Congrats Nelly, Thank you for informing :)

I would love to see this too

Hi Roman, As for me, I tried interpolating and extrapolating all-time series as this was the only logical thing in this data but it did not give expected results, so I got my best score without imputing missing values. And about feature engineering, my best single model score on private LB is 33.11 with 180 features. And it is ensembled with a model on 3115 features.

@Roman_Lents, for most of it we didn't try to do anything with the NaN values, let the classifier handle them. Manually imputing them increased the RMSE for our

Thank you for telling youngtard, this means most of us have got our scores with ml models only. So features and validation were the key. Cool.

Bernd that's the best cv I have heard so far. I ask assuming you used KFold. How many folds did you use in that case? By the way great work.

Thank you!!. Normaly i use 5 fold cv. But in this case i used 20 fold (computation was fast because of the smal data set).

Thanks, that explains the reason for such a CV :D