Time series transformations

The sktime.transformations module contains classes for data transformations.

All (simple) transformers in sktime can be listed using the sktime.registry.all_estimators utility, using estimator_types="transformer", optionally filtered by tags. Valid tags can be listed using sktime.registry.all_tags.

For pairwise transformers (time series distances, kernels), instead see _transformations_pairwise_ref.

Transformations are categorized as follows:

Category	Explanation	Example
Composition	Building blocks for pipelines, wrappers, adapters	Transformer pipeline
Series-to-features	Transforms series to float/category vector	Length and mean
Series-to-series	Transforms individual series to series	Differencing, detrending
Series-to-Panel	transforms a series into a panel	Bootstrap, sliding window
Panel transform	Transforms panel to panel, not by-series	Padding to equal length
Hierarchical	uses hierarchy information non-trivially	Reconciliation

Composition

Pipeline building

TransformerPipeline(steps)	Pipeline of transformers compositor.
FeatureUnion(transformer_list[, n_jobs, ...])	Concatenates results of multiple transformer objects.
ColumnEnsembleTransformer(transformers[, ...])	Column-wise application of transformers.
ColumnwiseTransformer(transformer[, columns])	Apply a transformer columnwise to multivariate series.
FitInTransform(transformer[, ...])	Transformer wrapper to delay fit to the transform phase.
MultiplexTransformer(transformers[, ...])	Facilitate an AutoML based selection of the best transformer.
OptionalPassthrough(transformer[, passthrough])	Wrap an existing transformer to tune whether to include it in a pipeline.
InvertTransform(transformer)	Invert a series-to-series transformation.
Id()	Identity transformer, returns data unchanged in transform/inverse_transform.
YtoX([subset_index])	Create exogeneous features which are a copy of the endogenous data.
IxToX([coerce_to_type, level, ix_source])	Create features based on time index or hierarchy values.
TransformByLevel(transformer[, groupby, ...])	Transform by instance or panel.
TransformIf(if_estimator[, param, ...])	Conditional execution of a transformer given a condition from a fittable object.

ColumnTransformer(transformers[, remainder, ...])

Column-wise application of transformers.

FunctionTransformer([func, inverse_func, ...])

Constructs a transformer from an arbitrary callable.

Sklearn and pandas adapters

Tabularizer()	A transformer that turns time series/panel data into tabular data.
TimeBinner(idx[, aggfunc])	Turns time series/panel data into tabular data based on intervals.

TabularToSeriesAdaptor(transformer[, ...])	Adapt scikit-learn transformation interface to time series setting.
PandasTransformAdaptor(method[, kwargs, ...])	Adapt pandas transformations to sktime interface.

Series-to-features transformers

Series-to-features transformers transform individual time series to a collection of primitive features. Primitive features are usually a vector of floats, but can also be categorical.

When applied to panels or hierarchical data, the transformation result is a table with as many rows as time series in the collection.

Summarization

These transformers extract simple summary features.

SummaryTransformer([summary_function, ...])	Calculate summary value of a time series.
WindowSummarizer([lag_feature, n_jobs, ...])	Transformer for extracting time series features.
SplitterSummarizer(transformer[, splitter, ...])	Create summary values of a time series' splits.

DerivativeSlopeTransformer()	Derivative slope transformer.
PlateauFinder([value, min_length])	Plateau finder transformer.
RandomIntervalFeatureExtractor([...])	Random interval feature extractor transform.
FittedParamExtractor(forecaster, param_names)	Fitted parameter extractor.

Shapelets, wavelets, and convolution

ShapeletTransform([min_shapelet_length, ...])	Shapelet Transform.
RandomShapeletTransform([...])	Random Shapelet Transform.
ShapeletTransformPyts([n_shapelets, ...])	Shapelet Transform, from pyts.

Rocket([num_kernels, normalise, n_jobs, ...])	RandOm Convolutional KErnel Transform (ROCKET).
MiniRocket([num_kernels, ...])	MINImally RandOm Convolutional KErnel Transform (MiniRocket).
MiniRocketMultivariate([num_kernels, ...])	MINImally RandOm Convolutional KErnel Transform (MiniRocket) multivariate.
MiniRocketMultivariateVariable([...])	MINIROCKET (Multivariate, unequal length).

DWTTransformer([num_levels])

Discrete Wavelet Transform Transformer.

