摘要
为了更加有效地进行企业财务困境预测,利用t检验、单因素方差分析、逐步判别分析、逐步逻辑回归和邻域粗糙集5种特征提取方法,结合支持向量机、多元判别分析、Logistic回归、分类和回归树等多种分类学习算法构造备选基本分类器。在此基础上,提出了基于精度前向搜索和后剪枝的多特征子集组合分类器财务困境预测方法。该方法无需计算单分类器之间的差异性,首先以系统预测精度最大化为原则进行前向搜索,然后通过后剪枝策略选择精度最高或满意的系统作为最终结果。实证研究中以中国上市公司为研究对象,以十折交叉验证精度为评价标准,结果表明,该方法构建的组合系统的分类预测精度明显高于个体最优模型,最优组合系统和最简洁组合系统为财务困境预测提供了更多的灵活性。
In order to predict financial distress effectively,This paper constructs base classifiers by integrating multiple learning algorithms such as support vector machine,multi-discriminant analysis,Logistic regression,and CART with multiple feature selection methods including t test,one way ANOVA,stepwise discriminant analysis,stepwise Logistic regression and neighborhood rough sets.And then it proposes a multiple feature subsets ensembles,which select its base classifiers using an accuracy-guided forward search and post-pruning strategy.This method does not need to calculate the diversity among the base classifiers. Firstly it implements the forward search based on the principle of maximizing the system prediction accuracy,and then chooses the combing system with the highest accuracy or a satisfied one as the final result.Using Chinese listed companies＇ real data as our sample data and 10 fold Cross-Validation as an assessment,an empirical study is carried out.By comparing the experiment result with the individual best base classifier and within its inside components,it indicates that this method can improve the prediction accuracy significantly and provide more flexibility to financial distress prediction.
出处
《系统管理学报》
CSSCI
北大核心
2010年第4期420-427,共8页
Journal of Systems & Management
基金
国家自然科学基金资助项目(70773028
70801054)
高等学校博士点专项科研基金资助项目(200802130048)
关键词
财务困境预测
多特征子集
选择性集成
多分类器系统
financial distress prediction
multiple feature subset
selective ensemble
multiple classifier system
