Glossary of Common Terms#

The glossary below defines common terms and API elements used throughout sktime.


A single-purpose piece of code that practitioners write to solve a particular applied problem. Compare with toolbox and framework.


A technique in ensemble learning where multiple models are trained on different subsets of the training data, and individual model outputs are averaged by some rule (e.g., majority vote) to obtain a consensus prediction.

Composite estimator#

An estimator that consists of multiple other component estimators which can vary. An example is a pipeline consisting of a transformer and forecaster. The term can refer both to the class and its instance. For composite estimators, a tag can depend on components, such as capability:missing_data, and a scitype that depends on the components’ scitypes, e.g., the scitype of a pipeline being a forecaster or a classifier, depending on whether its last element is a forecaster or a classifier. Users familiar with software engineering patterns should note that this term may be used in a different sense than “composite pattern”: in the context of scikit-learn, the “composite estimator” combines both the composite pattern and the strategy pattern.


Within a learning task endogenous variables are determined by exogenous variables or past timepoints of the variable itself. Also referred to as the dependent variable or target.

Ensemble learning:#

A technique in which multiple models are combined to improve the overall performance of a predictive model.


An algorithm of a specific scitype, implementing the python class interface defined by the scitype. Individual estimators correspond to concrete classes, implementing the interface defined by the base class for the scitype. For example, the ARIMA class is an estimator of scitype "forecaster". Users should distinguish the python class, which can be seen as a blueprint, from an instance, which is a concrete object created from the blueprint, with specific parameter settings, and which can be fitted or applied to data. Somewhat confusingly, both the class (blueprint) and the instance (concrete object) are often referred to as “estimator” in scikit-learn parlance. Users should also take note of the distinction between “concrete class” in software engineering terms, which is the ARIMA (python) class, as it implements BaseForecaster (the “abstract class”), and the “concrete object”, which is a python instance of a python class. Estimators are objects with a fit method - not all scitype-s in sktime are estimators, e.g., performance metrics.


Within a learning task exogenous variables are external factors whose pattern of impact on tasks’ endogenous variables must be learned. Also referred to as independent variables or features.

Extension templates#

sktime is designed to be easily extendable, with 3rd and 1st party additions in the form of API compliant objects. To facilitate this, sktime provides a set of extension templates for power users to implement their own objects, such as forecasters, transformers, classifiers. The extension templates are found in the extension_templates folder, these are fill-in-the-blank templates that can be used to create new objects compliant with the sktime API. Each template is specific to the scitype of the object to be implemented, and there are different templates for a given scitype, depending on simplicity vs feature richness. The templates instruct a power user on setting of tags, and implementation of scitype-specific methods. The methods are usually private, e.g., _fit, _predict, while boilerplate is taken care of by the base class. For further details and a step-by-step tutorial on 1st and 3rd party extensions, see the guide on Implementing Estimators. For power users familiar with software engineering patterns: the extension templates make use of the template pattern for the extension contract, ensuring compliance with the strategy pattern for the user contract, defined by the scitype specific interface.

Feature extraction:#

A technique used to extract useful information from raw data. In time series analysis, this may involve transforming the data to a frequency domain, decomposing the signal into components, or extracting statistical features.


A learning task focused on prediction future values of a time series. For more details, see the Introduction.


A collection of related and reusable software design templates that practitioners can copy and fill in. Frameworks emphasize design reuse. They capture common software design decisions within a given application domain and distill them into reusable design templates. This reduces the design decision they must take, allowing them to focus on application specifics. Not only can practitioners write software faster as a result, but applications will have a similar structure. Frameworks often offer additional functionality like toolboxes. Compare with toolbox and application.


The ability of a predictive model to perform well on unseen data. A model that overfits to the training data may not generalize well, while a model that underfits may not capture the underlying patterns in the data.


A parameter of a machine learning model that is set at construction. Usually, this affects the model’s performance. Examples include the learning rate in a neural network, the number of trees in a random forest, or the regularization parameter in a linear model.


A member of the set of entities being studied and which an ML practitioner wishes to generalize. For example, patients, chemical process runs, machines, countries, etc. May also be referred to as samples, examples, observations or records depending on the discipline and context.

