比較風險值方法之新模式

Abstract

本研究以Pérignon and Smith（2008）所提的多變量涵蓋機率檢定，作為回溯測試（Backtesting）的主要架構，以確認最佳的風險值估計方法。多變量涵蓋機率檢定是利用相同期間，不同的涵蓋機率或不同的信賴水準所得出之檢定量，以改進Kupiec（1995）單變量涵蓋機率檢定的不足。 本文採用臺灣加權股價指數、臺灣十年期公債指標殖利率之百元報價、美元外匯與臺指選擇權作為實證資料，運用變異數－共變異數法、歷史模擬法與蒙地卡羅模擬法來衡量資產組合的風險值，並以涵蓋機率檢定評估各風險值模型的準確性。變異數－共變異數法包括RiskMetrics法、GARCH法、GARCH－T法以及AR（1）-GARCH（1, 1）法。 本篇的實證研究所得到的主要結論是，多變量涵蓋機率檢定可以偵測到因為單變量涵蓋機率檢定無法全面考慮而導致的誤判，從而降低錯誤的風險值估計模型被接受的可能性。除此之外，實證結果也顯示變異數－共變異數法與歷史模擬法表現最佳，此結果與Pérignon and Smith的研究結論相符合。

This study adopts a new backtesting method of Pérignon and Smith (2008), which uses multivariate coverage test as the framework to evaluate the accuracy of VaR models. The multivariate coverage test is a multivariate generalization of Kupiec’s (1995) unconditional coverage test. Its basic idea is that instead of considering only a single coverage probability, the accuracy of a given VaR method should be assessed with different coverage probabilities within the same period. Weekly data of Taiwan Weighted Stock Index, 10-Year Government Bond, Currency Exchage Rates in US Dollars, and Taiwan Weighted Stock Index Options are used in this study. The results indicate that multivariate test improves the ability of univariate test to reject misspecified VaR models. Empirically, historical simulation method and parametric methods worked better for portfolio trading revenues. The results are consistent with that of Pérignon and Smith.