Recurrent neural network embedded fuzzy (RNNEFS) system with its applications in stock market forecasting & MACD trading strategies

• Main Author: Chen, Yinya
• Other Authors: Quek Hiok Chai
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/147965
• Institution: Nanyang Technological University

In recent years, artificial neural networks have been used extensively in many real-world applications. However, high accuracy, driven by the increased complexity, often comes with the sacrifice of interpretability – many ANN models have black-box behavior and fail to provide explanations for the predictions. Some techniques have been developed to solve this problem including neuro-fuzzy systems which use neural networks to determine the parameters for fuzzy sets and fuzzy rules. These techniques may suffer from limited accuracy and learning abilities. This paper proposes a novel architecture called Recurrent Neural Network Embedded Fuzzy System (RNNEFS) which combines fuzzy system and neural network in an innovative embedded manner. This new system aims to overcome the limitations of each isolated paradigm and combines their respective strengths. It has following advantages: (1) capable of capturing sequential characteristics from time series data; (2) high interpretability and transparency of the predicting process; (3) high robustness in various environments. RNNEFS transforms the crisp into fuzzy input and simultaneously feed it into the parallel recurrent neural network and fuzzy system. The embedded RNN is used to learn data patterns and extract useful information instead of fuzzy rules generation in the traditional neuro-fuzzy systems. Neurons in RNN are connected with fuzzy rule nodes through a tagging mechanism which provides high interpretability of the predicting process. The proposed model RNNEFS is incorporated with genetic algorithm and applied in stock market forecasting and trading. Genetic algorithm efficiently searches for the optimal hyperparameters used in the trading strategy. The system is capable of reducing time lags in indicators and generating timely buy and sell signals to maximize return. The performance of the proposed trading system is benchmarked against existing strategies and the results are highly encouraging.