| Abstract: | 本論文以台灣股價指數期貨及美國SPDRs自2001年至2008年之日資料為實證標的，全球金融海嘯(2008)為預測期間，進行波動性預測能力比較及風險值績效評估。若預測模型可以在金融危機期間具有良好表現，實務上應該具有相當的重要性。有別於傳統文獻大多使用報酬率的平方作為市場真實波動的代理變數，本論文改以PK變幅、GK變幅、RS變幅及已實現波動度(RV)，並同時採用對稱與不對稱損失函數評估模型的波動性預測績效。更進一步加入已實現波動為基礎的風險值模型(RV-VaR)，利用Kupiec(1995)提出之非條件涵蓋率檢定，比較RV-VaR與GARCH族為基礎的風險值模型之風險管理績效。
實證結果皆指出，以不對稱GARCH模型的波動性預測能力較佳，顯示不對稱的變異數方程式設定能提升波動預測績效，其中以EGARCH模型最佳，而GARCH模型表現最差。在風險值評估部份，台灣股價指數期貨的實證結果，發現RV-VaR模型有低估風險值之虞，以致未能通過回溯測試；反之，GARCH族模型卻能提供準確的風險值預測績效。而美國SPDR指數型股票基金的結果則顯示各模型皆通過回溯測試，其中以RV-VaR模型較能準確估算真實風險值。整體來說，EGARCH與RV-VaR 模型分別為TAIFEX與SPDRs的最佳模型，此結果可提供機構法人、執政當局、風險管理者、投資大眾在面對未來極端事件時的參考依據，並提升風險控管績效。
This thesis applies alternative GARCH-type models to daily volatility forecasting with Value-at-Risk (VaR) application to the Taiwanese stock index futures and Standard & Poor’s Depositary Receipts (SPDRs) that suffered the global financial tsunami that occurred during 2008. Instead of using squared returns as a proxy for true volatility, this thesis adopts four volatility proxy measures, the PK-range, GK-range, RS-range, and RV, for use in the empirical exercise. The volatility forecast evaluation is conducted with a variety of volatility proxies according to both symmetric and asymmetric types of loss functions regarding forecasting accuracy. These models are also evaluated in terms of their ability to provide adequate VaR estimates with the inclusion of realized-volatility-based VaR model. Moreover, the predictive performance of the RV-based VaR model is compared with various GARCH-based VaR models according to both unconditional coverage test (Kupiec,1995) and utility-based loss functions with respect to risk management practice.
Empirical results indicate that the EGARCH model provides the most accurate daily volatility forecasts, whereas the performances of the standard GARCH model are relatively poor. Such evidence suggests that asymmetry in volatility dynamics should be taken into account for forecasting financial markets volatility. Moreover, I find a consistent result that the forecasting performance of models remains constant across various volatility proxies for both empirical data in most cases. In the area of risk management,the RV-VaR model tends to underestimate VaR and has been rejected for lacking correct unconditional coverage for the TAIFEX returns data, while the GARCH genre of models is capable of providing satisfactory and reliable daily VaR forecasts. In particular, the asymmetric EGARCH model is the most preferred. For SPDRs case, while all models have passed the back-test, the RV-VaR is considered the optimal VaR model both for a regulator and for a firm at alternative confidence levels during the whole year of 2008. The empirical findings presented here provide crucial implications for market practitioners, such as, policy makers, institutional risk managers, and common investors in risk management. |
