匯率變動預測：組合預測之應用

匯率預測為國際貿易與跨國投資的重要課題之一，尤其對貿易依存度高的台灣而言更是如此。過去相關文獻提出以多種變數或模型設定預測未來匯率變動，但實證上單一模型難以在所有樣本外期間及不同往前預測期數上的預測獲得一致的成功。因此本文旨在將組合預測 (combining forecasts) 應用於新台幣匯率變動預測，期望達到較單一模型更為準確的預測。本文使用具漂移項的隨機漫步模型作為比較標準模型，並以MSFE評估預測表現，得到以下結論：(1) 單變數ARDL對新台幣匯率變動的預測能力隨時間產生不穩定性，尤其在遠期的預測上。(2) 多數組合預測成功在短期的預測上獲得一致的預測表現提升，而簡單的組合預測方法 (簡單平均法及無時間偏好的Discount-MSFE) 更是成功在所有樣本外期間及所有往前預測期數上一致勝過隨機漫步模型。

關鍵字

匯率變動；樣本外預測；隨機漫步模型； ARDL ；組合預測

並列摘要

Forecasting exchange rate is one of the most considerable issues in terms of international trade and finance, especially for Taiwan. The past researches found that empirically, individual models have inconsistent forecasting performance in different out-of-sample period or over different forecast horizons. Our study then consider several combining forecasts methods to forecast 1, 3, 6, 12 months ahead NTD/USD exchange rate returns. We use random walk with drift benchmark model and MSFE to evaluate forecasting performance in 4 out-of-sample periods. Our empirical results show that (1) individual ARDL has inconsistent forecasting performance during 2008:03-2020:03, especially over longer horizons. (2) all combining methods, except for principal component, consistently outperform random walk with drift model over shorter forecast horizons. And simple combining methods, such as simple mean and Discount-MSFE without time preference, consistently outperform random walk with drift model in every out-of-sample period and over every forecast horizons.