I will try to answer this in my opinion: the guys with high rmse and very good private lb never post- processed the outliers in the public leaderboard . A good example is this ID: ID_PMSOXFT4FYDW change it by multiplying by 10 for it to be near 8000 and you are at 140 to 150 range , that was the easiest of them all and thats why you see many people at the 140... range. so the trick was to find the other outlier ids in public to get a good public score and some guys were just so so good at it , hats off to them , and since the private, they have not used the outlier ids in calculation of the metric you don't see any signiicant changes. But if @Zindi decides to include them, trust me you will see a very different top 20.

Models were generally good to find non-outlier preds, but the outlier preds is what was the whole point of this competition and it's sad that they have decided not to take them into consideration for the metric calculation

Thanks. Yeah it was brilliant seaching for such ids. I saw the discussions about this and tried by myself also. But post processing the data like that doesnt felt right to me. So i choosed the score Good in my traintestsplit around 130, but was 450 + on public lb and got 149 on private lb. Great learning btw.