Smart Trading: A Novel Reinforcement Learning Framework for Quantitative Trading in Noisy Markets

The trading market's complexity is heightened as capital asset prices can be significantly influenced by traders' emotions. While the FinRL library provides a state-of-the-art reinforcement learning framework for training agents to trade in markets, it lacks the necessary approaches to counteract market noise and boost the agent's learning process in the complex environment. This paper proposes a novel reinforcement learning framework for quantitative trading to empower the agent to operate more effectively within the intricacies of noisy markets. Discrete features are used as inputs instead of continuous features to reduce the complexity of input features. This discretization process reduces market noise, thereby simplifying the agent's learning process as the feature space becomes more manageable after discretization. A theorem is introduced as a guide for choosing discrete features based on the available sample size. Within the trading environment, an adaptive scalar is employed to eliminate the influence of historical trends to prevent agents from blindly adhering to these trends over input signals. Additionally, a low-pass filter is applied before the computation of immediate rewards to facilitate the model's training process. Experiments are performed on different datasets to demonstrate that the agents trained with the proposed framework can earn excess returns across markets.