ChronosRL: embeddings-based reinforcement learning agent for financial trading

ChronosRL is a reinforcement learning (RL) framework for cryptocurrency trading that integrates transformer-generated embed-dings from the Chronos model. By leveraging pre-trained embeddings that capture complex temporal dependencies and nonlinear patterns, ChronosRL addresses the limitations of handcrafted feature-based strategies. Experiments across six cryptocurrency pairs over fve months show that ChronosRL achieves a mean proft of 102% with a 79% standard deviation, outperforming baselines such as Buy and Hold and Moving Average Bollinger Bands. Operating in a zero-shot setting without fine-tuning, ChronosRL demonstrates strong generalization capabilities. These findings highlight the potential of transformer-based embeddings to enhance RL performance in volatile financial environments, bridging advances in sequential representation learning and algorithmic trading.