In order to accurately predict the lateral deformation of a foundation pit, a multivariable stacking prediction model based on extreme gradient boosting (XGBoost), long short-term memory (LSTM) and linear regression (LR) was proposed. By using the XGBoost’s advantage of ensemble learning and the accuracy of the two-layer LSTM algorithm in the traditional foundation pit deformation prediction, the prediction accuracy and the generalization ability of the model were improved. In the data pre-processing stage, the K-nearest neighbors (KNN) interpolation algorithm was introduced to increase the total amount of data that can be effectively utilized, and the processing method of time information in the deep learning model Informer was used to deal with the problem of ignoring the different time intervals of time series data by supervised learning in the traditional algorithm. Taking a foundation pit under construction in Hangzhou as a practical engineering case, 616 missing data were interpolated, the time information was converted into three columns of time point feature information, and the proposed model was used for foundation pit deformation analysis. Existing measured data verified that the training accuracy and the generalization ability of the proposed model were greatly improved compared with both the two-layer LSTM model and the XGBoost model when predicting the maximum slope displacement of the foundation pit and the depth of the displacement point. The XGBoost model, which used time-point features, was more suitable for predicting time-sensitive indicators than the LSTM model.