基於深度學習概念之金融市場價格預測

Prediction of Price Trends in Financial Markets based on Deep Learning Technique

黃君平(Chun-Ping Huang)

指導教授 ： 陳安斌 ； 黃思皓

國立交通大學/管理學院/資訊管理研究所/碩士(2016年)

摘要

本研究將深度學習的概念與技術，應用於金融市場上價格趨勢的特徵提取，嘗試分類並預測未來的趨勢，並以台灣指數期貨的時間序列資料作為研究對象。本研究中透過四種不同的方式將一維時間序列資料進行二維化，使用有著在圖像資料分類辨識上有著卓越成果的卷積式類神經網路進行特徵的捕捉，並做出未來趨勢的分類，藉此達到預測未來的效果。研究中藉由四個階段的實驗達到結果的最佳化，並在最後與傳統反饋式類神經網路進行分類準確率的比較，以驗證卷積式類神經網路在由一維市場時間序列資料轉變而成的二維資料上的分類預測能力，並讓深度學習能有機會更進一步應用在金融領域中。
經過實證研究，卷積式類神經網路在本研究的資料上確實有其分類預測能力，並有機會可獲得與反饋式類神經網路同等或更好的結果；另外，將複數組在同一時間點所得出之二維資料一同作為輸入，以涵蓋更多資訊，將可能得到比原本的單組輸入更好的成果。本研究得出以上的結論，讓深度學習的概念將有更多機會可應用在財金領域中，也可能在其他的領域上獲得更好的成果。

關鍵字

深度學習 ； 資料二維化 ； 價格趨勢預測 ； 卷積式類神經網路 ； 機器學習

並列摘要

This paper proposed a novel financial time-series analysis method based on deep learning technique. The feature extraction step in intelligent trading system is an important data representation and modeling process. Traditional approaches are used to applying several technical indicators and expert rules to extract numerical features. The major contribution of this paper is to improve the algorithmic trading framework with a data-driven concept and convolutional neural networks (CNN). The simulation experiments are implemented and benchmarked in the historical prices of Taiwan Stock Index Futures. The first step of our proposed method is to clip the learning samples and testing samples from the historical data. Then, four different kinds of data visualization methods are proposed to transform the original one-dimensional time-series price data into more representative two-dimensional planes. Then, the deep learning algorithm based on CNN is implemented to model, classify, and predict the price trends in the future with repeatedly supervised and unsupervised learning. The discriminant feature extraction methods can also be trained automatically in the proposed framework. Finally, the benchmarks with various numerical experiments can evaluate the classification accuracy and optimize the settings of system parameters. The experimental results show that the deep learning can effectively classify the financial time series data without the domain knowledge of trading experts. The proposed two-dimensional data representation method and CNN learning algorithm also outperform various traditional approaches which used in the control groups of experiments. In summary, the deep learning technique is proven to be effective in our trading simulation application, and may have greater potentialities to handle the noisy financial data and complex problems in social science

並列關鍵字

deep learning ； data visualization ； trend prediction ； convolutional neural network ； machine learning

參考文獻

[2] Xu, Z., MacEachern, S., & Xu, X. (2015). Modeling Non-Gaussian Time Series with Nonparametric Bayesian Model. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 37(2), 372-382.

[3] Platanios, E., & Chatzis, S. P. (2014). Gaussian Process-Mixture Conditional Heteroscedasticity. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 36(5), 888-900.

[4] Bengio, Y., Lauzon, V. P., & Ducharme, R. (2001). Experiments on the application of IOHMMs to model financial returns series. Neural Networks, IEEE Transactions on, 12(1), 113-123.

[5] Tiňo, P., Schittenkopf, C., & Dorffner, G. (2001). Financial volatility trading using recurrent neural networks. Neural Networks, IEEE Transactions on,12(4), 865-874.

[6] Abu-Mostafa, Y. S. (2001). Financial model calibration using consistency hints. Neural Networks, IEEE Transactions on, 12(4), 791-808.