IPRIT 0.3
Developed innovative trading algorithms using machine learning for crypto assets, achieving notable accuracy in time-series predictions.
Challenge
Inspired by UniV3 research, I envisioned active liquidity provision on UniV3 as a significant enhancement to strategies using predictive models for price movement direction. This led to the development of a unique framework combining ML and RL, aimed at creating an unprecedented system with unconventional technologies.
Results
A state-of-the-art time-series prediction model, specifically tailored for market data, was developed. It achieved impressive accuracy, surpassing baseline solutions with 64% and 57% accuracy for 5 and 10-minute price predictions. An automatic trading system was then initiated. However, baseline solutions showed inconsistent results with high variance. Further research, based on existing frameworks, was required, but development stalled due to financial constraints and is currently paused pending improved financial circumstances.
Bulletpoints
- Pioneered a transformer-like encoder model for time-series prediction in PyTorch, achieving 64% and 58% accuracy for 5 and 10-minute forward price predictions, respectively, surpassing public benchmarks.
- Engineered a comprehensive quantitative trading framework leveraging tick-level exchange data, facilitating the development and deployment of robust trading algorithms via reinforcement learning.
- Implemented and managed a distributed, model-based reinforcement learning system, ensuring scalable and stable training across 16 nodes.
- Applied reinforcement learning models to enhance trading strategies, achieving a 20% improvement in trade execution efficiency and risk management compared to baseline.
- Achieved a 16-fold reduction in model training duration via strategic optimizations in cloud deployment on AWS EC2.