Dynamic self-organizing feature map-based models applied to bankruptcy prediction

Most bankruptcy prediction models used by financial institutions rely on single-period data, that is to say data that characterize firms at a given moment of their life. However, the financial literature devoted to bankruptcy considers this phenomenon to be a protracted process and time to be a fundamental explanatory variable of firm failure. The few studies that attempted to incorporate a temporal dimension into a forecasting model using multiperiod data often yielded results that did not really improve model accuracy. One may then suppose that the principles that ground the historical dimension of bankruptcy are less in question to explain the poor difference between the results estimated with these two types of data than the way time is modeled into prediction rules. This is why we propose a method that relies on a particular modeling of firm history using self-organizing neural networks and a segmentation of the data space, and which makes it possible to typify subsets of firms that share a common evolution of their financial situation over time. This method leads to models that are substantially more accurate than traditional ones, especially when it comes to forecasting the fate of firms the cost of misclassification of which is the highest for any financial institution.