Model selection:#

The process of selecting the best machine learning model for a given task. This may involve comparing the performance of different models on a validation set, or using techniques like grid search to find the best hyperparameters for a given model.


sktime supports multiple in-memory specifications for time series data and other objects. Such an in-memory specification is called mtype (short for “machine type”). Each mtype is represented by a string - e.g., pd-multiindex, which defines the data format and the data structure. For example, a pd-multiindex mtype is a collection of time series, represented as a 2-level MultiIndex-ed pandas.DataFrame, with columns representing variables, rows indexed by (instance, timepoint), where the timepoint level must be range-like or datetime-like. Each mtype implements an abstract data type, a (data) scitype, for instance Panel which refers to the abstract type of a collection of time series, with instance, timepoint and variable dimensions. In this terminology, the pd-multiindex mtype implements the (abstract) Panel scitype. Data containers can be checked for compliance with a given mtype using the sktime.datatypes.check_is_mtype function; all mtypes can be listed in sktime.datatypes.MTYPE_REGISTER. For more details on the general concept, see the datatypes and datasets user guide.

Multivariate time series#

Multiple time series. Typically observed for the same observational unit. Multivariate time series is typically used to refer to cases where the series evolve together over time. This is related, but different than the cases where a univariate time series is dependent on exogenous data.

Panel time series#

A form of time series data where the same time series are observed observed for multiple observational units. The observed series may consist of univariate time series or multivariate time series. Accordingly, the data varies across time, observational unit and series (i.e. variables).


Reduction refers to decomposing a given learning task into simpler tasks that can be composed to create a solution to the original task. In sktime reduction is used to allow one learning task to be adapted as a solution for an alternative task.

Scientific type#

See scitype.


Short for scientific type, denotes the abstract type of an sktime object, data container or estimator. One example of an estimator scitype is "forecaster", which denotes the abstract concept of a forecaster with (abstract) fit, predict, update methods. An example of a data scitype is Panel, denoting the abstract concept of an indexed collection of time series. Scitypes are represented by strings, and are implemented by concrete types. For data containers, concrete types are mtype-s (see there), for estimators, concrete types are python base interfaces, such as defined by BaseForecaster or BaseClassifier. Valid estimator scitypes, with their corresponding base classes, are listed in sktime.registry.BASE_CLASS_SCITYPE_LIST. All estimators of a given scitype can be listed using sktime.registry.all_estimators, and the scitype of a given estimator can be inferred by the sktime.registry.scitype utility. Compliance with concrete implementations of data scitypes can be checked using the sktime.datatypes.check_is_scitype utility; for estimators, compliance is checked using sktime.utils.check_estimator. For more details on data scitpyes, see mtype. For more details on estimator scitypes, see the user guides on individual learning tasks.


When a time series is affected by seasonal characteristics such as the time of year or the day of the week, it is called a seasonal pattern. The duration of a season is always fixed and known.


Is a setting where each timepoint of the univariate time series being measured for each instance are treated as features and stored as a primitive data type in the DataFrame’s cells. E.g., there are N instances of time series and each has T timepoints, this would yield a pandas DataFrame with shape (N, T): N rows, T columns.


Tags are string keyed value fields, used to identify properties of an object, or set flags for internal boilerplate. An example of a tag is capability:multivariate, a boolean flag, which indicates whether the object offers genuine support for multivariate time series. Objects with a given capability - that is, objects filtering by certain tag value - can be listed or filtered using sktime.registry.all_estimators. In sktime, most objects are scikit-base objects and implement the tag interface via get_tag or get_tags. Some tags are for internal or extender use only, e.g., X_inner_mtype, which allows an extender to specify the mtype of the inner data container they would like to work with. A list of all tags and their meaning, optionally filtered by the scitype of object they apply to, can be obtained from sktime.registry.all_tags. Further details on tags, for developers, can be found in the specification sheet that is part of the extension templates.

Time series#

Data where the variable measurements are ordered over time or an index indicating the position of an observation in the sequence of values.

Time series annotation#

A learning task focused on labeling the timepoints of a time series. This includes the related tasks of outlier detection, anomaly detection, change point detection and segmentation.

Time series classification#

A learning task focused on using the patterns across instances between the time series and a categorical target variable.

Time series clustering#

A learning task focused on discovering groups consisting of instances with similar time series.

Time series decomposition:#

A technique used to separate a time series into its underlying components, such as trend, seasonality, and noise. This can be useful for understanding the patterns in the data and for modeling each component separately.

Time series regression#

A learning task focused on using the patterns across instances between the time series and a continuous target variable.


The point in time that an observation is made. A time point may represent an exact point in time (a timestamp), a timeperiod (e.g. minutes, hours or days), or simply an index indicating the position of an observation in the sequence of values.


A collection of related and reusable functionality that practitioners can import to write applications. Toolboxes emphasize code reuse. Compare with framework and application.


When data shows a long-term increase or decrease, this is referred to as a trend. Trends can also be non-linear.

Univariate time series#

A single time series. While univariate analysis often only uses information contained in the series itself, univariate time series regression and forecasting can also include exogenous data.


Refers to some measurement of interest. Variables may be cross-sectional (e.g. time-invariant measurements like a patient’s place of birth) or time series.