Dynamic financial distress prediction using instance selection for the disposal of concept drift

Prior studies of financial distress prediction (FDP) all focus on static modeling and ignore whether the model is still suitable with time passing on. This paper devotes to the first investigation on what the concept of financial distress concept drift (FDCD) is, whether FDCD exists and how to dispose FDCD. We construct a dynamic FDP modeling based on instance selection for the disposal of FDCD. Dynamic FDP consists of instance selection, FDP modeling and future prediction. Instance selection methods including full memory window, no memory window, window of fixed size, window of adaptable size, and batch selection are used to tackle FDCD. For feature selection, we construct a wrapper by integrating forward and backward selections on Mahalanobis distance. Empirical results indicate that gradual and constant virtual concept drift does exist in FDP, and dynamic FDP models perform much better than static models. Meanwhile, window of fixed size and batch selection are more suitable for Chinese listed companies’ dynamic FDP.