A sparse multiple kernel learning method for listed companies financial distress prediction

Accurate prediction for listed companies before financial distress arising is always a critical issue in financial administration, especially improving the accuracy of financial distress prediction (FDP). The features from the financial data are just different indexes which represent financial status of the listed companies in different aspects. These heterogeneous features bring huge challenge to FDP. Multiple-kernel learning (MKL) has demonstrated better performance than the conventional support vector machine (SVM) for FDP. In this paper, a sparse multiple-kernel learning method is introduced for FDP. Firstly, an unsupervised learning is performed on predefined basis kernels. A cardinality constraint is then enforced on the linear combination of the basis kernels so as to improve learning performance and interpretability of the model learned. After that, an optimal kernel is unsupervisedly learned. Finally, the optimally combined kernel is used in SVM optimization and the final multiple-kernel predictor can be achieved for FDP. Experiments are conducted with 207 couples of normal and ST companies. The experimental results prove that the proposed sparse MKL algorithm outperforms the state-of-the-art and non-sparse MKL in FDP both at whole data set and different industry data set.