Interpretable fuzzy-embedded deep neural network with its application in stock trading

• Main Author: Koh, Amadeus Ying Jie
• Other Authors: Quek Hiok Chai
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166021
• Institution: Nanyang Technological University

Neural networks have become increasingly common over the years, being used in a wide range of applications. They are machine learning models designed to simulate the structure and function of the human brain, consisting of neurons which process and transmit information. Neural networks have the capability of recognizing patterns in data and learning complex functions, and therefore are able to form accurate predictions, commonly in regression or classification problems. They are undoubtedly a powerful tool and bring great value to many industries. Despite offering high-accuracy predictions, neural networks are not perfect. The main drawback of neural networks is that they are essentially a black box. Therefore, there is a lack of interpretability of their results. This is not ideal for applications where high compliance is needed, such as in the fields of traffic or financial management, where we would ideally want to interpret the predictions made by the neural networks and observe how it arrived at its output. In the following study, we explore the use of fuzzy logic, which will be incorporated into neural networks to enhance its interpretability through introducing human linguistic terms. Hence, we propose an interpretable Fuzzy-Embedded Long Short-Term Memory (FE-LSTM). The DBSTREAM clustering algorithm is incorporated into the FE-LSTM to cluster the input data, with the clusters formed being used as the fuzzy membership functions. DBSTREAM is an evolving clustering algorithm that can handle concept drift which is important as the distribution of data can change over time, compromising on the accuracy of the model. The fuzzy output is then defuzzified using centre-of-area defuzzification to return us a crisp output. The LSTM is responsible for mapping the fuzzy inputs to outputs and the tagging of the corresponding fuzzy IF-THEN rules is then performed, which helps us to interpret the outcome. Specifically, we will be experimenting with the FE-LSTM on stock price data. The FE-LSTM will forecast future stock closing prices and coupled with the appropriate trading indicators like Moving Average Convergence Divergence (MACD), trend reversals will be detected and acted upon, either buying or selling a stock position. We will hold multiple stocks in a portfolio which will be rebalanced over time using Reinforcement Learning to maximise profits. The results show that the portfolio management system was able to beat a conventional buy-and-hold strategy.