Produce a forward-looking forecast

Refits models on the full series, forecasts h steps ahead and combines them into an equal-weight ensemble. Accepts either an accidda_cv from get_cv — whose ranking selects the best top_n models — or an accidda_data / accidda_ncast, which forecasts every model in models.

Usage

get_fcast(x, models = default_models(), h = 4, top_n = 3)

Arguments

x

An accidda_cv (get_cv), accidda_ncast or accidda_data.

models

Named list of fable models. Defaults to default_models. For an accidda_cv, leave unset to forecast its top_n ranked models, or pass models to forecast a set of your own.

h

Integer. Forecast horizon, in reporting-interval steps (weeks for weekly data). Default 4; for an accidda_cv, defaults to the cross-validation horizon.

top_n

Integer. Number of top-ranked CV models to ensemble. Used only when x is an accidda_cv and models is unset. Default 3.

Value

An accidda_fcast object:

hub

Hub-format forecast (model_out_tbl, oracle_output).

score

Model ranking from the accidda_cv, or NULL.

meta

models, top_n, h, location, target, interval, nowcast.

Export with to_respilens.

Details

If the input carries ncast_lower / ncast_upper (from get_ncast), forecasts are pooled across the nowcast median and 95\

Examples

if (FALSE) { # \dontrun{
ncast <- get_data("covid", "ny", revisions = TRUE) |> get_ncast()
cv    <- ncast |> get_cv(eval_start_date = "2025-01-01", h = 4)

get_fcast(cv, top_n = 3)                 # reuse the cross-validation ranking
get_fcast(cv, models = default_models()) # forecast a different set; keeps $score
get_fcast(ncast)                         # or forecast the default models directly
} # }