Forecasting fluctuations in cryptocurrency trading volume using a hybrid LSTM–DQN reinforcement learning

Forecasting fluctuations in cryptocurrency trading volume is essential for developing effective trading strategies and risk management tools in volatile financial markets. This paper proposes a hybrid deep learning framework that integrates Long Short-Term Memory (LSTM) networks with Deep Q-Network (DQN) reinforcement learning to predict the directional movement of trading volume in cryptocurrencies. The LSTM component captures temporal dependencies in historical market data, while the DQN agent learns optimal actions based on state transitions in a custom-designed environment. The agent receives positive or negative rewards depending on the accuracy of its directional predictions (increase, decrease, or unchanged volume). The proposed model is trained and evaluated on historical data, with a focus on Bitcoin, and demonstrates progressive learning through reinforcement. The cumulative reward increased from − 93 to over 1800 within 200 training episodes, indicating effective policy optimization. The model achieved a Mean Squared Error of 0.4366 and a Root-Mean-Squared Error of 0.3552, reflecting high predictive accuracy. Furthermore, for directional volume prediction, the model achieved an accuracy of 87%, precision of 86%, recall of 88%, and F1-score of 87%, demonstrating its reliability in capturing upward and downward volume movements. The final system is deployed in a real-time web application using a Python-based dashboard, enabling continuous monitoring and visualization of predicted volume movements. The experimental results validate the effectiveness of combining temporal sequence modeling with reinforcement learning for volume fluctuation forecasting and demonstrate the model’s potential applicability in real-world financial decision support systems.