[snath-xoc]

Implementing suggestions:

• for the splitter: avoid having a final validation set with an arbitrary length. start the splits from the end of the historical data, going back with regular steps (and stop when train set reaches its minimum size)
• try post-hoc recalibration in the quantile prediction notebook
• plot a pareto front: coverage vs interval width
• randomized search:
• currently we try either temperature alone or all features. also consider temparature + wind speed
• currently we use the same hyperparameters for all HGB models in the quantile
      prediction. We can search for the best model for each quantile separately, then combine them.
• store an optuna db with a large number of iterations in the repo. in the notebook, resume the search to add a few iterations.
• try the quantile prediction with randomforest: FEA: DecisionTreeRegressor: add pinball loss support / quantile regression scikit-learn/scikit-learn#32903
• maybe try combination of features eg interactions between temp and humidity
• increase max_iter on the HGB
• store a skrub report in a static file in the jupyter book and show a link to it in the notebook. students who are in jupyterlite can see the static precomputed report. others can generate and open reports on their machine.
• we could also train & compare "fallback" models that use a more restricted set of features in case some information e.g. the weather forecast fails to be acquired in time once in "production".

@jerome-probabl @ogrisel