To solve the problems of too long compressive strength test cycle and too much material of cementitious sand and gravel （CSG） consumed， the model of extreme gradient boosting tree combined with long short-term memory network （XGBoost-LSTM） was used to predict the compressive strength of CSG. Firstly， the two input variables of “cement content” and “sand ratio” with strong correlation were selected into the XGBoost model for prediction， and the results were substituted into the LSTM model together with the original features. Another 94 sets of compressive strength data were used for training and validation. The results show that compared with the basic models XGBoost and LSTM， the coefficient of determination of the XGBoost-LSTM combined model is increased by 5.6% and 3.5% respectively. It has shown to be feasible to construct new features by the XGBoost model， and the XGBoost-LSTM combined model can accurately predict the compressive strength of CSG.