Object and estimator tags#

Every first-class object in sktime is tagged with a set of tags that describe its properties and capabilities, or control its behavior.

Tags are key-value pairs, where the key is a string with the name of the tag. The value of the tag can have arbitrary type, and describes a property, capability, or controls behaviour of the object, depending on the tag.

For instance, a forecaster has the tag "capability:pred_int": True if it can make probabilistic predictions. Users can find all forecasters that can make probabilistic predictions by filtering for this tag.

This API reference lists all tags available in sktime, and key utilities for their usage.

To search estimators by tags on the sktime webpage, use the Estimator Search Page

To search estimators by tags in a python environment, use the sktime.registry.all_estimators utility.

Inspecting tags, retrieving by tags#

Tags can be inspected at runtime using the following utilities:

to get the tags of an object, use the get_tags method. An object’s tags can depend on its hyper-parameters.
to get the tags of a class, use the get_class_tags method of the class. A class’s tags are static and do not depend on its hyper-parameters. By default, class tags that may vary for instances take the most “capable” value, in the case of capabilities.
to programmatically retrieve all tags available in sktime or for a particular type of object, at runtime, use the registry.all_tags utility
to programmatically retrieve all objects or estimators in sktime, filtered for values of tags, use the registry.all_estimators utility

`all_tags`([estimator_types, as_dataframe])	List all tags in sktime, for objects of a certain type.
`all_estimators`([estimator_types, ...])	List all estimators or objects in sktime, by scitype or tag.

General tags, packaging#

This section lists tags that are general and apply to all objects in sktime. These tags are typically used for typing, packaging and documentation purposes.

`object_type`	Scientific type of the object.
`maintainers`	Current maintainers of the object, GitHub IDs.
`authors`	Authors of the object, GitHub IDs.
`python_version`	Python version requirement specifier for the object (PEP 440).
`python_dependencies`	Python package dependency requirement specifiers for the object (PEP 440).
`env_marker`	Environment marker requirement for the object (PEP 508).
`requires_cython`	Whether the object requires a C compiler present, such as libomp, gcc.

Forecaster tags#

This section lists tags applying to forecasters ("forecaster" type). These tags are used to describe capabilities, properties, and behavior of forecasters.

`capability__exogeneous`	Capability: the forecaster can use exogenous data.
`capability__insample`	Capability: the forecaster can make in-sample predictions.
`capability__pred_int`	Capability: the forecaster can make probabilistic or interval forecasts.
`capability__pred_int__insample`	Capability: the forecaster can make in-sample probabilistic forecasts.
`capability__missing_values`	Capability: the estimator can handle missing data, e.g,, NaNs.
`capability__categorical_in_X`	Capability: If estimator can handle categorical variables in the X argument.
`capability__random_state`	Capability: the estimator can be derandomized using a random_state.
`requires_fh_in_fit`	Behaviour flag: forecaster requires forecasting horizon in fit.
`fit_is_empty`	Property: Whether the estimator has an empty fit method.
`property__randomness`	Property: Degree of randomness vs determinism of the estimator.

Tags for classifiers, regressors, clustering#

This section lists tags applying to time series classifiers, regressors, and clusterers ("classifier", "regressor", "clusterer" types). These tags are used to describe capabilities, properties, and behavior of these types of objects.

`capability__multivariate`	Capability: the estimator can handle multivariate time series.
`capability__multioutput`	Capability: the estimator can handle multi-output time series.
`capability__missing_values`	Capability: the estimator can handle missing data, e.g,, NaNs.
`capability__unequal_length`	Capability: the estimator can handle unequal length time series.
`capability__predict_proba`	Capability: the estimator can make probabilistic predictions.
`capability__feature_importance`	Capability: the estimator can provide feature importance.
`capability__contractable`	Capability: the estimator can be asked to satisfy a maximum time contract.
`capability__train_estimate`	Capability: the algorithm can estimate its performance on the training set.
`capability__random_state`	Capability: the estimator can be derandomized using a random_state.
`property__randomness`	Property: Degree of randomness vs determinism of the estimator.

