Because the accumulated advertisement logs in the Internet have the problems of sparse data, a large number of features and extremely unbalanced distribution of positive and negative samples, manual feature extraction is time-consuming and laborious, and it is difficult for a single prediction model to obtain better prediction performance. In response to these problems, this paper completes a click through rate prediction model based on GBDT model and stacking. This model uses GBDT model to automatically extract and construct features, and predicts and classifies click-through rate by Stacking model, which effectively improves the performance of the single prediction model. Experiments on real advertising data sets show that the GBDT-Stacking ensemble method increases the AUC value by at least 4% compared to the comparison model.