Unit:
Department of Informatics and TelecommunicationsΠληροφορική
Supervisors info:
Ιωάννης Παναγάκης, Αναπληρωτής Καθηγητής, Τμήμα Πληροφορικής και Τηλεπικοινωνιών, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών.
Original Title:
MACHINE LEARNING METHODS FOR PORTFOLIO OPTIMIZATION
Translated title:
Machine Learning Methods for Portfolio Optimization
Summary:
In this thesis, we consider the problem of Markowitz Portfolio Optimization. It is defined as attempting to minimize the variance of a diversified investment's returns. We use several conventional Machine Learning techniques to solve it, namely CVXpy, CVXpy-layers, Proximal and Projected Gradient Descent. We also propose a Deep Learning approach, which uses an LSTM unit. As investment units to train our models, we use the historic returns of 48 industry sector portfolios from 2019 to 2021(FF48 daily returns). Four of our models including our Deep Learning approach manage to surpass the performance of the equally weighted portfolio which is considered a tough benchmark in this problem. Finally, we propose modifications for further improvements.
Main subject category:
Technology - Computer science
Keywords:
markowitz portfolio, optimization, cvxpy, projected gradient descent, lstm
