LoginSign Up
← Back to all sparks
M

modeltime.ensemble

ANALYTICS
Velocity0.0

Time Series Ensemble Forecasting

github.com/business-science/modeltime.ensemble ↗

modeltime.ensemble wakes after four years, and the work is all tune 2.0 compatibility.

time series forecastingensemblestidymodelscompatibility maintenancer package
◆Current state
modeltime.ensemble builds average, weighted and stacked ensembles over modeltime forecast models. After a four-year gap it shipped twice in a fortnight during August and September 2025, both releases devoted to tracking breaking changes in tidymodels' tune package — new resampling column conventions, key uniqueness across resamples, recipe preparation. The tidyverse dependency was dropped in the same pass.
◆Where it's heading
This is a package whose forecasting capability was settled by 2021 — recursive ensembles, per-series calibration — and whose recent life is dictated entirely by upstream tidymodels churn. New contributors did that compatibility work, including one from the tidymodels side. It now requires tune 2.0.0 and modeltime.resample 0.3.0, pinning it to the current tidymodels generation rather than straddling versions.
◆Prediction
Expect the next release to follow the next tune or modeltime.resample breaking change rather than to introduce new ensembling methods.

◆Recent moves

  1. 11mo ago

    Realigned for tune 2.0.0 resampling changes

    The second of two compatibility releases in a fortnight, adapting to tune 2.0.0's resampling column names and key handling. It pins the package to tune 2.0.0 and modeltime.resample 0.3.0 rather than supporting both generations.

    View source ↗
  2. 11mo ago

    Drops the tidyverse dependency ahead of tune 2.0

    Dependency and CI cleanup from three first-time contributors, plus preparatory work for the next tune release. The jump from 0.4.2 marks the end of a four-year gap rather than a feature milestone.

    View source ↗
  3. 5y ago

    Per-series calibration IDs and parallel refitting

    Calibration gains an id feature so accuracy and confidence intervals can be produced per time series — the natural follow-on from the panel-data ensembles opened up in 0.4.0.

    View source ↗
  4. 5y ago

    Recursive ensembles for single and panel series

    ⚡ SPARK

    recursive() extends from individual models to ensembles, including panel data, with its own forecasting vignette. It is the last substantive capability change before the package went quiet for four years.

    View source ↗