深度學習法在理財機器人應用之研究： 台灣50成分股公司為例

Abstract

近年來資訊科技的發展與不斷被討論的金融科技創新議題，提高了理財機器人的討論度。不論是銀行提供給財富管理的客戶的需求，或是證劵公司與投信公司因人力縮編而考量自動化選股，因應這波金融科技創新的趨勢，國內外許多單位已經開始投入資源發展相關的服務機制，以符合未來數位化理財投資的需求。 要發展理財機器人，如果沒有深度學習的方法，單靠過去的股票分析機制，像基本面分析或是技術分析，無法透過單一種的分析機制，來取得股票獲利的機會。 深度學習法從1980年福島岡彥提出人工神經網路至今，這將近四十年的期間已經累積了許多方法。 本研究將透過2009年，由賽普•霍克賴特和于爾根•施密德胡伯提出的長短期記憶神經網路(LSTM)，透過實驗設計，考慮股票分析中基本面分析與技術面分析並同時加入型態分析中的酒田戰法的策略邏輯，借重LSTM深度學習的功能，建立高報酬的理財機器人的投資系統。

In recent years, due to the development of information technology and financial technology innovation, the application of financial robots, whether it is provided to the bank's wealth management customers at the bank, or the securities company and the investment company, they plan to develop the service of robo advisors to increase profits to customers and reduce cost for their operation and has begun to invest in resource development related service mechanisms and provide it to customers as one of virtual banking service. To develop robo advisors , if there is no mechanism for deep learning, relying on past stock analysis mechanisms, such as fundamental analysis or technical analysis, there is no way to obtain stock profitability through a single analysis mechanism. The deep learning method has been based on the artificial neural network proposed by Fukushima Okayama in 1980. This has accumulated many methods for nearly four decades. This study will consider both fundamental and technical aspects of stock analysis and simultaneous analysis of long- and short-term memory neural networks (LSTM) proposed by Sepp Hawkett and Jurgen Schmidh Huber in 2009. Joining the strategy logic of Jiuquan's tactics in the type analysis, through the experimental design, through the mechanism of deep learning, the investment system of high-paying wealth management robots is established.