以往針對風險值的估計，大致以GARCH模型估計波動度後，再以蒙地卡羅模擬法估計 風險值續效。本研究延伸過去的方式，試圖以Bi-GARCH模型估計波動度後，分別以狀 態空間及基因類神經網路兩種模型，應用於波動度濾化（Filtering)的處理，之後再與傅 統GARCH模型，藉由回溯測試以及Kupiec概似比檢定對風險值估計模型的續效進行 比較。本研究以台灣發行量加權股價指數與股價指數期貨結算價格為研究標的，取樣期 間為2002年1月1日至2004年12月31日之日資料，共計745筆。研究發現，無論在 台股指數現貨或期貨，經由狀態空間與基因類神經網路濾化後之風險值的模式續效，不 論短長期或多空階段之評比，皆遠優於傅統估算法。

Previous researchers usually use GARCH models in estimating volatility in evaluating value at risk (VaR) performance. In this study, Bi-GARCH models were adopted in estimating volatility. The estimated volatility is then filtered by using both State Space Models (SSM) and Generic Algorithm-Artificial Neural Network (GANN) models. The VaR performances of these models are compared using back-testings and Kupiec likelihood tests. A total of 745 daily data of Taiwan stock indexes of spot and futures ranging from Jan. 2, 2002 to Dec. 31, 2004 were collected. The results show that the filtered GANN and SSM models are better then traditional estimation methods for the evaluation of VaR for both stock spot and futures indexes.