The param parameter in the regression function provided by the user
is changed to a … parameter.
Extreme gradient boosting is supported using the xgboost
package.
utsf 1.3.2
The model tree is now built with the parameter committees set to
5
The param parameter used to set arguments to the regression models
in the create_model function is no longer needed, you can set the
arguments directly.
utsf 1.3.1
The pre-processing for dealing with trending series is now specified
in a simpler way.
The vignette has been improved.
The way in which tuning parameters are specified to user’s models
has changed.
utsf 1.3.0
The lags argument in the function for building the model (now
create_model) now can be an unordered integer vector.
The lags argument in the function for building the model (now
create_model) now must be an integer vector.
match.arg() is used so the options are visible to the user in the
help.
A main change is that the functionality of the forecast function,
that did a lot of things, is now distributed in several functions:
create_model() (build the model), forecast() (do the forecasts), efa()
for assessing forecast accuracy and tune_grid() for parameter
tuning.
Prediction intervals are optionally computed.
utsf 1.2.1
The default value of parameter transform_features in trend function
is again TRUE.
utsf 1.2.0
The estimated forecast accuracy per horizon is also computed.
Now it is possible to use only 1 lag with additive or multiplicative
transformation, if the features are not transformed.
Now it is possible to transform only the target (and not the
features) with the multiplicative transformation.
An error is produced if a too large autorregresive lag is used.
An error is produced in method KNN when k is greater than the size
of the training set.
A warning is produced when the time series is too short to estimate
forecast accuracy.
utsf 1.1.0
Improvements in estimation of forecast accuracy with rolling origin
evaluation.
The way in which pre-processings are specified has changed.