大數據分析與LSTM模型於股市投資策略之研究－以台灣各類股為例

Big Data analysis and LSTM Model investment strategy in the stock market – take Taiwan corporate stocks as Examples

蔣明哲(Ming-Zhe Jiang)

指導教授 ： 李鴻璋

淡江大學/商管學院/資訊管理學系碩士班/碩士(2021年)

https://doi.org/10.6846/TKU.2021.00783

摘要

低薪及低利率已成為台灣社會的常態，使得民眾開始尋找報酬較高的風險性資產投資，而台灣股票市場擁有報酬率高且進入門檻低等特性，使得進入股票市場投資的人數日益漸增。
股票投資的諸多決策以基本分析及技術分析為主，其中，基本分析需具備相關專業知識，而技術面則可較簡易的透過技術指標進行進出場決策，加上近年來隨者AI技術的發展，使得AI預測股價的方式興起，因此本研究將以技術指標策略，與AI股價漲跌預測兩部分作為探討。
本研究分兩部份探討，第一部份為探討對加權指數、各產業類股，所有上市櫃公司，以10種單一技術指標(RSI、KD)策略，於2004年~2019年進行模擬分析。第二部份則為探討長短期記憶模型(LSTM)之最佳模型參數與最佳交易策略，以2009~2015年為訓練期，價、量與技術指標(RSI、KD)為參數訓練LSTM模型，並回測2016~2019年類股資料，。
第一部份結果得出，於加權指數上，每年買入持有策略可獲得高於技術指標之報酬。於產業類股上，沒有單一技術指標策略可使所有產業之報酬率皆高於大盤。但於其中2個產業上(如存託憑證、農業科技業等)使用特定指標策略，卻可勝過加權指數及買入持有之報酬率。於個股上，使用技術指標策略之年化報酬率皆低於買入持有(7.09%)，但於股市下跌期間可使用RSI交叉策略獲得較買入持有高的報酬。
第二部分結果得出，在純價、量LSTM模型預測股價漲跌中，加入技術指標為參數，能提高模型預測之準確率，其中，於價、量模型加入RSI12指標後，測試準確率達77.62%，為最高準確率之參數組合。此外，本研究使用之7種模型，平均報酬率皆高於單一技術指標與買入持有策略之報酬率，而使用價、量與RSI12參數組合之模型，擁有13.33%之最佳報酬率，高出純價、量模型近2.2%之年化報酬率。

關鍵字

大數據 ； 深度學習 ； LSTM ； 技術指標 ； RSI ； KD ； 台股

並列摘要

Low salaries and low interest rates have become the norm in Taiwan’s society, making people start looking for risky asset with higher profit. Taiwan’s stock market has features such as high profit and low entry gap, which makes the number of investors entering the stock market increasing.
Many stock investment decisions are based on fundamental analysis and technical analysis. Among them, fundamental analysis requires relevant professional knowledge in human domain, while technical analysis can be much easier to make buy-sell decisions through pure technical indicators. Recently, with the development of AI modling, AI modle combines with technical parameters to depict stock price become more and more popular.
This study includes two parts. The first part is to discuss all the Taiwan Capitalization Weighted Stock Index(TAIEX), various industry indexes, and all listed companies etc., by using 10 kinds of technical indicators (e.g., RSI, KD) strategies, and conducts simulation analysis by using the historical datas from 2004 to 2019. The second part is to discuss the parameters and trading strategies in the long short-term memory model (LSTM). Seven kinds of the the parameters such as the price, volume and technical indicators (RSI, KD) are feed into the LSTM model as a training data for the period from 2009 to 2015 , and back-test the stock data for the period from 2016 to 2019.
The result of first part show that on TAIEX, using pure buy-and-hold strategy can get higher annual profit than those using technical indicators. As for kinds of industry indexes, technical indicators might get better result than that of buy-and-hold strategy, but these indicators vary among them. However, using specific technical indicator strategy in two of these industry indexes (e.g, depository receipts, agricultural technology) can get higher profit than TAIEX. For individual stocks, using the technical indicator strategy is lower annual profit than that of buy-and-hold, but during the downturn of the stock market, the RSI cross strategy can be used to obtain higher profit than that of buy-and-hold.
The second part of the results shows that adding technical indicators along with the price and volume, LSTM model can both improve the accuracy of the model prediction and get better results than those of using buy-and-hold strategy. Among them, after adding the RSI12 indicator to the price and volume model, the test accuracy is up to 77.62% is the parameter combination with the highest accuracy. In addition, the average profit of the seven kinds of input data combination obtain higher annual profit thanthose coming from both the profit of single technical indicator and buy-and-hold strategy. Among these seven ones, the input data compose with price, volume, and RSI12 get the best annual profit by 13.33%. This is nearly 20% increase than that of input data with the pure price and volume, which is previous studied.

並列關鍵字

Big Data ； LSTM ； KD ； RSI ； Taiwan Stock Market

參考文獻

英文文獻

[1] Nguyen Lu Dang Khoa, Kazutoshi Sakakibara and Ikuko Nishikawa, Stock Price Forecasting using Back Propagation Neural Networks with Time and Profit Based Adjusted Weight Factors, page 5, 2006.

[2] Stefan Angrick1 & Naoko Nemoto1, Central banking below zero: the implementation of negative interest rates in Europe and Japan, pages 4-23, 2017.

[3] Luca Di Persio and Oleksandr Honchar, Recurrent neural networks approach to the financial forecast of Google assets, pages 10-12, 2017.

[4] G.E. Hinton and R.R. Salakhutdinov. Reducing the dimensionality of data with neural networks.Science, 313(5786):504–507, 2006.