自身相關結構下台灣企業之企業評等預測

Abstract

本研究主要接續學者黃瑞卿等人(2009)的研究作後續探討。本文提議使用廣義估計等式(Generalized Estimating equations; 由學者Lipsitz 等人提出,1994)，來估計具有自我相關結構動態秩序機率模型的參數，預測公司信用評等。另一方面我們使用最大概似估計式來估計獨立結構的動態秩序機率模型參數。 本研究應用TCRI台灣信用評等資料，在獨立與自身相關結構下估計動態秩序機率模型的參數，並利用所估計之參數探討兩結構下動態秩序機率模型於樣本內分類正確率與樣本外預測能力的比較。 實證結果顯示，在獨立結構假設下的動態秩序機率模型有較高的樣本內分類正確率，且其分類正確率在整個樣本內其間較為穩定，這個結果可能的原因為一些技術上的限制。然而，自身相關結構假設下之動態秩序機率模型較假設為獨立結構之動態秩序機率模型， 有低的樣本內(in-sample)高估分類率，由於高估分類率會導致投資者有較重的損失，因此應較值得被投資者金管單位與企業注意。而在樣本外預測能力的比較，獨立結構下與自身相關結構下之動態秩序機率模型基本上都能得到一定程度的預測能力，但自身相關結構下之動態秩序機率模型會得到較低的樣本外高估預測率。 本文為國內首篇利用台灣信用評等資料來探討自身相關結構與獨立結構下的動態秩序機率模型的比較，最後結果指出因部分技術限制，自身相關結構於樣本內表現不如獨立結構，然自身相關結構下動態秩序機率模型可以與獨立結構動態秩序機率模型達到相同的樣本外預測正確率，同時也能得到較低的高估預測率。

This study mainly follows Hwang et al, (2009), used the generalized estimating equations to estimate the parameters of Dynamic Order Probit Model (DOPM), which assumed the autocorrelation structure of the error term. On the other hand, we used maximum likelihood estimator to estimate the DOPM parameters with independent error term. We then applied DOPM under these two assumptions (independent and autocorrelation error term) to model Taiwan Corporate Credit Risk Index data (TCRI). Then, we compared the in-sample classification accuracy and out-of-sample forecasting ability between DOPM with two structures. The empirical results reveal the preponderance of the independent structure DOPM for in-sample classification. Moreover, the predicting accuracy is more stable under independent structure DOPM than autocorrelation structure over time. The result can be attributed to some technical restrictions. However, the DOPM under autocorrelation structure presents fewer over rated cases which may result in more serious loss for investor, businesses, and regulators. The results of out-of-sample experiment show no better forecast ability of DOPM under each structure. However, more companies’ credit rating is over rated under independent structure. This thesis is the first research utilizes the Taiwan’s credit rating data to compare the DOPM under independent structure and autocorrelation structure. The empirical result shows that the DOPM under autocorrelation structure can achieve identical forecast accuracy with it under independent structure. And the out-of-sample forecast has fewer over rated cases.