The main problem faced by banks when granting credit is their inability to determine with any certainty whether or not the customer will honor its commitments. In fact, early detection of a company's difficulties is achieved by using default risk forecasting tools, all of which rely on past analysis to predict the company's future. This analysis is essentially based on the company's financial statements, which remain an essential source of information for detecting business difficulties.The aim of our work is to make a methodological contribution by proposing an effective tool for predicting the risk of corporate failure, based on the neural approach. The aim is to show how variable selection techniques adapted to neural networks offer a better choice in terms of failure prediction quality than those traditionally used in experimental studies, which remain insufficient to improve the degree of prediction of the Moroccan banker's failure models.