This paper proposes a hybrid approach to theforecasting of firms’ bankruptcy ofSpanish enterprises from the construction sector. Our proposal starts splitting the group of healthycompanies into two subgroups: borderline and non-borderline companies.Borderline companies are healthy companies with marked financial similaritieswith bankrupt ones. Then, each subgroup is divided in clusters according totheir financial similarities and then each cluster is replaced by a directorvector which represents the companies included in the cluster. In order to dothis, we use Self Organizing Maps (SOM). Once the companies in clusters havebeen replaced by director vectors, we estimate a classification model throughMultivariate Adaptive Regression Splines (MARS). Our results show that theproposed hybrid approach is much more accurate for the identification of thecompanies that go bankrupt than other approaches such as a multi-layerperceptron neural network and a simple MARS model.