I present an algorithm that can combine any number of return forecasting models into one without any retraining, hyperparameter tuning, or rebalancing, while being mathematically optimal and guaranteed to perform nearly as well as the best model.