Efficient Multi-Period Trading and Learning Techniques for Portfolio Optimization
Open Access
Author:
Diwan, Ishaan
Area of Honors:
Computer Science
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Mehrdad Mahdavi, Thesis Supervisor John Joseph Hannan, Thesis Honors Advisor
Keywords:
machine learning portfolio optimization stocks trading boosted decision trees decision trees Markowitz mean-variance portfolio theory
Abstract:
This thesis explores a new approach to multi-period portfolio optimization under the framework of the mean-variance portfolio theory developed by renowned economist Harry Markowitz. Two machine learning models, decision trees and gradient boosted trees, are introduced to predict the returns of a real-life portfolio consisting of three stocks. The models are learned using fundamental financial data for each company in the portfolio. These predicted returns are used to perform multi-period optimizations on the real-life portfolio over the course of a two-year investment horizon. The results of the multi-period optimizations using the predicted returns of the portfolio are compared to the results of the same optimizations performed using the historical returns of the portfolio. This thesis explores two different algorithms to rebalance the portfolio at the end of each period: a naive approach which randomly generates portfolio weights and the Markowitz mean- variance optimization scheme. The results of this thesis show that gradient boosted trees provide a promising model to predict the returns of the chosen portfolio over time as well as that fundamental financial data does not always prove fruitful when predicting the returns of individual stocks.
