[ENH] extend sktime ForecastingOptCV with broadcasting options and returned parameters

[anirudhsengar]

Summary

Extend ForecastingOptCV to expose sktime’s broadcasting controls plus richer post-fit diagnostics, aligning it more closely with ForecastingGridSearchCV while keeping Hyperactive optimizers in play. Also document the new knobs and assert the behavior via integration tests.

Changes

• Broadcasting flags: Added tune_by_instance/tune_by_variable init params. When sktime is available we reuse its delegated tagging (_set_delegated_tags + _extend_to_all_scitypes) so panel/multivariate broadcasting follows the native implementation rather than a bespoke copy.
• Returned attributes: fit now records scorer_, n_splits_, best_index_, best_score_, best_params_, best_forecaster_, refit_time_, and a cv_results_ table (the backtesting results for the winning configuration). We time refits with time.perf_counter and keep the optimizer’s best index when available.
• Docs & tests: README.md explains the new parameters/diagnostics. test_sktime_estimators.py gained regression tests to ensure the attributes populate and the tune-by flags adjust estimator tags.

Related Issues

Fixes #198

[fkiraly]

Nice!

Though, I think there is already automatic broadcasting in sktime - this should be activated, instead of reimplementing it.

Have a look at BaseGridSearch in sktime, and the __init__ code which handles tune_by_variable and tune_by_instance (the latter via _extend_to_all_scitypes).

[fkiraly]

Great!

Only very small comments above.

[anirudhsengar]

Done! Thank you @fkiraly for the feedback.

[anirudhsengar]

If approved, please merge sktime/sktime#9092 as well.

[fkiraly]

Great!

Small request - please move the example from the README to the docstring, and also update the docstring in sktime/sktime#9092 so it is identical.

[anirudhsengar]

Done! @fkiraly. Even updated sktime/sktime#9092.

[SimonBlanke]

merge the current main into your branch. This should help with the CI fails.