Portfolio optimization is a well-studied optimisation problem in both machine learning and computational finance. The task is to optimally allocate wealth across various assets such that the returns are maximized. In a world where market prices are dynamic and can often be considered adversarial, the assumptions of statistical machine Learning fail. Online learning provides a framework for studying such optimization problems. In this work, we compare and contrast different stateof-the-art online learning algorithms using a realworld dataset of different exchange-traded funds. We observe the superior performance of certain classes of algorithms over others during certain market trends. To take advantage of this discrepancy in the performance of different base experts, we propose meta-learning as a solution to combine these base experts. In this work, we present different Nth-degree meta-learners and a particleswarm-based meta-learner to assign the optimal weights to the decisions of experts. We present both a regret and run-time analysis for these methods. Experimental results on real-world stocks are also presented.
Contact Authors at hrithik_nambiar@brown.edu / lukas_jarasunas@brown.edu for further queries or for full project report.