Tags for ordinary transformers#

This section lists tags applying to ordinary transformers, i.e., objects that transform a single time series object ("transformer" type).

`scitype__transform_input`	The scitype of the input data for the transformer.
`scitype__transform_output`	The scitype of the input data for the transformer.
`scitype__transform_labels`	The scitype of the target data for the transformer, if required.
`requires_x`	Behaviour flag: transformer requires X in fit and transform.
`requires_y`	Behaviour flag: transformer requires y in fit.
`capability__missing_values`	Capability: the estimator can handle missing data, e.g,, NaNs.
`capability__unequal_length`	Capability: the estimator can handle unequal length time series.
`capability__unequal_length__adds`	Capability: transformer may produce unequal length series on equal length input.
`capability__unequal_length__removes`	Capability: the transformer produces equal length series on unequal length input.
`capability__random_state`	Capability: the estimator can be derandomized using a random_state.
`capability__inverse_transform`	Capability: the transformer can carry out an inverse transform.
`capability__inverse_transform__exact`	Capability: whether the inverse transform is an exact inverse to the transform.
`capability__inverse_transform__range`	Capability: the domain of invertibility of the transform.
`capability__bootstrap_index`	Capability: the transformer is a bootstrap that can return bootstrap idx.
`fit_is_empty`	Property: Whether the estimator has an empty fit method.
`transform_returns_same_time_index`	Property: transformer returns same time index as input.
`property__randomness`	Property: Degree of randomness vs determinism of the estimator.

Tags for detectors#

This section lists tags applying to time series detectors ("detector" types). These tags are used to describe capabilities, properties, and behavior of detectors.

`task`	Subtype tag for detectors: type of detection task.
`learning_type`	Learning type of the detection task.
`capability__update`	Capability: whether the estimator can be run in stream or on-line mode.
`capability__multivariate`	Capability: the estimator can handle multivariate time series.
`capability__missing_values`	Capability: the estimator can handle missing data, e.g,, NaNs.
`capability__random_state`	Capability: the estimator can be derandomized using a random_state.
`property__randomness`	Property: Degree of randomness vs determinism of the estimator.

Tags for metrics#

This section lists tags applying to time series metrics ("metric" type).

`lower_is_better`	Property: whether lower metric values are better.
`capability__sample_weight`	Capability: whether the estimator can handle sample weights.
`scitype__y_pred`	The scitype of the predicted target data, for probabilistic metrics.
`requires_y_true`	Behaviour flag: metric requires y_true in evaluate.
`requires_y_pred_benchmark`	Behaviour flag: metric requires y_pred in evaluate.
`requires_y_train`	Behaviour flag: metric requires y_train in evaluate.
`inner_implements_multilevel`	Extension tag: whether the metric implements multilevel evaluation.

Tags for time series aligners#

This section lists tags applying to time series aligners ("aligner" type).

capability__multiple_alignment capability__distance capability__distance_matrix property__alignment_type

Common developer tags#

This section lists tags that are used to control internal behaviour of objects, e.g., the boilerplate layer.

These are primarily useful for power users using the extension templates to create sktime compatible objects.

The tags below have limited use in retrieval or inspection of objects.

`x_inner_mtype`	The machine type(s) the estimator can deal with internally for X.
`y_inner_mtype`	The machine type(s) the estimator can deal with internally for y.
`visual_block_kind`	How to display html representation of a meta-estimator in a jupyter notebook.

Testing and CI tags#

These tags control behaviour of estimators in the sktime continuous integration tests.

They are primarily useful for developers managing CI behaviour of individual objects.

`tests__core`	Whether tests for this estimator are triggered by framework changes.
`tests__vm`	Whether to spin up a separate VM to test the estimator.
`tests__skip_all`	Whether all tests for this estimator should be skipped.
`tests__skip_by_name`	A list of test names that should be skipped for this object.