Models that have been studied in the financial literature and that are used to forecast bankruptcy are primarily default models: a firm goes bankrupt when it lacks sufficient resources to meet its financial obligations, hence when it becomes insolvent. Most empirical studies that have focused on bankruptcy prediction have therefore attempted to find measures that characterize a risk of default. The first models developed in the 1960s, following the study by Altman [4], have sought to assess this risk by estimating the distance between the financial situation of a given firm and a standard bankruptcy situation. Virtually all data-mining techniques that have been developed for classification purposes have been used to design failure models that share almost all the same characteristics: models are dichotomous, have good forecasting abilities and are easy to estimate. However, what can be considered the main factor of their success is also their main weakness. They essentially rely on a single rule and are estimated using financial data that solely characterize a unique period of firm life. This type of modeling reflects a rather rudimentary view of bankruptcy; it is considered the result of a a-historical process [39] that does not depend on time and that is reducible to a limited number of measures. But reality E-mail address: philippe.dujardin@edhec.edu. is a bit different. One knows that firms that apparently share the same financial profile, from the point of view of a model, may in reality have a very different probability of failure. Over time, some of them may have gained a certain resilience that gives them the ability to withstand failure. Some others may have received from their environment a sort of carrying capacity that has changed their fate at the very moment where their situation worsened, or have managed to recover even though nothing suggested they were able to do so [11]. All these factors, which can solely be analyzed over time, cannot be properly embodied by traditional models. The historical dimension of failure and the multiplicity of the situations that lead to bankruptcy have given rise to a large body of literature. We can find, on the one side, studies that focused on the temporal dimension of financial failure. They analyzed the way variables that measure firm activity over several years [22] may influence model accuracy, assuming that taking time into account with multi-period data would be sufficient to embody the dynamics of the phenomenon. We can also find, on the other side, studies that were interested in modeling the different financial situations that lead to bankruptcy. They especially analyzed how to embody at-risk situations using ensemble-based models, this time assuming that the multiplication of forecasting rules would make it possible to model the diversity of failure symptoms [32,46,61].