台灣遠期美元外匯風險溢酬之估測-MIDAS模型之應用

Abstract

本篇論文根據Ghysels (2005)所提出的混合數據抽樣模型(Mixed Data Sampling; MIDAS)，對台灣遠期美元外匯市場之風險溢酬進行估測。在研究假設上以Domowitz和Hakkio (1983)所定義之投資人持有遠期外匯契約的風險溢酬為基準，並引入跨期資本資產定價模型作為市場風險溢酬與風險間的基本關係。本篇論文觀察對象為10天期、30天期、60天期、90天期與180天期台灣遠期美元外匯，樣本期間選自1992年1月至2007年12月。研究發現，相較於傳統GARCH-in-mean模型與滾動視窗模型，MIDAS模型不但能穩定捕捉到外匯市場中時變風險溢酬的存在，並證實風險溢酬與其風險間具有正向變動的關係，對於抽樣後的風險溢酬之解釋能力亦較佳。此外，利用MIDAS模型可直接對投資人持有不同到期日之外匯契約的風險趨避程度做比較，實證結果顯示，當投資人持有到期日較長的外匯契約，面對外匯價格可能有較大的波動時，其風險趨避的程度也隨之增加。最後，進一步探討MIDAS模型中不同的波動預測因子和權重函數對外匯風險溢酬的預測能力，結果發現以變幅為基礎的波動預測因子對風險溢酬的解釋能力最佳，而不同的權重函數對MIDAS模型之預測能力並未有明顯不同的影響。本篇論文的研究結果可為投資人提供在資產配置上的一個參考依據。

This paper investigates the risk premiums of Taiwan’s U.S. dollar forward rates using Ghysels’ (2005) mixed data sampling (MIDAS) model. We introduce the assumption of forward contract’s risk premium defined by Domowitz and Hakkio (1983) as a starting point, and test intertemporal capital asset pricing model (ICAPM), which identifies the relationship of forward contract’s risk premium and holding risk. The NTD/USD exchange rates data from January 1992 to December 2007 are used to test for the time-varying risk premiums of 10, 30, 60, 90, and 180 days forward contracts. The empirical results show that the MIDAS model can robustly find existence of the time-varying risk premiums and positive relationship between the premium and risk. Compared with GARCH-in-mean and Rolling Window tests, the MIDAS model has better explanatory power in predictive regression for sampled risk premium. Moreover, we find evidence that the level of risk aversion increases steadily with the maturity horizon of forward contracts. This can be interpreted an indication that investors become more conservative when holding a long maturity contract. This paper further investigates the performances of the MIDAS specifications for five volatility predictors and two weight polynomials. We find that squared range provides a better forecast to the sampled risk premiums among five predictors. Two weight polynomials of the MIDAS regression perform similar results across all predictors.