Constructing Smart Beta Trading Strategies with Deep Learning: Evidence from Taiwan Stock Market

• Main Authors: Yu-Chan Hsieh, 謝育展
• Other Authors: Chou-Wen Wang
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/9cs5m4

碩士 === 國立中山大學 === 財務管理學系研究所 === 106 === In this research, Multilayer Perceptron (MLP) and Convolutional Neural Networks (CNN), belonging to Deep Learning, are designed as the investment models. Using Smart Beta factors and technical indicators as model inputs, this paper provides a Deep Learning Strategy Fund strategies. The purpose of this research is to verify whether the deep learning model performs well in the field of financial trading. The followings are the model settings: (1) The frequency of updating and training deep learning model is quarter base. (2)The rebalance of the deep learning portfolio is discussed separately in terms of quarterly frequency or monthly frequency. (3) Taiwan Weighted Index and benchmark portfolio based on Asness(2017) are used as comparing standards with deep learning strategy fund. Under conditions of model settings, the practical trading rules and benchmark standards, using Taiwan stock market as research data, this paper backtests related fund performance from 2007 to 2017. Based on the results of the quarterly balance of the portfolio, it can be seen that the performance of the Deep Learning Strategy Funds is better than the Taiwan Weighted Index, and the final cumulative return is also higher than the benchmark portfolio. In addition, through the feature filter method, the input factors can be refined well, which can further enhance the performance of the Deep Learning Strategy Funds. Under the monthly rebalance of the portfolio, it can be seen that the portfolios updated through the monthly revenue report can further enhance the overall performance compared with quarterly balance of the portfolios. Take the best-performing portfolio as a example, it’s annual return can up to 20.49%, and it’s annual Sharpe ratio is also as high as 116.48%. This result can verify the fact that the Deep Learning model has its feasibility in the practice of fund construction. Finally, the research also conducts a robust test of the Deep Learning Strategy Funds under large-scale fund. From the results, it can be found that the performance is still better than Taiwan Weighted Index and benchmark portfolio. This result also verify that Deep Learning Strategy Funds can not be influenced largely due to large-scale fund.