Development of an Automated Trading and Risk Management System Using AI
Applied reinforcement learning technologies and distributed computing to optimize trading strategies
- Client
- Robinhood
- Year
- 2023
- Services
- Algorithmic modeling and development
- Platform
- Python, Reinforcement Learning
About project
A large financial company approached us with the goal of automating trading and risk management on the stock exchange using artificial intelligence. The main task was to develop a system capable of independently making trading decisions and effectively managing risks by analyzing large volumes of historical data.
Solution
We developed a stock exchange emulator that replayed historical data, allowing the agent to learn in a controlled environment. During the research, more than 100 different strategies and approaches were tested. As agents, various reinforcement learning methods were tested, including: A2C (Advantage Actor-Critic) A3C (Asynchronous Advantage Actor-Critic) DQN (Deep Q-Network) Rainbow — a combination of DQN and numerous improvements DDQN (Double Deep Q-Network) -and others. The system was trained on a distributed computing cluster. Copies of agents and exchange emulators were distributed across dozens of servers. Each such setup collected millions of hours of simulated trading and transactions, after which the data were sent to a central server for training models on GPUs. Note: This technology was later used to improve models in other areas, including victories in esports competitions and the development of reinforcement learning methods with human feedback (RLHF), which are the foundation of modern chatbots.
Result
A fundamental project base was developed for further phased implementation into the platform's algorithms and trading bot products. The system demonstrated potential for increasing the efficiency of trading operations and improving risk management, providing the company with tools for more informed decisions on the stock exchange.
