But why is the question, I run about 400k models every 30 minutes at work. I have tried all other approaches, but xgboost model per timeseries has never been beaten before. Even if I wasn't targetting accuracy, Xgboost loading a pretrained model and predicting with it and still faster.