Hybrid deep neural network and deep reinforcement learning for algorithmic finance

• 主要作者: Ooi, Min Hui
• 其他作者: Quek Hiok Chai
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2022
• 主題: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• 在線閱讀: https://hdl.handle.net/10356/157234
• 機構: Nanyang Technological University

Deep learning is a recent breakthrough in the field of machine learning that has greatly improved predictive and modelling capabilities. While there are many significant achievements using deep learning in fields such as natural language processing and recognition problems, the application of deep learning in finance is still heavily being researched. Traditional prediction models utilise deep neural networks, but face difficulty achieving high levels of accuracy when solving complex problems. Additionally, such models lack interpretability which could prevent informed decision making using these models. This paper proposes a hybrid fuzzy deep neural network architecture. The proposed architecture consistently obtains high accuracy levels despite complex problem definitions and datasets. Furthermore, by embedding fuzzy logic, the model enables meaningful interpretations and insights surrounding the derivation of predictions through use of fuzzy rules. The proposed architecture was applied to the complex stock price prediction problem and maintained the high levels of accuracy, while increasing interpretability. The predicted stock prices were used in calculations of technical indicators such as the MACD to generate a better analysis of market trends and enable better informed trading decisions. Using deep learning as a method to solve complex problem often comes with error-prone and arduous development and debugging. This paper proposes a deep reinforcement learning (DRL) architecture that delivers good performance when dealing with complex problems. Furthermore, the proposed architecture is easily extendable to other complex problems, due to ability to change and adapt to environments. The proposed DRL architecture was applied to portfolio management. Different portfolio constraints were added to the environment, and the trade-offs between each portfolio decision and constraint under different market conditions were observed. An investor can use these results to weigh trade-offs and make more informed decisions in portfolio management.