Distance-based features

MatrixProfile([m])

Return the matrix profile and index profile for each time series of a dataset.

Dictionary-based features

SFA([word_length, alphabet_size, ...])

Symbolic Fourier Approximation (SFA) Transformer.

Moment-based features

SignatureTransformer([augmentation_list, ...])

Transformation class from the signature method.

Feature collections

These transformers extract larger collections of features.

TSFreshRelevantFeatureExtractor([...])	Transformer for extracting time series features via tsfresh.extract_features.
TSFreshFeatureExtractor([...])	Transformer for extracting time series features via tsfresh.extract_features.

Catch22([features, catch24, outlier_norm, ...])

Canonical Time-series Characteristics (Catch22).

Catch22Wrapper([features, catch24, ...])

Canonical Time-series Characteristics (Catch22 and 24), using pycatch22 package.

Series-to-series transformers

Series-to-series transformers transform individual time series into another time series.

When applied to panels or hierarchical data, individual series are transformed.

Lagging

Lag([lags, freq, index_out, ...])	Lagging transformer.
ReducerTransform([lags, freq, shifted_vars, ...])	Transformer for forecasting reduction.

Element-wise transforms

These transformations apply a function element-wise.

Depending on the transformer, the transformation parameters can be fitted.

BoxCoxTransformer([bounds, method, sp, ...])	Box-Cox power transform.
LogTransformer([offset, scale])	Natural logarithm transformation.

ScaledLogitTransformer([lower_bound, ...])

Scaled logit transform or Log transform.

CosineTransformer()

Cosine transformation.

ExponentTransformer([power, offset])	Apply element-wise exponentiation transformation to a time series.
SqrtTransformer([offset])	Apply elementwise square root transformation to a time series.

ScaledAsinhTransformer([...])

Hyperbolic sine transformation and its inverse [Rf00c2e2051c3-1].

Detrending and Decomposition

Detrender([forecaster, model])	Remove a trend from a series.
Deseasonalizer([sp, model])	Remove seasonal components from a time series.
ConditionalDeseasonalizer([...])	Remove seasonal components from time series, conditional on seasonality test.
STLTransformer([sp, seasonal, trend, ...])	Remove seasonal components from a time-series using STL.
MSTL(*[, periods, windows, lmbda, iterate, ...])	Season-Trend decomposition using LOESS for multiple seasonalities.

VmdTransformer([K, kMax, alpha, tau, DC, ...])

Variational Mode Decomposition transformer.

ClearSky([quantile_prob, bw_diurnal, ...])

Clear sky transformer for solar data.

Filtering and denoising

Filter(sfreq[, l_freq, h_freq, filter_kwargs])

Transformer that filters Series data.

BKFilter([low, high, K])

Filter a times series using the Baxter-King filter.

CFFilter([low, high, drift])

Filter a times series using the Christiano Fitzgerald filter.

HPFilter([lamb])

Filter a times series using the Hodrick-Prescott filter.

KalmanFilterTransformerPK(state_dim[, ...])	Kalman Filter is used for denoising data, or inferring the hidden state of data.
KalmanFilterTransformerFP(state_dim[, ...])	Kalman Filter is used for denoising or inferring the hidden state of given data.

ThetaLinesTransformer([theta])

Decompose the original data into two or more Theta-lines.

Differencing, slope, kinematics

Differencer([lags, na_handling, memory])

Apply iterative differences to a timeseries.

SlopeTransformer([num_intervals])

Slope-by-segment transformation.

KinematicFeatures([features])

Kinematic feature transformer - velocity, acceleration, curvature.

Binning, sampling and segmentation

TimeBinAggregate(bins[, aggfunc, return_index])

Bins time series and aggregates by bin.

TSInterpolator(length)

Time series interpolator/re-sampler.

IntervalSegmenter([intervals])	Interval segmentation transformer.
RandomIntervalSegmenter([n_intervals, ...])	Random interval segmenter transformer.

DilationMappingTransformer([dilation])

Dilation mapping transformer.

PAA([frames, frame_size])

Piecewise Aggregate Approximation Transformer (PAA).

SAX([word_size, alphabet_size, frame_size])

Symbolic Aggregate approXimation Transformer (SAX).

PAAlegacy([num_intervals])	Piecewise Aggregate Approximation Transformer (PAA).
SAXlegacy([word_length, alphabet_size, ...])	Symbolic Aggregate approXimation (SAX) transformer.

