運用多模式深度強化學習與語義影響模型於交易策略

Applying multimodal deep reinforcement learning with sentiment Influence model to the trading strategies

陳郁夫(Yu-Fu Chen)

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(2018年)

摘要

隨著近年來深度學習蓬勃發展，不少研究使用深度學習處理金融交易問題。其中
增強式學習由於其能夠隨時間動態產生相應決策，相近於人們投資決策之過程，除此
之外多型態學習藉由結合不同訊源之資料，能夠有效增強模型之預測力。另投資人們
常於投資決策過程中，常閱聽新聞等社群媒體之資料作為投資參考依據，而語意分析
能夠對這些資料做正負語意的量化，藉此判斷財金新聞放出的消息好壞。本研究結合
上述技術，訓練專員利用財金新聞進行多型態增強式深度學習於股票交易上。並對新
聞對股票市場之影響力進行量化及建模，觀察真實世界事件對於股票市場之影響力變
化。本研究以美國標準普爾500 指數(S&P 500) 之成分股公司為研究對象，並對Apple,
Google 進行多型態增強式股票交易，以最大化獲利報酬為目標，增加專員之投資獲益
能力。
本研究的實證結果可以得出以下結論:
1. 本研究所提出之多型態增強式股票交易系統，在長期間交易結果下，多型態專員
的交易結果表現優於僅關注價格資訊之專員，在對Apple 及Google 之股票交易
中，其獲利點數分別最高優於對照組143% 及1.36%。佐證多型態增強式交易為
可行，且金融新聞為反應市場要素有效之多訊源其一。除新聞資料外，我們更引
入Form 8-K 報告，以佐證本系統對於不同資料之泛用性。
2. 本研究所提出新聞對股票市場影響力之模型，能有效表現出其資訊傳遞及影響
力對市場之衝擊。並以此影響力模型投入多型態股票交易，能夠提升原有多型態
專員之獲利能力，於Apple 提升最高43% 之獲利能力，於Google 提升2% 獲利
能力。

關鍵字

多模式學習 ； 增強式學習 ； 語意影響力模型 ； 股票交易策略

並列摘要

With the growth of Deep Learning, there are more studies use deep learning on financial
problems with different kinds of popular techniques. Some people eager to develop agents for
stock trading to profit automatically and reinforcement learning is known as good at based on
different situations to make dynamic decisions that can be applied to trade problems.
In this thesis, we make reinforcement agents act like experienced investors, the agent should
not only concern the price fluctuations but also consider the news information. So we introduce
Multimodal learning which can merge different modalities data to enhance the performance of
model and Sentiment analysis for understanding the sentiment meaning of news to make trading
agents. Besides, our agents can have special insights on the impact news reports brings to the
market through our influence model.
We find out some conclusions listed below through experiments results using S&P 500 index
components companies:
1. In our proposed multimodal reinforcement stock trading system, multimodal agents outperform
than only price-concerns agents with at most 143% or 1.36% lead on average
profits points. It proves that multimodal reinforcement stock trading system is useful and
financial news is one of the effective modalities in our system. Furthermore, we introduce
the Form 8-K to prove our framework’s universality to different data.
2. Our proposed influence model has the ability to shape the news impact to the stock market
and it could be an aid for our multimodal agents to evaluate the status of the market. With
the assistance of influence model, the multimodal agents strengthen their profitabilities
with surpassing original multimodal agents with 43% and 2%.

並列關鍵字

Multimodal learning ； Reinforcement learning ； Sentiment influence ； Stock trading strategy

參考文獻

[1] M. Längkvist, L. Karlsson, and A. Loutfi, “A review of unsupervised feature learning and

deep learning for time-series modeling,” Pattern Recognition Letters, vol. 42, pp. 11–24,

2014.

[2] L. Di Persio and O. Honchar, “Artificial neural networks approach to the forecast of stock

market price movements,” International Journal of Economics and Management Systems,