運用機器學習方法預測外匯資料變化 以日圓對兌美元為例

Using Machine Learning Model in Predicting Japanese Yen to US Dollar Forex Data

蕭亦辰(Yi-Chen Shiao)

指導教授 ： 陳榮靜

朝陽科技大學/資訊學院/資訊管理系/碩士(2020年)

摘要

時間序列資料可以揭示現實世界的變動，例如溫度、雨量、外匯資料等等。因為時間序列有多種複雜的形式，所以模型也需要針對資料的不同而做變化。舉例來說，Auto Regressive 可以有效率的處理平穩型時間序列。但當遇到不平穩的時間序列，AR模型的效果就比較差。
在此研究，我們專注於外匯資料的處理。外匯的資料屬於不平穩的時間序列，也因此非常難以用統計模型與平穩行時間序列的模型來預測。因為影響到外匯資料的變數非常多，包含國民年均所得，中央銀行的基礎利率，與兩國之間的外交等等。如果單看收盤或者開盤價來預測，在不知道其他變數的變化情況下，數據就會趨向隨機變化。但透過近期整合機器學習模型如隨機森林等，可以預測得比以往的模型還要來的準確。
過去的研究大多著重在回歸的分析上，例如我們的前一個研究就是對於外匯資料的收盤價來進行預測。我們發現在資訊有限的情況下，也就是沒有其他特徵值的變化，要做出精確的預測收盤價的數值是比較困難的。所以此研究著重在分類外匯資料的未來走向。此研究運用了隨機森林與支援向量機和梯度提升機器來比較不同輸入特徵時的差別，並且針對多元分類與二元分類的輸出來進行比較。我們也提出了一個資料處裡的方法，來幫助模型更有效率的預測並分類。而結果顯示我們所提出的方法是可以有效幫助模型去分類資料的。

關鍵字

機器學習 ； 支援向量機 ； 隨機森林 ； 分類 ； 外匯資料 ； 時間序列

並列摘要

Time series data shows the dynamic changes of real world data, such as temperature, rainfall, Forex rate, etc.. Due to the complex changes of time series data, models differ from interpreting data best. For example, Auto Regressive (AR) models are commonly used in stationary data. As for non-stationary, it performs worse and cannot predict the data well.
In this research, we focus on the changes in Foreign exchange data. Forex data is a non-stationary time series, making it difficult to model with stochastic process and stationary time series models because the number of features influencing the Forex rate is large. Such as the Gross domestic product (GDP), base interest rate, and the relation of the two countries. If we only look at the closing price, the data might seem random because we cannot see the changes of the other features.
Unlike usual time series modeling uses regression methods, as our last research focused on predicting the closing value of each timestep. We found out that without the changes of other influencing features, it is hard to predict the closing price accurately. So, this research aims to classify the trend of future steps. Random Forest and Support Vector Regression are used to compare the efficiency of modeling both multi-class and binary class classification when having a different number of input features. We also proposed a method to process the data to make the classification more accurate. The result shows that our method can increase performance efficiently.

並列關鍵字

Machine learning ； SVM ； Random forest ； classification ； Foreign exchange rate ； Time series

參考文獻

[1] P. J. Brockwell and R. A. Davis, Introduction to Time Series and Forecasting, 3rd ed., 2016.

[2] T. Y. Kim, K. J. Oh, C. Kim, and J. D. Do, "Artificial neural networks for non-stationary time series," Neurocomputing, vol. 61, pp. 439-447, 2004.

[3] Z. Liu, "Chaotic Time Series Analysis," Mathematical Problems in Engineering, vol. 2010, 2010.

[4] R. Z. Aliber, "The Interest Rate Parity Theorem: A Reinterpretation," Journal of Political Economy, vol. 81, no. 6, pp. 1451-1459, 1973.

[5] D. Pradeepkumar and V. Ravi, "Soft computing hybrids for FOREX rate prediction: A comprehensive review," Computers & Operations Research, vol. 99, pp. 262-284, 2018.