實質匯率預測力-以台灣為例

A Study of The Predictability of Real Exchange Rate: Evidence from Taiwan

萬怡如(Yi-Ruh Wan)

指導教授 ： 吳致寧

國立中正大學/管理學院/經濟系國際經濟研究所/博士(2012年)

摘要

本論文旨在探討總體經濟模型是否能夠用來預測匯率？根據過去文獻，大多認為匯率是不可預測(Meese and Rogoff, 1983,1988)，少數則認為僅能做中、長期之預測(Mark, 1995, Killian and Taylor, 2003)。本文採用線性、非線性與學習模型三種方法針對新台幣兌美元之實質匯率進行匯率預測，證明選擇適當之總體經濟模型能夠對匯率作短期預測。在線性與非線性方法中，本文採用實質利率平價(URIP)與泰勒法則(Taylor rule)作為預測匯率之總體經濟模型，並且假設預測方程式中之解釋變數(regressor)為線性自迴歸模式(AR)與非線性指數平滑轉換自迴歸模式(ESTAR)兩種對匯率進行樣本內與樣本外預測(In- and out-of-sample prediction)。樣本內預測以t統計量衡量自變數與依變數之關係，其結果顯示，在實質利率平價之總體經濟模型下，實質匯率與兩國之實質利率差在線性與非線性模式下，無論短、中、長期均有顯著之關係；在泰勒法則之總體經濟模型下，未考慮利率平滑(interest rate smoothing)之典型的泰勒法則(conventional Taylor rule)模型顯示實質匯率與兩國之通貨膨脹差與產出缺口差在短、中或長期有顯著之關係，但對考慮利率平滑之泰勒法則模型則僅在長期顯著。至於樣本外預測能力則是採用DM、GM和CW統計量來衡量以實質利率平價與泰勒法則作為匯率預測的總體經濟模型和匯率隨機漫步模型(random walk model)兩者之預測正確性是否相同。研究結果顯示:無論是以實質利率平價或以泰勒法則作為預測匯率之總體經濟模型均可以打敗匯率隨機漫步模型成功的預測下一季(1-quarter ahead)之實質匯率，且總體經濟模型之均方根誤差(root mean square error, RMSE)均小於匯率隨機漫步模型。
有鑑於央行之貨幣政策(例如泰勒法則)常隨著時間而改變且通常不會對大眾公開，因此，人們只能經由過去經驗的學習來預測央行貨幣政策中之參數。在本文第三部份，採用Evans and Honkapohja (2001) 最小平方學習法則(least squares learning rule)對泰勒法則中之參數進行估計，然後再根據前述之方法進行匯率預測。在此學習模型下之樣本內預測結果顯示，實質匯率與兩國之通貨膨脹差與產出缺口差之關係並不顯著；但樣本外預測能力卻顯示以DM、GM和CW表示之統計量大多可以打敗隨機漫步模型成功的預測下一季(1-quarter ahead)之實質匯率，且均方根誤差(root mean square error, RMSE)亦小於匯率隨機漫步模型。
總結而言，本研究顯示在線性、非線性與學習模型三種方法探討下，無論是實質利率平價或泰勒法則作為匯率預測的總體經濟模型其樣本外預測能力在短期均優於匯率隨機漫步模型。

關鍵字

樣本外預測 ； 泰勒法則 ； 實質利率平價 ； 實質匯率 ； 學習模型

並列摘要

This dissertation aims for finding the answer of “Are exchange rates predictable based on macroeconomic fundamentals?” which is a longstanding problem in international finance over the past two decades and has not yet been convincingly explained. Three approaches, linear, nonlinear and learning, are used to investigate the real exchange rate predictability between US and Taiwan. The results show that real exchange rate can be predictable at short horizon by choosing the appropriate macroeconomic fundamentals. In the linear and nonlinear approaches, two different types of fundamentals based on real interest parity (URIP) and Taylor rule are applied and the data generating processes of regressors in the prediction equation are assumed in linear-AR and nonlinear-ESTAR models separately. The results show the evidence of in-sample real exchange rate-real interest differential link at all horizons and when the conventional Taylor rule is adopted, in-sample real exchange rate-inflation differential and output gaps differential link at all horizons except at long-horizon in the nonlinear approach. For the case of the Taylor rule with interest rate smoothing, in-sample real exchange rate - inflation differential and output gaps differential link only reveals at long-horizon. The out-of-sample predictability of the exchange rate models based on the URIP and Taylor rule fundamentals outperforms the random walk model which is statistically significant by the test of DM, GM and CW at 1-quarter horizon and the root mean square error of the exchange rate models is smaller than the error of the random walk model.
In the learning approach, considering the parameters of central bank’s monetary policy are unknown to the public and change over time, this dissertation uses least squares learning rules provided by Evans and Honkapohja (2001) to estimate the coefficients of Taylor rule, then investigates the real exchange rate predictability between US and Taiwan as previously mentioned. The results show no evidence of in-sample real exchange rate-inflation differential and output gaps differential link. But, the out-of-sample predictability of the exchange rate models based on the learning Taylor rule fundamentals outperforms the random walk model which is also statistically significant by the test of DM, GM and CW at 1-quarter horizon and the root mean square error of the learning Taylor rule fundamentals is also smaller than the error of the random walk model.

並列關鍵字

Taylor-rule ； URIP ； real exchange rate ； out-of-sample forecast ； learning

參考文獻

Chen, S., Wu, T., (2010), “Assessing Monetary Policy in Taiwan,” Academia Economic Papers 38 (1), 33-59.

Andrews, D., (1991), “Heteroscedasticity and Autocorrelation Consistent Covariance Matrix Estimation,” Econometrica 59 (3), 817-858.

Becker, R., Enders, W., Hurn, S., (2004), “A General Test for Time Dependence in Parameters,” Journal of Applied Econometrics 19 (7), 899-906.

Becker, R., Enders, W., Lee, J., (2006), “A Stationary Test in the Presence of An Unknown Number of Smooth Breaks,” Journal of Time Series Analysis 27 (3), 381-409.

Christopoulos, D.K., Leon-Ledesma, M.A., (2010), “Smooth Breaks and Non-linear Mean Reversion: Post-Bretton Woods Real Exchange Rates” Journal of International Money and Finance 29, 1076-1093.