Several models combined

The weights are learnt from the series itself: the last origins periods are forecast by every member from the data before them, and the errors decide how much each member counts. Then the members are fitted on the whole series and their forecasts combined. As a candidate in a backtest(), an ensemble learns its weights again at every origin.

Usage

model_ensemble(
  members,
  weighting = c("inverse_error", "equal", "median", "stacked"),
  origins = NULL,
  horizon = NULL,
  top = NULL
)

Arguments

members

A list of models.

weighting

"inverse_error" (in inverse proportion to the mean squared error), "equal", "median" or "stacked" (the weights, none negative and adding up to one, with the smallest squared error).

origins

Periods forecast to learn the weights (default 12).

horizon

How far ahead those forecasts go (default: one seasonal cycle, at most 12).

top

Keep only the members with the smallest error.

Value

A model specification, to use with fit_model(), forecast_model() or backtest().

See also

Other models: model_arima(), model_auto_arima(), model_box_cox(), model_croston(), model_decomposed(), model_ets(), model_holt_winters(), model_log_linear(), model_mean(), model_prophet(), model_tbats(), model_theta()

Examples

fit <- fit_model(model_ensemble(candidates_default(), top = 3), AirPassengers)
fit$params
#>      weight_holt_winters        weight_log_linear weight_log_arima_011_011 
#>                0.4335252                0.3251547                0.2413201