Co-authored-by: Joao Barros joao.f.barros@tecnico.ulisboa.pt

Reference Issues/PRs

Implements #31344.

What does this implement/fix? Explain your changes.

This implements a new class, MultiHorizonTimeSeriesSplit, to enhance time series cross-validation by supporting multiple prediction horizons per split.
The existing TimeSeriesSplit class is limited to generating a single, contiguous test set per split, which is insufficient for scenarios requiring forecasts over multiple future steps (e.g., predicting 1, 3, and 5 days ahead).
This new class addresses that limitation by enabling users to evaluate time series models across various forecast horizons in a single, streamlined workflow.

Here’s a breakdown of the modifications:

• Added MultiHorizonTimeSeriesSplit as a subclass of TimeSeriesSplit.
• Introduced a horizons parameter (a list of positive integers) to specify the prediction steps for the test sets.
• Modified the split method to generate test indices based on the provided horizons.
• Added checks to ensure horizons is a non-empty list of positive integers and that the input data length supports the specified horizons and number of splits.
• Created tests to verify the correctness of splits across different configurations (e.g., varying horizons, gaps, and test sizes).

Any other comments?