Stock selection using machine learning

• Main Author: Shen, Nan.
• Other Authors: Vivekanand Gopalkrishnan
• Format: Final Year Project
• Language: English
• Published: 2009
• Subjects: DRNTU::Engineering::Computer science and engineering::Computer applications::Computers in other systems; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/17042
• Institution: Nanyang Technological University

The final year project involves an empirical investigation of the predictability of stock returns and feasibility of building a stock selection application using machine learning. The purpose of the project is to build a successful application of machine learning on S&P 500 data. The proposed application is capable to predict future stock returns and make wise decisions to select stocks and construct portfolios. Econometric analysis and empirical financial modeling involving machine learning methods are widely studied and adopted both in literature and practice. However, prediction models employing artificial intelligence techniques such as neural networks and decision trees are used arbitrarily or implemented in ad-hoc scenarios, such as feature selection, data preprocessing, portfolio construction methods, and evaluation criteria etc. There has never been a proper and comprehensive survey particularly on stock selection applications. Also, economic indicators, which proved to be an important factor in predicting stock returns by economist, have rarely been used in machine learning applications. In this project, a comprehensive and detailed survey on machine learning applications on stock selection is presented. A framework is proposed to take considerations of all possible choices and parameters in relevant practical issues. Thereafter, a model-tree based stock selection application using technical, fundamental and economic indicators is built with a Greedy Search procedure in parameter selection. We show its validity and superiority by comparing against both S&P500 benchmark and several previous machine learning applications. And finally we apply our model for a case study on the financial crisis periods and demonstrate its capability to provide economically significant profits for investors.