News
- Date: 2022-05-12
- Added the tsdistributions package. Estimation uses automatic differentiation (AD)
as in the other packages.
- Date: 2022-04-30
- Missing value handling now available in the tsvets package for both
estimation and filtering.
- Date: 2022-04-13
- Most of the packages have replaced the use of snow and foreach with the
functionality of the future
package. This eliminates the cores argument in many functions (such as
tsbacktest)
and hence may break some backwards compatibility. However, the user can now
provide a future plan (including using remote servers) and is generally
much more flexible. Additionally, tracing is now handled by the progressr
package which requires the user to provide handlers for the type of output
they wish (including beeps).
- A lot of additional functionality has been added, some of it experimental
such as the
tscalibrate method for empirical distribution calibration via
backtesting with support in the predict function for the output of this
calibration. In other places we have added extra features such as in the
tsdecompose method to return a simplifed state space set of components
(Trend, Seasonal, ARMA if available, Irregular). Additionally, we have
changed the indexing of the decomposed states so that summing them up
will return the exact predictive distribution.
- The logit transformation is part of the
tstransform method, from the tsaux
package, which also includes the Box-Cox. Unifying the transformations under
this auxiliary function will allow us to expand to potentially other methods
going forward. The logit was chosen in order to model certain series which
are bounded in the [0,1] domain, and makes an assumption that they follow
a Logit-Normal distribution.
- Date: 2021-11-19
- The tsvetsad package now implements
automatic differentiation (AD) for the VETS model implemented in the tsvets
package. An argument use_autofiff is now available
in the
estimate method which will then dispatch to the estimate_ad method.
Gradients and Hessian are implemented, but for the latter an additional option use_hessian
needs to be set to TRUE and it currently defaulted to FALSE in the tsvetsad package.
The package is still in beta testing as speed optimization need to be undertaken
to perform more competitively to the non-AD method.
- Date: 2021-06-20
- The tsissmad package now implements
automatic differentiation (AD) for the ISSM model implemented in the tsissm
package. An argument use_autofiff is now available
in the
estimate method which will then dispatch to the estimate_ad method.
Gradients and Hessian are implemented, but for the latter an additional option use_hessian
needs to be set to TRUE and it currently defaulted to FALSE in the tsissmad package.
Only trigonometric seasonality is implemented but regular seasonality may be included
in a future update. Additionally, only the nlminb solver is used for autodiff.
- Date: 2021-05-17
- The tsetad package now implements
automatic differentiation (AD) for all ETS models implemented in the tsets
package. An argument use_autofiff is now available
in the
estimate method which will then dispatch to the estimate_ad method.
Gradients and Hessian are implemented, and it is suggested that the nlminb
solver be used which makes use of both. Speedups for highly parameterized
or difficult (powerMAM) models has been observed to be upto 16x when using
AD. Addition of AD to the tsissm and tsvets packages may be imlemented in due
course.
- Missing values are now handled during model estimation using the predict
step and only calculating the likelihood for non-missing values. Applies
to the tests and tsissm packages. For the tsvets package we have
not yet decided on the implementation. This may come at a future date.