Hi Alex! You’re right – the difference in the mean absolute error between the train and test sets can be measured in “hundreds of times”. Notes taken, thank you for your feedback!

Thanks, metodinikolov, for your feedback! 🙂
The removal of some of the temperature features is something that I completely overlooked when I started the modelling part, and therefore unwittingly I’ve quadrupled the temperature weight in the models. Definitely the first thing to rework.
Also agree about the need of outliers analysis, and also thought about the creation of a binary (or even a numeric) feature flagging whether (or to what extent, if numeric) the record is an outlier. So it can hint the model in what conditions it should predict extreme figures.
Sadly, some good ideas came too late 🙂
Cheers!

Hi alex-efremov, many thanks for your feedback! It’s really encouraging to hear that I’m on the right track! 🙂

Thank you very much for your feedback and the good and detailed suggestions – I agree with everything you say. I will have a look at the project in the next hours and will see what of these (or what else, if any) I’ll be able to include. Thanks a lot for your appreciation! 🙂