NEWS
mlr3forecast 0.2.0 (2026-08-24)
- BREAKING CHANGE: Key columns are no longer features by default. Restore the feature role explicitly when needed.
- BREAKING CHANGE:
po("fcstavg") was renamed to po("fcst.avg") to match the id prefix of the other PipeOps.
- feat:
DirectForecaster gained $importance(), $selected_features(), and $oob_error() methods returning one result per horizon model, named like $native_model.
- feat: New learner
fcst.ar wrapping stats::ar(), fitting autoregressive models by Yule-Walker, Burg, OLS, or maximum likelihood with optional AIC order selection.
- feat:
fcst.mean gained the bootstrap and npaths parameters for empirical quantiles resampled from the residuals.
- feat: New learner
fcst.sparma wrapping smooth::sparma(), fitting sparse ARMA models whose orders map to specific lags instead of expanding polynomials.
- feat:
forecast() now validates newdata as a data frame with unique column names.
- feat:
partition() now validates ratio before partitioning a TaskFcst.
- feat:
mlr3::set_threads() support: the num.cores parameter of fcst.arfima, fcst.auto_arima, fcst.nnetar, fcst.bats, and fcst.tbats now carries the "threads" tag, and setting num.cores to a value greater than one enables the corresponding parallel switch at train time.
- fix:
fcst.nnetar now declares nnet in its packages so parallel training finds predict.nnet on the main process.
- feat:
pipeline_fcst_local() now accepts any object supported by as_graph() and validates key.
- feat:
PredictionFcst now stores explicit roles for extra columns in $col_roles, replacing type-based detection (#52).
- feat:
RecursiveForecaster now supports validation and internal tuning (configure with set_validate()) and delegates $importance(), $selected_features(), and $oob_error() to the wrapped graph.
- feat:
TaskFcst now accepts character or integer keys, while tsibble, tsf, and tsbox converters preserve their types.
- fix: Numeric
freq values now represent the seasonal period, while the grid step is inferred from the order column.
- fix: Both forecasters no longer advertise learner properties they cannot honour, fixing failures when tuning with
AutoTuner. This drops the hotstart properties for both and additionally validation, internal tuning, importance, selected features, and OOB error for DirectForecaster.
- fix:
default_measures("fcst") now returns regr.mse, so forecast resampling and benchmark results can be aggregated without an explicit measure.
- fix:
DirectForecaster now rejects empty or duplicate horizons values.
- fix:
fcst.arima, fcst.auto_adam, fcst.ets, fcst.gum, fcst.rlgt, and fcst.stlm parameter definitions now match the wrapped functions' defaults, ranges, and dependencies.
- fix:
fcst.mase, fcst.msis, and fcst.rmsse now infer period from task$freq unless it is set.
- fix:
fcst.nnetar now supports quantile predictions and uses bootstrap, npaths, and innov when simulating their prediction intervals.
- fix:
fcst.prophet now supports logistic growth through a required cap task feature and an optional floor task feature.
- fix:
fcst.sma now fits the complete training task and no longer exposes the incompatible holdout parameter.
- fix:
fcst.tslm now omits season from its generated default formula for nonseasonal tasks.
- fix:
PipeOpFcstAvg now declares its required packages and the "fcst" tag instead of dropping them.
- fix:
rsmp("fcst.holdout", n = 0) now puts no observations into the training set instead of all of them.
mlr3forecast 0.1.0 (2026-07-22)