Skip to contents

Creates an Ensemble Model using Weighted Averaging in the Modeltime Nested Forecasting Workflow.

Usage

ensemble_nested_weighted(
  object,
  loadings,
  scale_loadings = TRUE,
  metric = "rmse",
  keep_submodels = TRUE,
  model_ids = NULL,
  control = control_nested_fit()
)

Arguments

object

A nested modeltime object (inherits class nested_mdl_time)

loadings

A vector of weights corresponding to the loadings

scale_loadings

If TRUE, divides by the sum of the loadings to proportionally weight the submodels.

metric

The accuracy metric to rank models by the test accuracy table. Loadings are then applied in the order from best to worst models. Default: "rmse".

keep_submodels

Whether or not to keep the submodels in the nested modeltime table results

model_ids

A vector of id's (.model_id) identifying which submodels to use in the ensemble.

control

Controls various aspects of the ensembling process. See control_nested_fit().

Details

If we start with a nested modeltime table, we can add ensembles.

nested_modeltime_tbl

# Nested Modeltime Table
Trained on: .splits | Model Errors: [0]
# A tibble: 2 x 5
  id    .actual_data       .future_data      .splits         .modeltime_tables
  <fct> <list>             <list>            <list>          <list>
1 1_1   <tibble [104 x 2]> <tibble [52 x 2]> <split [52|52]> <mdl_time_tbl [2 x 5]>
2 1_3   <tibble [104 x 2]> <tibble [52 x 2]> <split [52|52]> <mdl_time_tbl [2 x 5]>

An ensemble can be added to a Nested modeltime table.

ensem <- nested_modeltime_tbl %>%
    ensemble_nested_weighted(
        loadings       = c(2,1),
        control        = control_nested_fit(allow_par = FALSE, verbose = TRUE)
    )

We can then verify the model has been added.

This produces an ensemble .model_id 3, which is an ensemble of the first two models.

# A tibble: 4 x 6
  id    .model_id .model         .model_desc                   .type .calibration_data
  <fct>     <dbl> <list>         <chr>                         <chr> <list>
1 1_3           1 <workflow>     PROPHET                       Test  <tibble [52 x 4]>
2 1_3           2 <workflow>     XGBOOST                       Test  <tibble [52 x 4]>
3 1_3           3 <ensemble [2]> ENSEMBLE (WEIGHTED): 2 MODELS Test  <tibble [52 x 4]>

We can verify the loadings have been applied correctly. Note that the loadings will be applied based on the model with the lowest RMSE.

Note that the xgboost model gets the 66% loading and prophet gets 33% loading. This is because xgboost has the lower RMSE in this case.

-- Modeltime Ensemble -------------------------------------------
    Ensemble of 2 Models (WEIGHTED)

# Modeltime Table
# A tibble: 2 x 6
  .model_id .model     .model_desc .type .calibration_data .loadings
      <int> <list>     <chr>       <chr> <list>                <dbl>
1         1 <workflow> PROPHET     Test  <tibble [52 x 4]>     0.333
2         2 <workflow> XGBOOST     Test  <tibble [52 x 4]>     0.667