Tabular model always predicts 0 - M5 Forecasting accuracy competition

Can someone help me to understand why this happens. I have used features referring


Are you still working on the competition ?
Could that be due to the fact that some items are mostly flat series ?

Hi @louis
Yes, Still I do.
It was because trying argmax. My bad … Then i just submitted the results considering as regression problem. It worked.

I am trying to use the tabular learner to get all the nice features of fastai out of the box, but I face a couple of problems. Would be nice to know if you have faced similar issues.

I mostly struggle to wrap my thoughts about the validation strategy using rolling features: i am using rolling features (last 7 mean sales for example). So when i comes to the validation set I cannot include days which are more than 1 day ahead of the last day of my training set. One way i to write my own loop to compute the roling feature on top of the predicted value, this for every level of days ahead in my validatino set.

It’s seems quite cumbersome, but I cannot see anyway to mitigate overfitting than doing this…