More thoughts on neutralization/residualization
The idea of neutralization is very powerful for MMC because you can see exactly what makes your prediction unique and you can test if that unique quality is good very directly. But we hope that some of the more advanced users will be able to take this tool and find other interesting ways to train models even for focus on the main tournament as well. For instance, you could try neutralizing the targets to your predictions, and then train a whole new model which will be independent of your first model! You can tweak the proportion parameter so that you’re only 50% neutral to certain factors. So many options out there that we hope you all can explore. We can’t wait to see the results of your experimentation.