信用違約機率之預測－Binary Regression Quantiles的應用

Abstract

本研究預測違約機率的方法為:Binary Regression Quantiles(二元分量迴歸)，此理論基礎與預測方式是使用美國學者Grigorios Kordas(2004)的方法，將分量迴歸運用在應變數為二元的屬質變數上之計量方法。\n最小平方法是目前最常見到的迴歸分析，但在古典線性迴歸模型中，應變數的解釋是來自於自變數的相對應的平均變化，而忽略了不同規模與分配下應變數的邊際變化，本文試圖以此方法和以最大概似估計法所建構出的Logit模型做一比較，而研究資料為台灣於民國85年至93年曾被列為全額交割類股的上市公司。\n 本研究發現Kordas (2004)的方法，雖然能將分量迴歸應用在屬質二元變數上，但是在預測方面相較於傳統Logit方法卻沒有出現較佳的預測能力。

The method implemented in PD calculation in this study is “Binary Regression Quantiles”. The foundation of the research and the way to forecast is according to the Ph.D Thesis of Grigorios Kordas(2004). He apply the binary variable for Quantile Regression.\n The Ordinary Least Square is the most common way to regression analysis, but in the classic linear regression the change of dependent variable comes from the independent variable averagely. It neglects the marginal change of the dependent variable according to different scale and distribution. We want to compare the Binary Regression Quantiles with the Logit Regression. \n Although we could apply the binary variable for Quantile Regression successfully, the outcome of the forecast is not as efficient as the Logit Regression.