Missing value treatment

Imputer([method, random_state, value, ...])

Missing value imputation.

DropNA([axis, how, thresh, remember])

Drop missing values transformation.

Seasonality and Date-Time Features

DateTimeFeatures([ts_freq, feature_scope, ...])

DateTime feature extraction, e.g., for use as exogenous data in forecasting.

HolidayFeatures(calendar[, holiday_windows, ...])	Holiday features extraction.
CountryHolidaysTransformer(country[, ...])	Country Holidays Transformer.
FinancialHolidaysTransformer(market[, ...])	Financial Holidays Transformer.

TimeSince([start, to_numeric, freq, ...])

Compute element-wise time elapsed between time index and a reference start time.

FourierFeatures(sp_list, fourier_terms_list)

Fourier Features for time series seasonality.

FourierTransform()

Simple Fourier transform for time series.

PeakTimeFeature([ts_freq, peak_hour_start, ...])

PeakTime feature extraction for use in e.g. tree based models.

Auto-correlation series

AutoCorrelationTransformer([adjusted, ...])	Auto-correlation transformer.
PartialAutoCorrelationTransformer([n_lags, ...])	Partial auto-correlation transformer.

Window-based series transforms

These transformers create a series based on a sequence of sliding windows.

MatrixProfileTransformer([window_length])

Calculate the matrix profile of a time series.

HOG1DTransformer([num_intervals, num_bins, ...])

HOG1D transform.

Multivariate-to-univariate

These transformers convert multivariate series to univariate.

ColumnConcatenator()

Concatenate multivariate series to a long univariate series.

Augmentation

InvertAugmenter()	Augmenter inverting the time series by multiplying it by -1.
RandomSamplesAugmenter([n, ...])	Draw random samples from time series.
ReverseAugmenter()	Augmenter reversing the time series.
WhiteNoiseAugmenter([scale, random_state])	Augmenter adding Gaussian (i.e. white) noise to the time series.

FeatureSelection

These transformers select features in X based on y.

FeatureSelection([method, n_columns, ...])

Select exogenous features.

ElbowClassSum([distance])	Elbow Class Sum (ECS) transformer to select a subset of channels/variables.
ElbowClassPairwise()	Elbow Class Pairwise (ECP) transformer to select a subset of channels.

Subsetting time points and variables

These transformers subset X by time points (pandas index or index level) or variables (pandas columns).

ColumnSelect([columns, integer_treatment, ...])	Column selection transformer.
IndexSubset([index_treatment])	Index subsetting transformer.

Adapters to other frameworks

Generic framework adapters that expose other frameworks in the sktime interface.

TemporianTransformer(function[, compile])

Applies a Temporian function to the input time series.

Panel transformers

Panel transformers transform a panel of time series into a panel of time series.

A panel transformer is fitted on an entire panel, and not per series.

Equal length transforms

These transformations ensure all series in a panel have equal length

PaddingTransformer([pad_length, fill_value])

Padding panel of unequal length time series to equal, fixed length.

TruncationTransformer([lower, upper])

Truncate unequal length panels to lower/upper bounds.

Dimension reduction

PCATransformer([n_components, copy, whiten, ...])

Principal Components Analysis applied to panel of time series.

Series-to-Panel transformers

These transformers create a panel from a single series.

Bootstrap transformations

MovingBlockBootstrapTransformer([n_series, ...])	Moving Block Bootstrapping method for synthetic time series generation.
SplitterBootstrapTransformer([splitter, ...])	Splitter based Bootstrapping method for synthetic time series generation.
STLBootstrapTransformer([n_series, sp, ...])	Creates a population of similar time series.

Panel-to-Series transformers

These transformers create a single series from a panel.

Merger([method, stride])

Aggregates Panel data containing overlapping windows of one time series.

Outlier detection, changepoint detection

HampelFilter([window_length, n_sigma, k, ...])

Use HampelFilter to detect outliers based on a sliding window.

ClaSPTransformer([window_length, ...])

ClaSP (Classification Score Profile) Transformer.

Hierarchical transformers

These transformers are specifically for hierarchical data and panel data.

The transformation depends on the specified hierarchy in a non-trivial way.

Aggregator([flatten_single_levels])

Prepare hierarchical data, including aggregate levels, from bottom level.

Reconciler([method])

Hierarchical reconciliation transformer.