自從G30、BIS等權威機構推薦風險值作為量化市場風險的方法後，風險值已成為現今市場風險管理的重要工具。本研究以台股加權指數為實證對象，評比不同模型所估計出的風險值與預測績效。模型分為無母數模型與有母數模型兩類。本文所使用的無母數模型為分位數迴歸模型 (Koenker及Bassett 1978， 1982) ，並根據 Taylor (1999) 的方法選出適合的解釋變數。有母數模型包括J. P. Morgan的 Riskmetrics 模型及含有異質變異的GARCH模型。由於有母數模型假設資產報酬率為常態分配，但多數資產報酬率具有厚尾的現象，本文也嘗試以訓練資料標準化後取分位數替代標準常態分配的臨界值，探討是否能有效捕捉資料的厚尾現象。本研究得到的結論包括： (1) 分位數迴歸模型搭配由有母數模型估計出的一步預測標準差，通常會改進風險值估計的效果。 (2) 對於長天期風險值的估計，分位數迴歸模型是值得參考的。 (3) 報酬率左端有很明顯的厚尾現象，使用報酬率標準化後的經驗分位數 (Empirical Quantile) 會較使用標準常態 的臨界值要好。 (4) 趨勢向下與趨勢向上的預測期間會影響風險值的估計績效。在趨勢向上時，每種模型的向前測試結果大致都很好，但在趨勢向下時則不盡然。

Since the authoritative organizations, such as G30 and BIS, recommended the Value at Risk (VaR) as a way to quantify marketing risks, VaR has recently became an important tool on market risk management. In this research, we take Taiwan stock exchange index as empirical data and the estimated VaR and predicting effectiveness for different models are compared. There are two types of VaR models: non-parametric models and parametric. The non-parametric model in this paper is the quantile regression model (Koenker and Bassett, 1978, 1982) in which the independent variables are chosen by Taylor's (1999) method. Parametric models consist of J. P. Morgan's Riskmetrics and GARCH model. The quantile of the standardized training data instead of the critical value of the standard normal distribution to catch the fat-tailed phenomenon is used. The conclusions include: (1) One-step-ahead standard deviation forecast, estimated by parametric models and combined with quantile regression model, usually improve the VaR prediction. (2) Quantile Regression model is generally good for long-holding periods. (3) When the left end of returns presents fat tail, applying empirical quantile is obviously better than using the critical value of the standard normal distribution. (4) Upward or downward returns may significantly influence the effects of the estimation of VaR. When it goes up, most of the estimation results are accurate, but not in the case of a downward trend.