Cycle-based trading & portfolio management system

• Main Author: Zhan, Xiaoying
• Other Authors: Quek Hiok Chai
• Format: Final Year Project
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/66823
• Institution: Nanyang Technological University

This research project aims to apply machine learning techniques in the area of financial investment. By adopting data-driven objective methods, some common human biases known to prevent investors from making rational decisions could be reasonably avoided. The strategy is built upon cyclical movements in the stock market, which is mainly induced by business cycles. Thus, the target horizon is mid to long term. By selecting stocks at their troughs and investing capitals during the rising phases, capitals could be utilized more efficiently to preserve values and generate returns. To predict the inflection points in stock prices, Takagi-Sugeno-Kang fuzzy neural network is adopted due to its accuracy. To improve its performance, Evolutionary Algorithms (EA) are applied to fine tune the model’s parameters. In addition, angular coding scheme is used to conquer the problem of limited search space associated with the designing of TSK Fuzzy Rule-Based System with EAs. After the longer term inflection signal is given, entry/exit points are confirmed by shorter-term signals such as MACD, which reflects more recent market changes. Maximum reward reinforcement learning is also incorporated to estimate the potential rising amplitude in order to avoid entering into unprofitable trades while taking into account transaction costs. The cycle-based stock selection approach is combined into the design of a portfolio management system based on Markowitz Portfolio Theory. The system constructs portfolios with the objective of maximizing return while maintaining overall risk at a predefined target level. Rebalancing is scheduled according to the Larry Swedroe 5/25 rules, which enables prompt response to significant market changes. The proposed cycled-based strategy achieves average annual return of around 14%. Compared to the benchmark (S&P) annual return of 9% during the same back-test period, the system makes a significant improvement.