本篇文章主要研究在不同長短的估計期間下，風險值績效表現是否會因而有所影響。研究標的採用美國道瓊工業指數於次級房貸風暴發生時期進行風險值評估。在模型配適部分利用平均-變異數模型與AR(1)-GARCH(1,1)配適，由於傳統統計推論假設資料是服從常態分配，但經過大量的實證研究證明金融資料的特性其實非常態，故將波動模型之誤差項設定為一般化偏態t分配，並與常態分配分別比較估計結果。而估計方法則以有無採用拔靴法計算信賴區間做區分，融入移動視窗概念估計變異數動態參數過程，以利提升估計風險值的準確度。 實證結果顯示當估計期間為250天，在回溯測試檢定或者是模型正確性檢定下，兩者的風險值績效表現都略低於估計期間較長的情況，也就代表進行風險值評估時，須要針對估計期間加以考量，且以採取長天期的估計期間較能夠有效的正確估計風險值，但估計期間設定過長則可能會發生資料存在結構性轉變的情況，故本研究提供風險管理者於實務上應針對期間設定加以分析，以提升風險值的績效表現。 This thesis examines different estimation periods that would affect the performance of VaR. Because of the reports of the Basel committee, estimating VaR should be performed on a long and short term basis. As such, the estimation periods will be 250 days, 500 days and 750 days in this thesis. The empirical data applies to the Dow Jones Industrial Average historical return rate that is provided to compare the performance of each model. We will also use the GARCH model, capturing heteroskedasticity effects to increase suitability. Also taking into account the characteristics of financial data, we use generalized skewed t distribution to evaluate the VaR, comparing it with normal distribution. The estimation method applies a rolling bootstrapping method to obtain an appropriate confidence interval. The empirical results demonstrate that, in both the back and LR testing, the performance of 250 days is worse compared to longer term periods. Consequently, this shows that using longer estimation period is superior to shorter term estimation. Concerning the empirical results, it is important to consider the difference of estimation periods to assess VaR. Empirical results provide greater accuracy in VaR forecasting. However, empirical results would require structure changes in long estimation periods. Hence, this thesis offers the conception of choosing estimation periods in order to increase performance. This technique is aimed to be helpful to investors or risk managers to make appropriate strategies.
