now that the competition over, I’d like to know if someone could share their approach to build a learning model with this data.
With time series, the data to predict is generally in the future compared to the training set. This dataset (with test periods in the past compared to some training data) was definitively a challenge.
My approach was to run a model for each ForecastId.
This was kind of slow but at least I was sure that my prediction was based on past data (well I did use future weather data as it was allowed).