Hybrid CNN-LSTM Model for Bitcoin Limit Order Book Prediction

Abstract This study improves the accuracy of price movement prediction in financial markets by developing a hybrid model that combines convolutional neural networks (CNNs) and a (LSTM) network. Focusing on limit order book (LOB) data, which reflect real-time market conditions, a highly accurate prediction model was constructed by effectively extracting spatial and temporal features from the high- long short-term memory dimensional and complex LOB data. Using XBT/USD data from the BitMEX exchange and building upon previous research, we also implemented a labeling method that excludes past data to improve the model’s long-term prediction accuracy and examine its applicability to different markets. Experimental results demonstrate that the proposed hybrid CNN-LSTM model obtain an accuracy of 61%, an F1-score of 69% (for the upward trend), and an AUC-ROC score 0.6180, thereby demonstrating superior performance (particularly when predicting the upward trends). The transition to binary classification and the data balancing method, which preserves the sequential nature of time-series data, contributed significantly to improved prediction accuracy. In addition, the evaluation of a simple trading strategy based on the proposed model yielded Sharpe ratio and profit factor values of 0.2158 and 1.5347, respectively, which indicates limited but positive risk-adjusted returns. These results suggest that the proposed hybrid model can predict the direction of price movement with a certain level of accuracy based on LOB data and holds potential for practical trading strategies.