You will be scored using the MAE calculated on a holdout dataset. Make sure to not overfit to the public leaderboard which is a *small random* subset of this holdout set.