Hi, thanks for your comments. Not having done a detailed study of the input features, one should not expect the final result to be realistic. I see this more as a platform to start doing some machine learning applied to sports. In fact, if you train the model multiple times, not surprisingly you get different predictions for the same bracket. An important feature that in my opinion is missing but can be added somehow is how the teams performed during the regular season when they played against each other. I leave this as an exercise, the information should be available from basketball-reference.com .