[ENH] `skforecast` integration for time series hyperparameter tuning

[Omswastik-11]

Summary

This PR adds a full integration with skforecast, allowing Hyperactive to optimize hyperparameters of skforecast forecasting models using any of its optimization algorithms.

Implementation Details

SkforecastExperiment (skforecast_forecasting.py)

• Inherits from BaseExperiment.
• Uses skforecast.model_selection.backtesting_forecaster inside _evaluate() to perform time-series cross-validation for each parameter set.
• clones the forecaster and applies new parameters with set_params() before every evaluation.

SkforecastOptCV (skforecast_opt_cv.py)

• sklearn-style estimator (inherits from BaseEstimator).

• Works with ForecasterRecursive and other compatible skforecast forecasters.

• fit():

  • Builds a SkforecastExperiment with user settings (steps, initial_train_size, metric, etc.).
  • runs Hyperactive’s optimizer to search for the best hyperparameters.
  • refits the best forecaster on all available data.

• predict():

  • delegates to best_forecaster_.predict() for easy forecasting after optimization.

Configuration

• added skforecast as an optional dependency in pyproject.toml under the integrations extra.

Verification

• added skforecast_example.py showing a HillClimbing search with ForecasterRecursive + RandomForestRegressor.
• added unit tests to verify parameter handling, experiment execution, and integration flow.

Closes

Fixes #199

[Omswastik-11]

Hi @fkiraly !! Modified the docs string. As you suggested .

added get_test_params() verified all the tests .

tested the pre-commit on changed files

kindly verify this.

[Omswastik-11]

Hi @fkiraly !! commited the changes as you suggested

1. added skforecast to sktime-integration
2. https://github.com/SimonBlanke/Hyperactive/actions/runs/19598614085/job/56149302140?pr=208 Modified CI to avoid disk storage issue in CI build
3. cheked SkforecastExperiment against all tests

[JoaquinAmatRodrigo]

Hi, I’ve taken a look at the code related to skforecast, and it looks good. Thanks for the work, @Omswastik-11!

[Omswastik-11]

Hi @JoaquinAmatRodrigo @fkiraly !!! Thanks 👍. Can you trigger the workflow to see if it works perfectly ?

[JoaquinAmatRodrigo]

It needs to be triggered by one of the repository's maintainers.

[Omswastik-11]

Hi @JoaquinAmatRodrigo, @SimonBlanke, and @fkiraly — I’d appreciate your suggestions on how to proceed here.

The issue is that skforecast currently does not support Python 3.14, which affects the new integration. I am a bit unsure about two things:

1. Where exactly should skforecast be declared? – Only in pyproject.toml under optional extras? – Or in the _tags?
2. When should test:vm be enabled or disabled?

---

What I have done so far

1. Fixes for the test_examples workflow

Example test run: https://github.com/SimonBlanke/Hyperactive/actions/runs/19608701548/job/56518055022?pr=208

I checked the examples/ directory and noticed that many dependencies (e.g., torch, tensorflow) were not actually needed.

Changes:

pyproject.toml

• Added a new test_examples optional dependency group including:

  • pytest
  • integrations used in examples (sklearn, sktime, skforecast)
  • optuna

Makefile

• Added a new target for running example tests.

test.yml

• Updated the test-examples job to install only the required libraries .

---

2. Handling the Python 3.14 compatibility issue

skforecast fails to install on Python 3.14, so I added conditional version constraints:

skforecast; python_version < "3.14"

This has been added to all optional dependency groups that require skforecast.

Related CI runs: https://github.com/SimonBlanke/Hyperactive/actions/runs/19608701548/job/56518055064?pr=208 https://github.com/SimonBlanke/Hyperactive/actions/runs/19608701548/job/56518055077?pr=208

---

local testing

Thanks!

[JoaquinAmatRodrigo]

The integration works, nice!

The remaining issues relate to dependency isolation, you have to do it in two places:

• the get_test_params function, see above
• in the tags (I think you also need to add a few others)

Besides this, the "higher is better" property needs to be inferred from the metric and set as a tag in __init__.

