Designing Ensemble-Based Models Using Neural Networks and Temporal Financial Profiles to Forecast Firms’ Financial Failure

Most bankruptcy prediction models that have been analyzed in the literature rely solely on variables that measure firms’ financial health over a single year. However, it has long been known that a firm’s history plays a decisive role in its ability to survive and that, at the same time, variables used to embody this history in a prediction model often lead to marginal improvements in model accuracy. This apparent contradiction suggests that it is perhaps not so much the principle of using history as an explanatory variable that is in question, but rather the way in which this history has been captured and modeled up until now. This is why we propose a methodological framework that makes it possible to efficiently embody a firm’s history using a quantification process, and use the result of this process to improve model accuracy. It relies on the estimation of typical temporal financial“profiles” that govern the evolution of firms’ financial situations over time using an ensemble of neural networks. These “profiles” are designed in such a way as to be able to capture “general” patterns of evolution that tend to affect firms in a rather similar way, as well as “specific” patterns that may have different effects on various sub-groups of firms. They are used to build an ensemble of classification rules and make forecasts. The results achieved in this study show that this technique leads to forecasts that are more accurate than those of traditional methods at different time horizons.