Changelog
Source:NEWS.md
modeltime.ensemble 1.0.0
CRAN release: 2021-10-19
NEW Nested Modeltime Ensembles
In modeltime
1.0.0, we introduced Nested Forecasting as a way to forecast many time series iteratively. In modeltime.ensemble
1.0.0, we introduce nested ensembles that can improve forecasting performance and be applied to many time series iteratively. We have added:
-
ensemble_nested_average()
: Apply average ensembles iteratively -
ensemble_nested_weighted()
: Apply weighted ensembles iteratively
modeltime.ensemble 0.4.1
CRAN release: 2021-05-31
- Improvements for parallel processing during refitting (available in
modeltime
0.6.0). - Requires
modeltime
0.6.0 andparsnip
0.1.6 to align with xgboost upgrades.
modeltime.ensemble 0.4.0
CRAN release: 2021-04-05
Recursive Ensembles
-
recursive()
- Therecursive()
function is extended to recursive ensembles for both single time series and multiple time series models (panel data). -
“Forecasting with Recursive Ensembles” - A new forecasting vignette for using
recurive()
with ensembles.
Fixes
-
modeltime_forecast()
now returnsNA
when missing values are present in the sub-model predictions.
modeltime.ensemble 0.3.0
CRAN release: 2020-11-06
Panel Data
- Improvements made to
ensemble_average()
,ensemble_weighted()
andensemble_model_spec()
to support Panel Data (i.e. when data sets with multiple time series groups that have possibly overlapping time stamps).
Changes
-
modeltime.ensemble
now depends onmodeltime.resample
for themodeltime_fit_resamples()
functionality. -
modeltime_fit_resamples()
moved to a new packagemodeltime.resample
. -
ensemble_weighted()
: Now removes models that have no weight (e.g. loading = 0). This speeds up refitting.
modeltime.ensemble 0.2.0
CRAN release: 2020-10-09
Stacked Ensembles (Breaking Changes)
The process for creating stacked ensembles is split into 2 steps:
- Step 1: Use
modeltime_fit_resamples()
to generate resampled predictions - Step 2: Use
ensemble_model_spec()
to apply stacking using amodel_spec
Note - modeltime_refit(stacked_ensemble)
is still one step, which is the best way to handle refitting since multiple stacked models may have different submodel compositions. An additional argument, resamples
can be provided to train stacked ensembles made with ensemble_model_spec()
.