@JoaquinAmatRodrigo, is there a programmatic way to do this?

We do not have a programmatic strategy for this. So far, all the regression metrics that Skforecast allows to be passed as a string are intended to be minimised.

"mean_squared_error": mean_squared_error,
"mean_absolute_error": mean_absolute_error,
"mean_absolute_percentage_error": mean_absolute_percentage_error,
"mean_squared_log_error": mean_squared_log_error,
"mean_absolute_scaled_error": mean_absolute_scaled_error,
"root_mean_squared_scaled_error": root_mean_squared_scaled_error,
"median_absolute_error": median_absolute_error,
"symmetric_mean_absolute_percentage_error": symmetric_mean_absolute_percentage_error

If the user passes a custom function as a metric, they need to indicate whether it is a maximisation or minimisation.

[Omswastik-11]

Hi @JoaquinAmatRodrigo !! Thanks for the clarification . It means it would be better if we pass the metric in constructor like metric = 'mse' , scoring = 'lower' as default ? where in metric user can pass their own metric functions and change scoring to 'lower ' or 'higher' if they want to minimize or maximize that metric

[JoaquinAmatRodrigo]

Looks like a great solution. You might want to double-check the keywords, usually libraries use 'maximize' or 'minimize'. Not sure what is the ones used in this library.

[Omswastik-11]

Thanks @JoaquinAmatRodrigo !! In my recent changes I have added a boolean parameter named 'higher_is_better' . https://github.com/SimonBlanke/Hyperactive/pull/208/commits/ca013a4c83ce583a8fa0766dd6416cb7e2953e5c

I checked sktime_forecasting.py they used scoring as parameter for both metrics and higher_or_lower part but as we have one separate parameter for metric so i just added a boolean param . https://github.com/SimonBlanke/Hyperactive/blob/main/src/hyperactive/experiment/integrations/sktime_forecasting.py#L171

@fkiraly and @SimonBlanke any suggestions ?

[Omswastik-11]

Hi @fkiraly !!! Have a look at this . I explained why I did changes https://github.com/SimonBlanke/Hyperactive/pull/208#issuecomment-3585153351

[fkiraly]

Have a look at this . I explained why I did changes

What you say in the comment you mention is not consistent with the actual changes in the pyproject.

If you are using AI, please watch what it is doing.

[Omswastik-11]

Hi @fkiraly Sorry for the issue It seems I forgot to remove the initial changes . The followings

  - name: Free Disk Space (Ubuntu)
    if: runner.os == 'Linux'
    run: |
      sudo rm -rf /usr/share/dotnet
      sudo rm -rf /usr/local/lib/android
      sudo rm -rf /opt/ghc
      sudo rm -rf /opt/hostedtoolcache/CodeQL
      sudo docker image prune --all --force

rest is the same as I written in comments .

as you mentioned about the changes made in pyproject.toml This was the commit here I made changes . https://github.com/SimonBlanke/Hyperactive/pull/208/commits/3d92878e05a64569ffc33eb960fded1a3726f18e Can you pin where I did the mistakes ?

all tests are passing currently btw https://github.com/SimonBlanke/Hyperactive/pull/208/checks?sha=ca013a4c83ce583a8fa0766dd6416cb7e2953e5c

[Omswastik-11]

Hi @SimonBlanke !! Can you re run the workflow ? I have reverted the changes and added this which last time solved the storage issue .

refer this https://github.com/hyperactive-project/Hyperactive/actions/runs/19744054270/job/56593197910?pr=208

  - name: Free Disk Space (Ubuntu)
    if: runner.os == 'Linux'
    run: |
      sudo rm -rf /usr/share/dotnet
      sudo rm -rf /usr/local/lib/android
      sudo rm -rf /opt/ghc
      sudo rm -rf /opt/hostedtoolcache/CodeQL
      sudo docker image prune --all --force