A hybrid Deep belief network approach for Financial distress prediction

Abstract

After the subprime crisis in 2008, an efficient Financial Distress Prediction (FDP) model has become necessary. Many research works have attempted to provide a model using statistical or intelligent methods. In this respect, this paper adopts a two-stage hybrid model that integrates Deep Learning and Support Vector Machine as a FDP modeling method. Local receptive fields is a technique used in order to select the nodes for each layer of our deep network. Then, stacked Restricted Boltzmann Machine is applied to form a Deep Belief Network as pre-training. Subsequently, Support Vector Machine follows for classification. An experiment over a sample of French firms offers accuracy details about this method. The proposed model actually provides a result of 76,8% instances that are correctly classified. Henceforth, the new technique could prevent Financial distress before it happens. As such, investors, managers, banks, decision makers and others could benefit from its significant impact.