本文研究對象為台灣加權股價指數、美元兌新台幣匯率、台積電股價、新竹商銀股價等日資料，分別以Gaussian GARCH、GARCH-t、GARCH-NoVaS等3種模型來進行實證，並以MAD作為比較基準，探討當金融資產報酬率存在高峰厚尾現象時，對於日報酬平方而言，何種模型的預測能力較佳。 實證結果證明GARCH-NoVaS模型的預測能力較Gaussian GARCH以及GARCH-t為佳，亦即當金融資產報酬率存在高峰態與厚尾現象時，GARCH-NoVaS不僅可以解決Gaussian GARCH所無法捕捉到的厚尾現象，亦可修正GARCH-t的低峰態的缺點，對於資產報酬率波動性之GARCH殘差的設定，比過去常使用的常態分配與t分配更為適當。 This research introduce three different GARCH models, they are Gaussian GARCH, GARCH-t, and GARCH-NoVaS. To evaluate and compare the predictive ability of three different GARCH models with respect to MAD, we focus on four well-know datasets, they are Taiwan weighted stock index, U.S. exchange rate, and stock price of Taiwan Semiconductor Manufacturing Co. and Hsinchu International Bank. We also discuss which model’s performance is better when the price return is leptokurtic and fat-tailed. The result show that the predictive ability of GARCH-NoVaS is much better than the others. GARCH-NoVaS can correct not only fat-tailed property which Gaussian GARCH cannot describe, but also the defect of low kurtosis of GARCH-t. The assumption of GARCH residual in GARCH-NoVaS is more appropriate than Gaussian GARCH and GARCH-t.
