Stale blocks are not avoidable in blockchain, such as the Bitcoin network, when proof-of-work is used as the consensus protocol. However, as the economic loss to the miners and the security risk to the network cannot be ignored, research is needed to identify and analyse stale blocks. By analysing the factors influencing the generation of stale blocks, the authors propose a new machine learning model based on XGBoost. They propose a new data collection method for bitcoin nodes to obtain real data for training prediction model. Then, based on the model, they generate optimal mining strategies and analyse the economic benefits. The experimental data and application cases show that the real-time data detection and machine learning model that they propose can accurately identify and predict the generation of stale blocks and generate an economically optimal mining strategy in the Bitcoin network with the presence of stale blocks.