To address the issue that there are many characteristics influencing the stray current of a subway track, the conventional feature selection method affects the prediction accuracy of the model, and the interpretability of the model results is poor, this paper proposes a stray current prediction model based on optimal feature improved eXtreme Gradient Boosting (XGBoost). Using the flexibility and the strong searchability of the genetic algorithm, we found the first M features that minimizing the mean square error (MSE) of the objective function generation by generation in the set containing the original V features. Simultaneously, the stray current prediction model under the optimal feature selection method (OFS-XGBoost) is established. To address the issue that the prediction results of the OFS-XGBoost are good, however, the machine learning black-box model has an insufficient explanatory ability for the prediction results, an attribution analysis framework based on SHAP theory is proposed to show the influence of feature set on the prediction results of the model in an understandable way based on the marginal contribution of stray current feature samples to improve the inference accuracy. The results show that the prediction error of the proposed model is only 1.684%, which is lower than the prediction models such as random forest and back propagation (BP) neural network under the same optimization strategy. The attribution analysis method based on SHAP value explains the impact of input characteristics on stray current prediction results from a global and individual perspective, helping intelligent subway health management based on improving model interpretability.