本文利用一個兩階段的檢驗程序,探討臺灣股價指數報酬率之GARCH模型設定方式及其干擾項分配假設。這個實證發現,在一組相互競爭的GARCH模型中,以具有ARCH-in-mean設定及條件一般化誤差分配 (conditional generalized error distribution) 之指數化GARCH (exponential GARCH, EGARCH) 模型,最能解釋臺灣股價指數報酬率的動態行為。樣本內的報酬率變異性比較顯示,當市場波動劇烈時,這個EGARCH模型會產生迥異於GARCH模型的波動性詮釋。而樣本外的比較則顯示,相較於較GARCH及RiskMetrics模型,本文所選取的模型能更準確地預測報酬率分配的尾端行為。此外,這個研究亦提供一些具有經濟計量意涵的實證結果。首先,如同Chen and Kuan (2002) 針對美國股價指數報酬率所得到的研究結果,本文亦發現GARCH研究中常用的Ljung-Box, McLeod-Li, 及BDS檢定並無法有效地檢測出忽略波動不對稱性之誤設的GARCH模型。再者,此項誤設有可能被錯誤認定成GARCH模型干擾項的非對稱性。
In this paper, we concentrate on the discrimination between competing GARCH type models and innovation distribution assumptions for the Taiwan stock index returns. By using a two-step detection procedure, we find that the EGARCH with the ARCH-in-mean effect and the generalized error distributed innovations is the most promising one of a set of representative models. The in-sample comparison shows that the selected model generates different interpretations on the volatility from the GARCH model (with the normally distributed standardized innovations) when the market is volatile. The out-of-ample comparison demonstrates that the selected model outperforms the GARCH and RiskMetrics models for predicting the tails of the return distribution. This study also provides some empirical evidence with important implications on GARCH modeling. First, as shown by Chen and Kuan (2002) for the U.S. stock index returns, we find that the Ljung-Box, McLeod-Li, and BDS tests are unable to discover a misspecified GARCH model with the neglected asymmetric volatility effect for the Taiwan stock index return. Second, if the misspecification of the GARCH model is overlooked, then the neglected asymmetric volatility effect may be confused with the asymmetry of innovation distribution.