[FEAT] Evaluation to utils

[elephaint]

PR:

• Softly deprecates the HierarchicalEvaluation class in favor of a function evaluate which is a small wrapper around utilsforecast's evaluate which can work with utilsforecasts's evaluation functions. This further unifies the HF API across our packages.
• The deprecation is soft in the sense that for now, we keep the HierarchicalEvaluation class and loss functions but remove it from the examples, the docs and I added a deprecation notice to each.
• Updates all examples to show the new evaluation method.

Note:
utilsforecast scaled_crps is different from what is currently in hierarchicalforecast, as the former normalizes on a per-series basis, whereas the GluonTS implementation (which is followed in hierarchicalforecast currently) normalizes the CRPS using the norm based on all timeseries. Therefore, when recalculating our examples using the scaled crps of utilsforecast, we get different error metrics (and different conclusions!)

Todo / to solve:

• Using utilsforecast, we evaluate slightly differently than we currently do with HF, in particular when using benchmark models in the evaluation metric. We previously would compute a relative score as follows: (overall_scalar_loss / overal_scalar_loss_benchmark), whereas using utilsforecast's loss functions, we compute the relative loss per timeseries, and compute the mean across these relative losses. Added a benchmark possibility in evaluate to obtain the correct behavior

The following loss functions should be added to utilsforecast to achieve parity with current HF functionality:

• log_score: requires a set of inputs that's difficult to achieve in utilsforecast without some fiddling. Need to think about what this score adds beyond scaled_crps and energy_score Don't think it adds much tbh.
• energy_score Deferred for now, the loss function can still be used via the old API, and it's removed from the examples, keeping only scaled_crps.
• msse Already in utilsforecast
• rel_mse not including this one in utilsforecast; users should use mse in conjunction with the benchmark-attribute in evaluate, to get the proper relative mse.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB