Yang Li, Jingxiang Li, Shuangge Ma

【Abstract】

In the stock market, there are far fewer “ST” corporations than normal ones, thus financial early warning system, as a classification model, will always be trained with a seriously imbalanced dataset. However, common models like logistic regression cannot deal with such kind ofdata. Thus our study develop a financial early warning system, which is immune to data imbalance, by using weighted L1 regularized support vector machine （w-L1SVM）: On the one hand, w-L1SVM avoid the negative effects caused by imbalanced training data viaweighting samples from two classes separately; On the other hand, by introducing LASSO penalty into w-L1SVM, the w-L1SVMis capable to select significant features while being trained. Through numerical study, we verify the w~L1SVM’s performance on prediction and feature selection under imbalanced dataset. In the real study part, we develop a financial early warning system for Chinese corporations in machinery manufacturing industry using w-L1SVM. Result shows that our model is capable to select significant indicators and to detect “ST” corporations accurately. More, it has better prediction performance than other traditional methodologies, which certify the feasibility of w-LISVM in financial early warning system.

【Keywords】

weighted L1 regularized support vector machine, imbalanced dataset, feature selection, financial early warning system