Forecast with a pretrained time-series model

FOUNDATION() provides access to large pretrained forecasting models within fable. Unlike traditional statistical models, these models do not require training on the supplied data. The fitting step stores the observed history, and forecasts are generated directly from the pretrained model.

Usage

FOUNDATION(
  formula,
  backend = c("chronos", "timesfm", "sundial", "moirai"),
  model_id = NULL,
  device = c("cpu", "cuda"),
  n_samples = 200L
)

Arguments

formula: The series to forecast, for example observation. For count data, use log(observation) if variance stabilisation is required. fable automatically back-transforms forecasts. Additional predictors are not supported.
backend: Pretrained model backend to use: "chronos", "timesfm", "sundial", or "moirai".
model_id: Optional Hugging Face model identifier. If NULL, the default model for the selected backend is used.
device: Computation device: "cpu" (default) or "cuda".
n_samples: Number of forecast samples to generate. Defaults to 200.

Value

A fable model specification for use with model, get_cv, or get_fcast.

Details

Forecasts are returned in the same format as other fable models, allowing them to be used with functions such as get_cv and get_fcast.

The models run through Python using reticulate. On the first forecast in a session, required Python dependencies and model weights are installed and downloaded automatically. These are cached for subsequent forecasts.

Available models (backend):

"chronos": Amazon Chronos. Default: amazon/chronos-t5-small.
"timesfm": Google TimesFM. Default: google/timesfm-2.5-200m-pytorch.
"sundial": Tsinghua Sundial. Default: thuml/sundial-base-128m.
"moirai": Salesforce Moirai. Default: Salesforce/moirai-1.1-R-small.

Some models provide only a limited number of quantiles, so extreme prediction intervals may be approximate.

Author

Cyril Geismar

Examples

if (FALSE) { # \dontrun{
ncast <- get_data("covid", "ny", revisions = TRUE) |> get_ncast()
get_fcast(ncast, models = c(
  default_models(),
  list(
    CHRONOS = FOUNDATION(log(observation), "chronos"),
    MOIRAI  = FOUNDATION(log(observation), "moirai")
  )
))
} # }