Skip to content

Chronos-2 via brulee

Source: R/tabular_chronos_brulee.R
details_tabular_chronos_brulee.Rd

brulee::brulee_chronos() uses a pretrained time-series model to make quantile or point forecasts.

Details

For this engine, there are multiple modes: quantile regression and regression

Tuning Parameters

This model has no tuning parameters. Chronos-2 is a pretrained forecasting model with fixed weights: no training is performed, the historical (“context”) data is ingested at fit time, and the model forecasts a fixed horizon. On first use, the engine downloads the pretrained weights (about 500MB) and caches them locally.

Forecast configuration is supplied through engine arguments that mirror brulee::brulee_chronos():

  • prediction_length: the number of future time steps to forecast.

  • id_column: the name of the column that identifies each series.

  • timestamp_column: the name of the column with the time values.

  • model_id and revision: the pretrained model to use and its revision.

  • device: the torch device to use (e.g., "cpu").

  • cache_dir: where the pretrained weights are cached.

The quantile_levels are taken from the mode (via set_mode()) and forwarded to the fit automatically. Note that the underlying model only allows a specific set of 21 quantile levels: (0.01, 0.05, …, 0.95, 0.99).

Translation from parsnip to the original package (quantile regression)

tabular_chronos() |>
  set_engine("brulee", prediction_length = 14) |>
  set_mode("quantile regression", quantile_levels = (1:9) / 10) |>
  translate()

## tabular chronos Model Specification (quantile regression)
##
## Engine-Specific Arguments:
##   prediction_length = 14
##
## Computational engine: brulee
##
## Model fit template:
## brulee::brulee_chronos(formula = missing_arg(), data = missing_arg(),
##     prediction_length = 14, quantile_levels = quantile_levels)

## Quantile levels: 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, and 0.9.

Translation from parsnip to the original package (regression)

The regression mode returns the median point forecast.

tabular_chronos() |>
  set_engine("brulee", prediction_length = 14) |>
  set_mode("regression") |>
  translate()

## tabular chronos Model Specification (regression)
##
## Engine-Specific Arguments:
##   prediction_length = 14
##
## Computational engine: brulee
##
## Model fit template:
## brulee::brulee_chronos(formula = missing_arg(), data = missing_arg(),
##     prediction_length = 14)

Preprocessing requirements

There are no preprocessing requirements; the data are used as the historical context for the forecast. However, any date column that is passed to the data is converted to a numeric value (similar to what as.numeric(date) would do). It might be beneficial to pass factors or indicators for cyclic characteristics of the date like year, week, day of the week etc.

The parsnip interface forecasts a single series: predict() returns one row per horizon step, which cannot unambiguously represent more than one series, so supplying data with multiple id_column values is an error. For multi-series forecasting, call brulee::brulee_chronos() directly.

Case weights

The underlying model implementation does not allow for case weights.

Prediction types

parsnip:::get_from_env("tabular_chronos_predict") |>
  dplyr::filter(engine == "brulee") |>
  dplyr::select(mode, type)

## # A tibble: 2 x 2
##   mode                type
##   <chr>               <chr>
## 1 quantile regression quantile
## 2 regression          numeric

References

  • Ansari, A. F., Shchur, O., Küken, J., Auer, A., Han, B., Mercado, P., et al. (2025). Chronos-2: From univariate to universal forecasting. arXiv preprint arXiv:2510.15821.