Small release to resolve check issues with the development and patched versions of R. The release includes some minor improvements to the output consistency of initial states in
ETS() models, the passing of arguments in
ARIMA() models, and handling of missing values in
- Display of ETS initial states now use a
state[t] notation to describe the state’s position in time (#329, #261).
- Allowed specifying
method argument in
- Improved handling of missing values in
- Fixed error with forecasting and simulating from
NNETAR() estimated using a short series (#326).
AR() fitted values not being re-scaled to match original data (#318).
The release of fabletools v0.3.0 introduced general support for computing h-step ahead fitted values, using the
hfitted(<mdl>, h = ???) function. This release adds model-specific
hfitted() support to ARIMA and ETS models for improved performance and accuracy.
This release adds improved support for refitting models, largely in thanks to contributions by @Tim-TU.
It is also now possible to specify an arbitrary model selection criterion function for automatic
ARIMA() model selection.
refit() method for NNETAR, MEAN, RW, SNAIVE, and NAIVE models (#287, #289, #321. @Tim-TU).
hfitted() method for ETS and ARIMA, this allows fast estimation of h-step ahead fitted values.
generate() method for AR, the
forecast() method now supports bootstrap forecasting via this new method.
- Added the
selection_metric argument to
ARIMA(), which allows more control over the measure used to select the best model. By default this function will extract the information criteria specified by the
trace argument for tracing the selection procedure used in
- Fixed unnecessary warning when forecasting short horizons using
generate() method for NNETAR models when data isn’t scaled (#302).
refit.ARIMA() re-selecting constant instead of using the provided model’s constant usage.
- Fixed use of exogenous regressors in
This release coincides with v0.2.0 of the fabletools package, which contains some substantial changes to the output of
forecast() methods. These changes to fabletools emphasise the distribution in the fable object. The most noticeable is a change in column names of the fable, with the distribution now stored in the column matching the response variable, and the forecast mean now stored in the
.mean column. For a complete summary of these changes, refer to the fabletools v0.2.0 release news: https://fabletools.tidyverts.org/news/index.html
- Forecasts distributions are now provided by the distributional package. They are now more space efficient and allows calculation of distributional statistics including the
- The uncertainty of the drift parameter in random walk models (
SNAIVE()) is now included in data generated with
- Added Syntetos-Boylan and Shale-Boylan-Johnston variants of
- Performance improvements.
- Fixed issue with approximation being used when refitting ARIMA models and when a specific model is requested.
TSLM() models when the data contains missing values.
- Fixed typo in
glance() output of
- The sample path means are now used instead of analytical means when forecasts are produced from sample paths.
- Added autoregressive modelling with
- Better handling of rank deficiency in
- Added bootstrap forecast paths for
ARIMA() specials now allow specifying fixed coefficients via the
- Documentation improvements.
CROSTON() for Croston’s method of intermittent demand forecasting.
- Documentation improvements
- Fixed NNETAR & VAR handling of missing values (#215).
- Fix ETS forecasting with forecast horizons less than the seasonal period (#219).
- Fixed season() special for non-seasonally based time indices (#220)
- Fix issue with simulation forecasting from damped ETS models.
- Added interpolation method for
MEAN() model (#203).
- Added rolling mean option for
MEAN() model (#204).
- Corrected forecast standard error for drift models.
- Support for 9 models and relevant methods: