近來與波動預測有關的文獻中，使用極值資料為基礎的預測績效傾向優於以傳統報酬率為基礎的預測模型，如GARCH 類模型。然而，有關以極值為基礎的波動預測之實證成果仍然有限。承續與過去研究相同的興趣，本研究旨在同時使用統計與財務之損失函數對於較廣泛的資料集合，例如新興市場股價指數、商品指數與個股股價指數，重新評估以極值為基礎的樣本外波動預測績效。與過往研究不同之處在於，本研究使用Lopez (2001)提出的機率評分法則做為統計損失函數。與傳統的統計損失函數不同之處在於機率評分法則是針對模型對於可觀測的事件進行預測績效評估，而傳統的統計損失函數是針對不可觀測的波動進行預測績效評估。Lopez (2001)指出此評估方式較能夠貼近於預測者的真實決策進行波動預測的績效評估。關於財務觀點的波動績效評估，本文根據 Ferreira and Lopez (2005)在風險值的架構下，使用準確性與效率性檢定探討以變幅波動為基礎的波動預測。
In recent literature related to volatility forecasts, the forecasting performances based on extreme value data tend to outperform those produced by traditional return-based model, such as GARCH-type model. However, the empirical achievements about extreme value-based volatility forecasts are still limited. Following the same interest of previous researches, this study aims to contribute to the literature by adopting both statistical and financial loss functions to reexamine out-of-sample performance of extreme value-based volatility forecasts for a variety data set, such as emerging market stock indices, commodities indices and individual stocks. Different from previous studies, probability scoring rules of Lopez (2001) is employed as the statistical loss function. The key difference between scoring rules and standard statistical loss functions is that scoring rules compare a model’s forecasts to observable events, rather than a proxy for the true unobservable variance. As indicated by Lopez (2001), this evaluation method can be tailored to the actual decision problem of a forecast user. With regards to evaluation of financial perspective, following the work of Ferreira and Lopez (2005), the extreme value-based volatility forecasting is evaluated under VaR framework, which involves both accuracy and efficiency tests.
