Dynamic Multi-Pair Trading Strategy in Cryptocurrency Markets with Deep Reinforcement Learning

Battle Count: 81

Highly relevant to quantitative trading. Introduces a practical hybrid architecture combining classical statistical arbitrage (cointegration, Hurst filtering) with modern DRL execution policies. Addresses critical challenges in crypto pair trading: divergence risk, regime shifts, and policy instability. Demonstrates safe RL via deterministic shielding applicable to real-world algorithmic trading. Provides extensive sensitivity analysis, bootstrap significance testing, and ablation studies. The 'Fixed Risk, Adaptive Mean' model and hierarchical pair selection are directly implementable in production trading systems.

High complexity due to multi-component architecture: (1) Dynamic monthly pair selection with Engle-Granger cointegration, Hurst exponent, and beta filtering across 4,950 combinations; (2) Custom 'Fixed Risk, Adaptive Mean' execution engine with state-dependent Z-score, frozen volatility anchors, and live mean adaptation; (3) Three-tier stop-loss architecture (deterministic SL, SL Lock, Time Decay SL); (4) PPO-LSTM training with Gymnasium environment, multiple reward schemes, and observation space variants; (5) Deterministic risk management overlay for deployment; (6) 10x leverage mechanics with liquidation engine. Requires expertise in both quantitative finance and deep reinforcement learning.

Full source code available on public GitHub repository. Custom simulation engine built from scratch with Python 3.12, Poetry dependency management, Docker containerization, Hydra-Pydantic configuration, Gymnasium API compliance, Stable-Baselines3, and Weights & Biases integration. Fixed random seeds (42), detailed hyperparameter tables, and comprehensive sensitivity analyses provided. Data sourced from public Binance Data Vision archive.