| 摘要: | 有別於以往文獻使用傳統單變量檢定迴歸式對遠期匯率不偏性假說進行檢定，本文採取4種panel unit root檢定對6個國家在1、2、3、6個月等不同合約期間下的遠期溢酬進行檢定。本文首先對LL、IPS、Fisher P_lambda與考慮結構性變動的panel LM檢定做一番簡要的導覽與回顧，瞭解各檢定提出的原始動機與其基本精神之後，便對各檢定的實證結果進行分析與探討。除了比較panel unit root檢定與單變量的ADF-GLS檢定之間的差異之外，還觀察並比較4種panel unit root檢定結果之間是否一致，並藉此瞭解1992年至2002年間新台幣相對於外幣在外匯市場內的變化情形。 由檢定結果可發現，不僅單變量的ADF-GLS檢定與4種panel unit root檢定之間存在差異，panel unit root檢定之間也因有無考慮資料具結構性變動特質而產生不同的檢定結果。在ADF-GLS檢定結果裡，不論是有無包含時間趨勢項，英鎊、歐元、加拿大幣與日圓大都傾向不拒絕單根的虛無假設。而在LL、IPS與Fisher P_lambda等檢定結果裡則除了歐元之外，其餘國家幾乎都可以拒絕虛無假設。造成這兩種檢定結果相差甚遠的原因，相信是因為使用了包含更多訊息的panel模型之後，影響了對個別數列的統計推論。而在進一步使用考慮資料具有結構性變動特質的panel LM檢定後，發現在5%的顯著水準下，能夠不受有無考慮時間趨勢因素的影響下皆能拒絕單根的有：加拿大幣、新加坡幣、日圓與台幣，而在LL、IPS與Fisher P_lambda等3種檢定中幾乎都能夠顯著拒絕虛無假設的英鎊，在考慮結構性變動後幾乎不能拒絕。 Unlike the earlier research, which used the conventional univariate regression to test the forward rate unbiased hypothesis, we adopt 4 panel unit root tests to examine forward premium series of 6 currencies over panels of 1-, 2-, 3-, and 6-month forward contract maturities. In the beginning of the thesis, we briefly present and review the LL, IPS, Fisher P_lambda and panel LM test, which consider the structure change of data, to realize the motivation of adducing these methods, and their original meanings, then we discuss and analyze our empirical findings. Not only compare the differences between panel unit root tests and the univariate ADF-GLS test, we also observe and compare the outcome of 4 panel unit root tests, therefore we can understand how NT dollar vary in the foreign exchange market from 1992 to 2002, by contrasting with other currencies. From the test results we find not only the univariate ADF-GLS test is different to the 4 panel unit root tests, but also, the 4 panel unit root tests have different outcome, either. In the results of ADF-GLS test, no matter with or without the time trend term, British pound (BP), Euro, Canadian dollar (CD) and Japanese yen (JY) all can reject the null of unit root. However, in the results of LL, IPS and Fisher P_lambda test, almost all currencies can reject the null except Euro. The reason of making the huge difference between the 2 kinds of tests might be by using the panel model, which can include more information, can effect the conclusion of statistical inference. Furthermore, by applying panel LM test with allowing structure breaks of data, we find no matter with or without the time trend term, CD, Singapore dollar, JY and NT dollar can reject the null hypothesis at 5% significance level. However, BP, which can reject the null in LL, IPS and Fisher P_lambda test, almost cannot reject when considering the structure change of data. |
