Coastal structures are often designed to a maximum allowable wave overtopping discharge, hence accurate prediction of the amount of wave overtopping is an important issue. Both empirical formulae and neural networks are among the commonly used prediction tools. In this work, a new model for the prediction of mean wave overtopping discharge is presented using the innovative machine learning technique XGBoost. The selection of features to train the model on is carefully substantiated, including the redefinition of existing features to obtain a better model performance. Confidence intervals are derived by tuning hyperparameters and applying bootstrap resampling. The quality of the model is tested against four new physical model data sets, and a thorough quantitative comparison with existing machine learning methods and empirical overtopping formulae is presented. The XGBoost model generally outperforms other methods for the test data sets with normally incident waves. All data-driven methods show less accuracy on oblique wave data, presumably because these conditions are underrepresented in the training data. The performance of the XGBoost model is significantly improved by adding a randomly selected part of the new oblique wave cases to the training data. In the end, this new model is shown to reduce errors on all data used in this work with a factor of up to 5 compared to existing overtopping prediction methods.