匯率預測的困難，可以由截至目前尚未有任何單一模型，得以在短區間預測打敗隨機漫步模型得到印證。文獻上常用以預測匯率的模型，包括代表長期均衡的貨幣學派模型，以及源自於市場無套利條件的遠期外匯溢酬模型。本文利用Liu and Kuo（2014）的模型平均（model averaging）法，透過組合這些單一模型，對美元兌新臺幣匯率進行預測。組合模型進行預測，除了充份利用所有可得的訊息外，也免去先驗選擇單一模型可能產生的風險。在極小化預測均方誤的前提下決定各模型的最適權重後，本文採用的組合預測可以在樣本期間內，得到顯著優於前述單一模型及隨機漫步模型之表現。其大幅降低單一模型的累積預測平方誤（cumulative sum of squared forecasting error），甚可達隨機漫步模型的90%以上。這樣的預測表現，源自於資訊的充份應用，以及能依據各單一模型表現的偏誤與變異，調整與時而異的組合權重。我們的樣本期間，包含了金融風暴及美國採行量化寬鬆貨幣政策等重大經濟事件，也使得單一模型對匯率的準確預測更顯困難。本文同時發現，各單一模型偏誤與變異表現，在這些經濟事件發生時，發生明顯的轉折。這意謂在某一時間點為「最適」的單一模型，可能隨時空改變而非最適。組合預測模型也正因為能適時依據這些變化，調整各模型的權重，而能更有效結合訊息，提供較任一單一模型為佳的預測結果。

The difficulty of forming accurate exchange rate forecast has manifested itself by inabilities of some existing models, including monetary model and forward premium model, to beat a random walk. The paper adopts a different approach to the forecasting exercise that combines these existing models. The approach not only makes best use of available information, but also is free of model selection risks. The forecast performance of the combination model is found to outperform those of any single aforementioned model and a random walk in the samples. Specifically the cumulative sum of squared forecating errors of our combination model is remarkably reduced. The reductions in forecast errors can be attributed to the time-varying weights that are assinged according to the relative magnitudes of bias and variance of each considered model. Moreover, the samples span over US subprime crisis and quantitative easing, where each of the considered models finds it not easy to yield good forecast on exchange rate movements. Associated with the finding is that the corresponding bias and variance of each considered model display dramatic shifts in these recent global economic events, implying that the combination is able to extract useful information from each considered model alone to yield more accurate exchange rate predictions.