Predicting default probabilities is at the core of credit risk management and is becoming more and more important for banks in order to measure their client's degree of risk, and for firms to operate successfully. The SVM with evolutionary feature selection is applied to the CreditReform database. We use classical methods such as discriminan analysis (DA), logit and probit models as benchmark On overall, GA-SVM is outperforms compared to the benchmark models in both training and testing dataset.