- 學位論文
Google 的搜尋熱度是否有助於預測匯率？
Does Google Trends-ruled Exchange Rate Predictions Beat Random Walk?
摘要
在匯率預測的研究上，Meese and Rogoff（1983）發現基本面模型的預測力甚至不及隨機漫步模型，經歷了多方研究的驗證後，隨機漫步模型逐漸立於不敗之地。Rossi（2013）從這些研究歸結出兩點：只有在長天期預測時，基本面模型才能打敗隨機漫步，且其預測能力因時而變。本文以創新的方式，試圖捕捉因時而變的市場訊號。 我們選擇使用Google Trends搜尋熱度（下稱GTI），每日提取外匯市場的訊號，進行匯率變動的方向性預測。不同於過去研究時常採用的中至長天期預測，本文憑藉著GTI能被即時觀察及取得的性質，進行短天期預測。我們最大的創新在於，透過非線性的雙層預測方法及簡單的加總預測，捕捉並統合外匯市場因時而變的特質。此雙層方法係以GTI在第一層篩選預測模型，並透過模型於第二層預測匯率變動方向。 儘管雙層預測的結果再次確認了預測能力因時而變的性質，最終在簡單的加總預測中，本文的預測在統計上取得了相對隨機漫步更高的勝出率。這確認了GTI在追蹤此變動性質的總體力量，並證實以GTI在短期捕捉此變動性質的合理性。
並列摘要
Since Meese and Rogoff (1983), fundamental models’ inabilities to beat random walk in exchange rate predictions have been widely documented. It is concluded by Rossi (2013) that the fundamental models can only beat random walk at long horizons, and the predictive ability at most be time-varying and occasional. Our research invokes another new attempt that deals directly with the varying predictive ability for fundamental models across time. We ask whether the time-varying nature can be tracked straightforward, when signaling to market. If the signals can be detected and received, there is a higher chance to improve the exchange rate predictability. Tapping into the information extracted from Google Trends, we check if the signals are captured and reflected. Unlike past studies where predictions were usually conducted at medium-to-long horizons, we focus on out-of-sample daily predictions at short horizons. Given the observable and obtainable real time Google Trends index (GTI), we justify the high forecast frequency. Aside from that, our predictions greatly differentiate from past studies with an easy yet novel 2-layer approach following an aggregated result. For the 2-layer approach, we have GTI-ruled model selection in the first layer and predictive models in the second layer. Rather than evaluating the performance in statistical sense, our study places an emphasis on that of the directions of change. The results for the 2-layer predictions though reaffirm the time-varying nature, by counting the statistical success and failure, a higher rate of beating the random walk confirms GTI’s aggregate power in tracking the such nature. The aggregated predictions further legitimate the use of Google Trends search intensity in capturing such time-varying nature at short horizons.