Forecast and evaluate time series models

Selects and evaluates multiple time series models via expanding-window cross-validation, then produces a final forward-looking forecast.

Usage

get_fcast(df, eval_start_date, h = 4, top_n = 3, extra_models = NULL)

Arguments

df

An accidda_ncast object from get_ncast, or an accidda_data object from check_data or get_data. If the underlying data frame contains ncast_lower and ncast_upper columns, nowcast uncertainty propagation is enabled automatically.

eval_start_date

Date or string coercible to Date. First date at which forecasts are evaluated. At least 52 weeks of data must precede this date.

h

Integer. Forecast horizon in weeks. Default is 4.

top_n

Integer. Number of top ranked models included in the ensemble. Default is 3.

extra_models

Named list of additional forecasting models passed to fabletools::model(). Each element must reference the observation column and be compatible with fabletools::forecast().

Custom models can be constructed using the fable modelling framework, see the fabletools extension_models vignette.

The name of each list element is used as the model label in the output.

Value

An object of class accidda_fcast containing:

hubcast

Hub-format forecast object (model_out_tbl and oracle_output).

score

Model ranking based on rolling origin WIS.

plot

ggplot2 object showing forecasts and prediction intervals.

Details

Cross-validation. From eval_start_date onwards, models are repeatedly refitted on all data up to each cutoff and used to forecast the next h weeks. Models are ranked by WIS; the best top_n form an equal-weight ensemble.

Nowcast uncertainty. When an accidda_ncast object is supplied (or an accidda_data whose data frame contains ncast_lower and ncast_upper columns), cross-validation runs on the observation column (the median corrected series). The final forecast is then produced three times (from the median, lower 95\ and the resulting distributions are pooled, so prediction intervals reflect both model uncertainty and nowcast uncertainty.

Examples

if (FALSE) { # \dontrun{
extra_models <- list(
  CUSTOM_ARIMA = ARIMA(observation ~ pdq(1, 1, 0)),
  PROPHET = fable.prophet::PROPHET(observation),
  EPIESTIM = EPIESTIM(observation, mean_si = 3.96, std_si = 4.75)
)

get_fcast(
  df,
  eval_start_date = "2025-01-01",
  h = 4,
  top_n = 3,
  extra_models = extra_models
)
} # }