves better results than those estimated with a single model designed using the base classifier of the ensemble, but it is not uncommon that the results of the former model do not outperform those of a single model when estimated with any other classifier. Indeed, an ensemble of decision trees is almost always more accurate than a single tree but not necessarily more than a neural network or a support vector machine. We know that the accuracy of an ensemble used to forecast firm bankruptcy is closely related to its ability to capture the variety of bankruptcy situations. But the fact that it may not be more efficient than a single model suggests that current techniques used to handle such a variety are not completely satisfactory. This is why we have looked for a method that makes it possible to better embody this diversity than current ones do. The technique proposed in this article relies on the quantification, using Kohonen maps, of temporal patterns that characterized the financial health of a set of companies, and on the use of an ensemble of incremental size maps to make forecasts. The results show that such models lead to better predictions than those that can be achieved with traditional methods.