# Deep Reinforcement Learning in Quantitative Algorithmic Trading: A Review

**Authors:** Tidor-Vlad Pricope  
**arXiv:** [2106.00123](https://arxiv.org/abs/2106.00123)  
**Date:** 2021-05-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.00123v1/

## Key points

- Deep reinforcement learning (DRL) has been explored for automated low-frequency quantitative stock trading
- Prototype systems developed using DRL have shown success in specific markets
- Existing studies highlight limitations and challenges faced by DRL agents, including lack of profitability and real-time testing on online trading platforms
- DRL has potential in algorithmic trading, particularly in minute to daily timeframes
- The question of whether DRL can outperform professional human traders is explored, with some studies suggesting it can compete in terms of risk-adjusted return rates
- Further research is needed to fully understand the extent of DRL's power in this environment
- The paper provides an overview of reinforcement learning (RL) concepts and discusses a specific experiment involving market data and a supervised learning system called RCNN for sentiment analysis
- Multiple approaches proposed in the literature for creating successful trading agents using RL and deep neural networks are presented
- Further research is necessary to fully exploit the potential of DRL in algorithmic trading.

## Summary

This paper explores the applicability of deep reinforcement learning (DRL) in automated low-frequency quantitative stock trading. Prototype systems developed using DRL have demonstrated great success in specific markets. The authors review the progress made so far in this subdomain of AI in finance and discuss the limitations and challenges faced by existing studies. They highlight that while DRL agents have shown significant improvements in performance compared to established baseline strategies, they have not achieved a decent level of profitability. The lack of real-time testing on online trading platforms and meaningful comparisons between different types of DRL or human traders is also noted. The paper emphasizes that DRL has enormous potential in algorithmic trading, particularly in minute to daily timeframes. The question of whether DRL can lead to a super-human agent capable of outperforming professional human traders is also explored. While some studies suggest that DRL systems can compete with professional traders in terms of risk-adjusted return rates, further research is needed to fully understand the extent of DRL's power in this environment. In the background section, the paper provides an overview of reinforcement learning (RL) concepts for readers unfamiliar with the topic. It discusses a specific experiment involving market data on a daily timeframe for one stock and a supervised learning system for sentiment analysis called RCNN. The RCNN model was trained using news headlines from various companies and achieved satisfactory accuracy on testing. Overall, this paper aims to present multiple approaches proposed in the literature for creating successful trading agents using RL and deep neural networks. It concludes that while DRL shows promise, further research is necessary to fully exploit its potential in algorithmic trading.