標題: | 藉深度學習與市場輪廓開發智慧型台指期交易模型 Intelligent Trading Model developed by Market Profile and Deep Learning in Taiwan Index Futures |
作者: | 吳昇晏 陳安斌 Wu, Shen-Yen Chen, An-Pin 管理學院資訊管理學程 |
關鍵字: | 深度學習;卷積式類神經網路;市場輪廓理論;技術指標;TensorFlow;Deep Learning;Convolutional Neural Network;Market Profile;Technical Indicators;TensorFlow |
公開日期: | 2017 |
摘要: | 傳統技術分析在各個國家流傳已久,各種判斷方式層出不窮,部分所謂的名師或預測軟體亦透過各種平台闡述自己的分析教學並藉此收取費用,然而真正穩定獲利的策略卻僅在少數。再加上傳統技術指標皆以計算過去歷史資料的統計手法導出特定公式,其公式通常無法涵蓋市場背後真正隱含之意圖、或只涵蓋部份意圖,造成交易人員誤用而虧損連連。
而市場輪廓理論主要核心著重在於在市場交易人員(包括自己)的心態本質,例如貪婪跟害怕等。因此若能透過市場輪廓理論來強化技術指標,或能建立一個互補之交易模型,以利投資者進行使用。
本研究將以十五分鐘K線為主,搭配各種傳統技術指標之組合,如隨機指標(KD)、移動平均線(MA)、 指數平滑異同移動平均線 (MACD)等,以及市場輪廓理論中的特徵值來建構台指期當沖交易策略,再將各種表現優良之策略透過卷積式類神經網路進行學習,最後檢視各種策略之損益情況,藉此確認卷積式類神經網路是否能透過上述策略來建構一智慧型交易模型。
本研究實證傳統技術指標確實能透過市場輪廓理論進行強化,將原本每口損益1.23點提升至2.26,增幅179%;若策略單純以市場輪廓為架構,則每口損益相較於傳統技術指標增幅至少740%。而透過卷積式類神經網路所建構之模型,績效亦有所成長。 Traditional technical indicators have been widely used in various investment decision making. Some securities analysts or investment assistant software even tried to charge the usage fees of their secret indicators. However, few traders or investors would make robust and effective profits by reference to this kind of tools. The formulas of traditional technical indicators were deriving by statistical methods from the historical data in financial markets. These formulas are not enough to represent the real intention of the markets, which causes the trading information or buy/sell signals don’t have reproducibility. The core of the Market Profile focuses on the modeling of human behaviors and psychology, such as greedy and fear. Thus, it have the opportunity to develop a complementary trading model by using Market Profile Theory to enhance traditional technical indicators. First, this thesis examined Market Profile features and traditional technical indicators, such as Stochastic (KD), moving average (MA), Moving Average Convergence Divergence (MACD), Value Area (VA) and Point of Control (POC), by a trading simulation in historical datasets. The valid and profitable factors will be collected into the learning feature pools and deep convolutional neural networks will be applied to construct the trading decisions and strategy. The experimental results show that the intelligent trading system can be improved by Market Profile Theory dramatically, which can earn 2.26 points per lot in TAIEX future trading compared to 1.23 points of its original rule-based version. |
URI: | http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070463406 http://hdl.handle.net/11536/141786 |
顯示於類別: | 畢業論文 |