tidy() now reports extra regressor estimates as
coefficients on the original data scale (additive regressors in units of
the response, multiplicative regressors as proportional effects), rather
than prophet’s internal scaled parameters. Other terms remain on the
internal scale, as documented in ?tidy.fbl_prophet.forecast() no longer forwards ... to
prophet::predictive_samples(), which doesn’t accept
additional arguments. ... is now documented as unused.generate() for prophet models (#4), simulating
sample paths (trend including future changepoints, seasonality, holidays
and regressors) with normally distributed or, with
bootstrap = TRUE, resampled-residual observation noise.
Works in-sample and out-of-sample, including with MCMC estimation and
transformed responses.refit() for prophet models (#30). By default the
estimated parameters are kept and the fitted values, residuals and
components are recomputed for the new data;
reestimate = TRUE fits the model again on the new data
using the original estimation arguments. Prophet cannot be updated
incrementally, so there is no stream() method.growth() now supports type = "flat", for a
constant trend without changepoints.holiday() gains a country argument to
include prophet’s built-in country holidays (via
prophet::add_country_holidays()), with or without a
holidays table. tidy() now names (and orders)
holiday terms from the fitted model, so country holidays and repeated
holiday dates are supported.season() gains a condition argument (a
bare logical column name) for conditional seasonality (via
prophet::add_seasonality(condition.name = )). The column
must also be present in new_data when forecasting.prophet() now documents and supports the
mcmc.samples and backend arguments (passed to
prophet::prophet()) for MCMC estimation and choosing the
"rstan" or "cmdstanr" Stan backend.
tidy() and glance() summarise MCMC draws by
their posterior means.forecast() now supports times = 0, which
returns point forecasts (as degenerate distributions) without simulating
sample paths. The times argument is now documented,
including its speed and memory trade-off (#26).components() now includes the holidays,
extra_regressors_additive and
extra_regressors_multiplicative terms when they are present
in the model.stan.fit), reducing model size (#26).log(x) or
I(x^2)) are now supported, using make.names()
to name them in the model and in tidy() (#32).?prophet and
the introduction vignette.uncertainty.samples = 0
(previously the argument was misspelt and ignored), so fitting no longer
computes unused uncertainty samples.forecast() (#33).Small patch for compatibility with fabletools v1.0.0.
forecast(),
components(), fitted(),
residuals().