This paper presents a novel deep reinforcement learning framework for market making in cryptocurrency markets, termed DRLMM. It addresses the stochastic inventory control challenges faced by market makers by employing two advanced policy gradient-based algorithms that interact with a simulated environment. The environment is constructed using limit order book data and order flow arrival statistics, with a forward-feed neural network serving as the function approximator. The study compares two different reward functions to optimize agent performance, evaluating outcomes through daily and average trade returns.

The research demonstrates the effectiveness of deep reinforcement learning in enhancing market making strategies, particularly in the volatile and data-rich context of cryptocurrencies. By leveraging policy gradient methods and neural networks, the framework enables agents to learn adaptive trading policies that manage inventory risks and capitalize on market opportunities. The findings highlight the potential of DRLMM to improve profitability and efficiency in high-frequency trading, contributing to the growing body of literature on AI applications in quantitative finance.