A Reinforcement Learning Framework to Minimize Reward Latency in Quantitative Trading

Nawaz, Sharjeel (2025) A Reinforcement Learning Framework to Minimize Reward Latency in Quantitative Trading. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (922kB) | Preview
Preview	PDF (Configuration Manual) Download (640kB) | Preview

Abstract

Quantitative Trading uses mathematical models to execute trades based on pattern learn from historical stock data. In recent research, most of the mathematical models are reinforcement learning models .However, In reinforcement learning (RL), a key challenge arises from delayed rewards, particularly in quantitative trading environments. For example, the Model may buy a stock but never sell it, so there's no actual reward to evaluate whether the decision was good or not. Therefore, it is difficult for a model to optimize long-term profit by learning from delayed rewards. This presents a significant hurdle in training RL agents effectively for financial market. This research proposes a Reinforcement Learning Framework that uses reward shaping technique to reduce the delay in rewards so the agent can get faster and clearer feedback on the stock purchased. In this research, different deep learning architectures including Deep Q-Networks (DQN) ,Double DQN and Dueling Double DQN was used with our custom reinforcement learning Framework. the purpose of this reward shaping is to discourage excessive inactivity. The introduction of our reward shaping based Reinforcement Learning Framework markedly increased agent activity—by as much as a factor of four relative to unpenalized baselines — and contributed to faster convergence. For the DQN architecture, a penalty magnitude of 5 produced the highest observed profit, reaching 1,400 after 10,000 episodes compared to 350 in the baseline configuration. In the Double DQN architecture, a similar penalty setting yielded a final profit of 1,200, substantially exceeding the 250 achieved without penalization. The Dueling Double DQN architecture attained its highest profit without penalties (from 200 to 1,100 over training), yet still demonstrated measurable gains under appropriately tuned penalty parameters. Overall, these findings suggest that judiciously calibrated penalties can effectively mitigate reward latency, increase decision-making frequency, and improve profitability in reinforcement learning-based quantitative trading systems. However, an excessively high penalty resulted in degrading the model’s performance. This research benefits both academia and industry. For academic research, it contributes a systematic evaluation of reward shaping techniques in RL trading systems.

This research provide this systematic evaluation by giving insights into how delayed rewards can be mitigated through carefully tuned penalties. For the financial industry, the research provide a practical method to improve the efficiency and profitability of reinforcement learning based trading agent by encouraging model to make more responsible trading actions according to market changes.

Item Type:	Thesis (Masters)
Supervisors:	Name Email -, - UNSPECIFIED
Subjects:	H Social Sciences > HG Finance Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence H Social Sciences > HG Finance > Investment > Stock Exchange
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	04 Jul 2026 14:00
Last Modified:	04 Jul 2026 14:00
URI:	https://norma.ncirl.ie/id/eprint/9474

Actions (login required)

View Item