基于改进的MRMR算法和代价敏感分类的财务预警研究

The Research on Financial Early Warning Based on the Improved MRMR Algorithms and Cost Sensitive Classification

针对上市公司财务预警数据呈现出的高维和不平衡的双重特性,基于改进的MRMR算法和代价敏感分类构建财务预警模型并进行实证分析。首先,为了克服财务预警数据的不平衡性对特征选择和分类的不利影响,使用组合采样技术SMOTE+ENN进行数据平衡化处理。其次,利用绝对值余弦度量构建改进的MRMR算法并进行特征选择。最后,将支持向量机、L2-逻辑回归和CART决策树及其对应代价敏感模型作为比较模型进行财务预警研究。通过大量实证分析显示,SMOTE+ENN的引入有效提升了ST公司样本及其对应特征的重要性。在不影响财务预警模型总体分类性能的前提下,改进的MRMR算法可以得到更为简洁的预测特征集,且组合模型MRMR_FDAQ+CSSVM的预测结果最优,因此建议优先将该模型应用于上市公司财务危机的预测。

With the rapid development of the global economic integration and the market economy,Chinese enterprises,especially listed companies,are faced with huge risks and uncertainties in operation.Aiming at the dual characteristics of high dimensional and unbalanced financial early warning data of listed companies,the financial early warning model based on the improved MRMR algorithm and cost sensitive classification is considered and the corresponding empirical analysis is conducted.Firstly,in order to overcome the adverse effect of the imbalance of financial warning data in the feature selection and classification,the data sampling method is performed by the combined sampling technique SMOTE+ENN.Secondly,the improved MRMR algorithms based on the absolute cosine are used for feature selection.Finally,support vector machine,2-norm logistic regression and CART decision tree and its corresponding cost-sensitive models are used as comparative models for financial early warning research.A large number of empirical studies show that the SMOTE+ENN effectively improves the importance of ST company samples and their corresponding features;using the improved MRMR algorithms,a more concise prediction feature set can be obtained without affecting the classification performance of the early warning model.The combination model MRMR_FDAQ+CSSVM has the best prediction result,and it is recommended to be applied to the prediction of financial crisis of listed companies.