| 摘要: | 由於金融資產具有波動叢聚與異質變異的特性，欲估計出真實波動度而言是相當困難的，故本研究擬使用善於捕捉金融資產波動叢聚與異質變異能力之GARCH模型架構下，考慮二類波動模型：（1） GARCH（1,1）模型、（2） 在GARCH的條件變異數方程式中分別加入日變幅（PK、GK與RS）、已實現波動（Realized volatility, RV）、已實現雙冪次變異（Realized bipower variation, RBP）、隱含波動（Implied volatility, IV）與隔夜波動（Overnight volatility, ONV）之增廣GARCH模型，進行台灣加權股價指數（TWSE）與台灣櫃檯指數（OTC）日報酬率之波動預測，探討隱含於各波動估計式的日/日內價格交易資訊能否提升GARCH模型的波動預測能力。本研究在實證上擬以絕對報酬率（Absolute returns, ARET）、 PK日變幅與5分鐘之RV作為真實波動代理變數。在MAE及LL二種損失函數中，使用三種波動代理變數來建構樣本外波動預測績效的評估。特別是，本研究進一步使用benefit統計方法來檢測七種波動估計式對提升GARCH模型之波動預測的訊息價值。實證結果顯示，除了以ARET為波動代理變數之預測績效結果外，PK與RV之預測績效結果幾乎一致，皆以GARCH-RBP模型之預測能力最佳，即對GARCH模型能提升較多的波動預測之準確性。
Estimating the true volatility of assets returns is a difficult task since financial assets are well known to have stylized characteristics of volatility clustering and heteroskedasticity. Based on the GARCH （generalized autoregressive conditional heteroskedasticity, GARCH） framework, this thesis considers two GARCH volatility model specifications: （i） the traditional GARCH（1,1） model, （ii） the GARCH-X model which augments the traditional GARCH model by respectively incorporating daily price ranges （PK, GK, and RS）, realized volatility （RV）, realized bipower variation （RBP）, implied volatility and overnight volatility （ONV） as explanatory variable into the GARCH variance equation. These models are used to investigate the information value of the daily/intraday trading prices that is embodied in the aforementioned volatility estimators for improving forecasts of TWSE and OTC stock markets volatilities at daily horizon. This study adopts ARET （absolute returns）, PK range and RV volatility proxy measures for used in empirical exercise. The out-of-sample forecast evaluation is conducted using various proxy measures in terms of MAE and LL loss error statistics. Particularly, this study also employs benefit statistics to further examine the information values of the various estimators for improving GARCH-based volatility forecasts. The empirical results show that to predict fluctuations in performance results of the ARET proxy variables except, both the prediction of PK and RV performance results are almost the same, the predictive power of the model begin with GARCH-RBP are the best, i.e. GARCH model can enhance more accuracy of the volatility forecasting. |
