To develop a high-performance model for predicting the spontaneous combustion tendency of coal, on the basis of multiple gas indicators and industrial analysis parameters, four machine learning approaches (random forest, neural network, support vector machine, and Stacking ensemble) were used to predict spontaneous combustion temperature and natural ignition period, thereby evaluating coal spontaneous combustion risk. The findings indicate that the Stacking ensemble model exhibits superior generalization capability. Furthermore, feature importance analysis reveals that volatile matter and ethylene are the most influential predictors for natural ignition period and spontaneous combustion temperature, respectively. Model performance evaluation suggests that increasing data volume significantly enhances the predictive generalization of all four methods for spontaneous combustion temperature. However, expanding data alone yields only marginal improvement in predicting the natural ignition period. Enhancing feature representation is therefore necessary to further improve model performance.