Forecasting bankruptcy using biclustering and neural network-based ensembles

Abstract

Most bankruptcy prediction models that have been analyzed in the literature, and that are estismated using ensemble-based techniques, are still not able to fully embody the true diversity of firm bankruptcy situations. Indeed, these models try to assess all bankruptcy situations either mostly using the same set of variables (bagging, boosting), or using the same set of observations (random subspace). In the first case, an ensemble assumes that any symptom of failure has the same origin. In the second case, it assumes that any financial situation that can lead to failure is the same for all firms. However, there are many situations where these two assumptions do not hold and where a state of bankruptcy may be specific to a given subgroup of firms or may be explained by a particular subset of variables. Certain methods, such as random forest or rotation forest, which combine the characteristics of both random subspace and bagging appear as solutions to this issue. However, they do not always perform significantly better than other ensemble models do. This is why we propose a modeling method that attempts to overcome the limitations of the previous models. It is based on a biclustering technique that seeks out groups of firms that are each characterized by a well-defined subset of variables and on an ensemble technique that is used to embody the full diversity of all bankruptcy situations that belong to each bicluster as precisely as possible. We show how the complementarity between these two techniques can improve forecasts.