Noise to insight: guided portfolio optimization with technical confluence and intrinsic signal shaping

Abstract

In quantitative finance, the realm of the cryptocurrency market, remains relatively underexplored compared to US equities. This project aims to implement a profitable trading solution using reinforcement learning (RL) to outperform baseline strategies such as passive investing and traditional RL methods in the cryptocurrency market, on top of the US equities market. Although much research has been done in the area of reinforcement learning for quantitative finance, traditional approaches often rely on feeding technical indicators as raw features to the agent, assuming it can interpret their utility. This project hypothesizes that explicitly teaching the agent the functional significance and inter-dependencies of these indicators will enable faster convergence to optimal policies. Fundamentally, a key challenge of reinforcement learning is balancing the trade-off between exploration and exploitation during training. Excessive exploration may hinder the agent’s ability to converge to a meaningful policy, while rapid convergence may result in suboptimal policies due to local optima. To address this, our hypothesis proposes leveraging intrinsic rewards to guide the agent towards exploring optimal states early in training. This is based on the observation that agent performance varies significantly with different random seeds, where certain seeds allow the agent to discover effective strategies quickly. While the project is still in its preliminary stages, the proposed methods aim to advance the application of reinforcement learning in portfolio management, by providing insights into effective exploration strategies and improving the interpretability of technical indicators for RL agents. This research has the potential to contribute to the growing field of quantitative finance by offering innovative solutions for navigating the complexities of the financial markets.