关于一类广义可加违约概率模型的探讨  

Study on evaluation of default probability based on generalized additive models

作　　者：王小明[1] 

机构地区：[1]上海财经大学统计学系,上海200433

出　　处：《系统工程理论与实践》2008年第6期52-58,共7页Systems Engineering-Theory & Practice

基　　金：教育部社会科学基金(05JA910004);上海财经大学十一五"211"工程项目

摘　　要：在现代商业银行的信用风险评估中,违约概率度量具有核心地位.传统的违约概率模型解释性强、计算强度小而应用方便,但容易产生模型设定偏差;现代人工智能模型预测精度高,却又存在解释性差、计算强度高和过度拟合等问题.为此,提出一类基于广义可加模型的违约概率模型,并对这类模型的拟合精度和预测效果进行实证比较.研究表明,广义可加违约概率模型不仅具有很高的预测精度,而且具有良好的可解释性和计算效率.另外,还从实际应用的角度出发,对该类模型拟合过程中可能存在的限制和困难进行探讨和分析,以利于对该方法的进一步的理论研究和实际应用.Evaluation of default probability （DP） is one of the keys to finance risk management in modem commercial banks. Traditional DP models are easy to be used in practice for their good interpretation and computational cheapness, but easy to produce model bias. On the other hand, modem artificial intelligence models often have a higher prediction accuracy, but drawbacks of lack of explanation capacity, high computational cost and easy to be over fitted. In this article, we introduce a new effective method for evaluation of DP based on generalized additive models. Real data analysis and comparison with some other exist methods indicate that these models perform well on discriminant accuracy as well as interpretation and computational effect. From the view of practice, some limits of the models are discussed on behalf of the further theoretical and practical studies.

关 键 词：信用风险度量 违约概率模型 广义可加模型 

分 类 号：F830.33[经济管理—金融学]