Added a wrapper for bsts models that takes the old internals of forecast_rt. The idea is to put in a vector y, samples, and a time horizon and to get out a data.frame of samples (rows) by horizon (columns).
Updated all examples to work with the new wrapper
Updated docs to match change.
Stop raw returns form evaluate_model being the default
Add bsts_model wrapper for fable (fable_model).
Add other models to docs and examples.
Added support for uncertain observations plotting (via reducing alpha and plotting observations as points).
Provide docs on how to add generic models.
Comment
fable has several dependencies for different models that are required in order for it to run them. The error checking in compare_timeseries screens out these messages which can make it hard to debug. The model implementation process should include an initial check that the fable model works at the lowest level (i.e. with fable_model as per the examples). Similarly fable natively uses the future package for parallelisation but also requires the future.apply package which is not installed by default. If a future is detected (i.e. for use with compare_timeseries) then future.apply must be installed in order for a fable model to be run, even though it is not running in parallel in this instance. Obviously these features are not ideal and quite annoying on initial set up if not aware. Some flagging has been added to the fable_model docs to indicate the problem.
bsts
models that takes the old internals offorecast_rt
. The idea is to put in a vectory
, samples, and a time horizon and to get out a data.frame of samples (rows) by horizon (columns).evaluate_model
being the defaultbsts_model
wrapper forfable
(fable_model
).Comment
fable
has several dependencies for different models that are required in order for it to run them. The error checking incompare_timeseries
screens out these messages which can make it hard to debug. The model implementation process should include an initial check that thefable
model works at the lowest level (i.e. withfable_model
as per the examples). Similarlyfable
natively uses thefuture
package for parallelisation but also requires thefuture.apply
package which is not installed by default. If afuture
is detected (i.e. for use withcompare_timeseries
) thenfuture.apply
must be installed in order for afable
model to be run, even though it is not running in parallel in this instance. Obviously these features are not ideal and quite annoying on initial set up if not aware. Some flagging has been added to thefable_model
docs to indicate the problem.