Get access

Feature Selection for Support Vector Machine in the Study of Financial Early Warning System

Authors

  • Jingxiang Li,

    1. School of Statistics, Renmin University of China, Beijing, China
    Search for more papers by this author
  • Yichen Qin,

    1. Department of Operations, Business Analytics and Information Systems, University of Cincinnati, Cincinnati, OH, USA
    Search for more papers by this author
  • Danhui Yi,

    1. School of Statistics, Renmin University of China, Beijing, China
    2. Center for Applied Statistics, Renmin University of China, Beijing, China
    3. Statistical Consulting Center, Renmin University of China, Beijing, China
    Search for more papers by this author
  • Yang Li,

    Corresponding author
    1. School of Statistics, Renmin University of China, Beijing, China
    2. Center for Applied Statistics, Renmin University of China, Beijing, China
    3. Statistical Consulting Center, Renmin University of China, Beijing, China
    • Correspondence to: Yang Li, School of Statistics, Renmin University of China, Beijing, China.

      E-mail: li@ruc.edu.cn

    Search for more papers by this author
  • Ye Shen

    1. Department of Epidemiology and Biostatistics, University of Georgia, Athens, GA, USA
    Search for more papers by this author

Abstract

In this article, we introduce the L1 regularized support vector machine (L1-SVM) as an effective feature selection technique into the modeling of financial early warning system (EWS), for the purpose of establishing compact financial EWSs for Chinese-listed companies. By introducing LASSO penalty into the SVM framework, the L1-SVM is a capable methodology to select causative features in classification problem. We evaluate the feature selection performance of L1-SVM under different circumstances through numerical simulations and find it suitable for selecting features for financial EWS. In the real study, we establish four financial EWSs with features selected by L1-SVM and compare them with those trained with full features. The empirical result illustrates that our EWSs, with only minority of features, outperform significantly than full ones in the respect of generalization performance, which indicates the feasibility of L1-SVM in real applications. Copyright © 2014 John Wiley & Sons, Ltd.

Ancillary