Synthetic minority over-sampling technique (SMOTE) has the problem of over fitting in improving the imbalanced samples’learning ability of support vector machine (SVM) ． In this paper，adaptive synthetic sampling approach (ADASYN) and optimization of decreasing reduction approach (ODR) are assembled into an ODR-ADASYN to overcome the blindness in generating new samples and the limitations in processing the object． Combining SVM with ODＲ-ADASYN，an improved SVM，named ODＲ-ADASYN-SVM，is put forward to predict extremely financial risks; T-test is also applied to the significance test of the difference of the prediction accuracy of all models and to the evaluation of the prediction stability of all models． The result illustrates that the ODR-ADASYN-SVM can not only significantly improve the imbalanced samples’learning ability of SVM，but also overcome the problem of over fitting for SMOTE effectively． Hence，the ODＲ-ADASYN-SVM has a superior ability to predict extremely financial risks．