Portfolio Management via Reinforcement Learning

Period: January 2019 – December 2021
Funding: Innovation and Technology Fund (ITF), Hong Kong, Reference No. UIM/365
Industry partner: Magnum Research Limited (AQUMON)

Problem

Standard portfolio rebalancing rules are based on fixed-weight targets and threshold-trigger mechanisms that do not adapt to evolving market conditions or explicitly account for transaction costs. A principled sequential decision-making framework is needed to learn rebalancing policies that remain effective under diverse market regimes and realistic cost structures.

Method

The portfolio rebalancing problem is formulated as a Markov Decision Process (MDP) with state variables capturing portfolio composition, market features, and risk indicators; actions corresponding to rebalancing decisions; and a reward function that penalises both tracking error and transaction costs. We explored value-based and policy-gradient reinforcement learning algorithms, including recurrent and sequence-based architectures that condition on the trajectory of market states. Emphasis was placed on sample efficiency, robustness to distribution shift, and deployability in a production rebalancing pipeline.

Outcome

The project produced a research-to-production foundation for RL-based portfolio management, with documented methodology, experimental benchmarks, and a reusable framework for systematic strategy development and evaluation at AQUMON.