Towards optimized AI strategies for cryptocurrency trading

This thesis investigates the application of advanced AI techniques, specifically Double Deep Q-Networks, for developing optimized cryptocurrency trading strategies. We explore the efficacy of two distinct neural network architectures as Q-function approximators: a novel time-series model, TimesNet, and a conventional 1D CNN. The study employs historical price data (OHLCV) for Bitcoin and Ethereum, evaluating performance with both a basic feature set and an extended set incorporating technical indicators and blockchain-specific metrics. Our experiments demonstrate that DDQN agents, particularly those leveraging the TimesNet architecture, can learn complex trading behaviors and achieve significant outperformance compared to traditional Buy & Hold benchmarks and CNN-based agents across diverse market conditions. The research highlights the potential of sophisticated temporal modeling and reinforcement learning in navigating the volatile cryptocurrency landscape, while also underscoring the nuanced impact of feature engineering on strategy performance across different market regimes and assets.