Forecasting

The sktime.forecasting module contains algorithms and composition tools for forecasting.

All forecasters in sktime can be listed using the sktime.registry.all_estimators utility, using estimator_types="forecaster", optionally filtered by tags. Valid tags can be listed using sktime.registry.all_tags.

A full table with tag based search is also available on the Estimator Search Page (select “forecaster” in the “Estimator type” dropdown).

Base

BaseForecaster()	Base forecaster template class.
ForecastingHorizon([values, is_relative, freq])	Forecasting horizon.

Pipeline composition

Compositors for building forecasting pipelines. Pipelines can also be constructed using *, **, +, and | dunders.

make_pipeline(*steps)

Create a pipeline from estimators of any type.

TransformedTargetForecaster(steps)	Meta-estimator for forecasting transformed time series.
ForecastingPipeline(steps)	Pipeline for forecasting with exogenous data.
ColumnEnsembleForecaster(forecasters)	Forecast each series with separate forecaster.
MultiplexForecaster(forecasters[, ...])	MultiplexForecaster for selecting among different models.
ForecastX(forecaster_y[, forecaster_X, ...])	Forecaster that forecasts exogeneous data for use in an endogeneous forecast.
ForecastByLevel(forecaster[, groupby])	Forecast by instance or panel.
GroupbyCategoryForecaster(forecasters[, ...])	Compositor that utilizes varying forecasters based on inferred category.
Permute(estimator[, permutation, steps_arg])	Permutation compositor for permuting forecasting pipeline steps.
HierarchyEnsembleForecaster(forecasters[, ...])	Aggregates hierarchical data, fit forecasters and make predictions.
TransformSelectForecaster(forecasters[, ...])	Compositor that utilizes varying forecasters by time series data's nature.
FhPlexForecaster(forecaster[, fh_params, ...])	Uses different parameters by forecasting horizon element.
IgnoreX(forecaster[, ignore_x])	Compositor for ignoring exogenous variables.
FallbackForecaster(forecasters[, verbose, ...])	Forecaster that sequentially tries a list of forecasting models.

Reduction

Reduction forecasters that use sklearn regressors or sktime time series regressors to make forecasts.

Concurrent tabular strategy

Uses exogeneous data at the same time stamp - simple reduction strategy.

YfromX(estimator[, pooling])

Simple reduction predicting endogeneous from concurrent exogeneous variables.

Direct and recursive - sktime native 1st generation

1st generation direct and recursive reduction forecasters, numpy based.

Different strategies can be constructed via make_reduction for easy specification.

make_reduction(estimator[, strategy, ...])

Make forecaster based on reduction to tabular or time-series regression.

DirectTabularRegressionForecaster(estimator)	Direct reduction from forecasting to tabular regression.
DirectTimeSeriesRegressionForecaster(estimator)	Direct reduction from forecasting to time-series regression.
MultioutputTabularRegressionForecaster(estimator)	Multioutput reduction from forecasting to tabular regression.
MultioutputTimeSeriesRegressionForecaster(...)	Multioutput reduction from forecasting to time series regression.
RecursiveTabularRegressionForecaster(estimator)	Recursive reduction from forecasting to tabular regression.
RecursiveTimeSeriesRegressionForecaster(...)	Recursive reduction from forecasting to time series regression.
DirRecTabularRegressionForecaster(estimator)	Dir-rec reduction from forecasting to tabular regression.
DirRecTimeSeriesRegressionForecaster(estimator)	Dir-rec reduction from forecasting to time-series regression.

Direct and recursive - sktime native 2nd generation

2nd generation rearchitecture of direct and recursive reduction forecasters, pandas based.

DirectReductionForecaster(estimator[, ...])

Direct reduction forecaster, incl single-output, multi-output, exogeneous Dir.

Direct and recursive - 3rd party

SkforecastAutoreg(regressor, lags[, ...])

Adapter for skforecast.ForecasterAutoreg.ForecasterAutoreg class [Rab014607127f-1].

DartsRegressionModel([lags, ...])	Darts Regression Model Estimator.
DartsLinearRegressionModel([...])	Darts LinearRegression Estimator.
DartsXGBModel([past_covariates, ...])	Darts XGBModel Estimator.

Naive forecasters

NaiveForecaster([strategy, window_length, sp])

Forecast based on naive assumptions about past trends continuing.

ForecastKnownValues(y_known[, method, ...])

Forecaster that plays back known or prescribed values as forecasts.

Prediction intervals

Wrappers that add prediction intervals to any forecaster.

SquaringResiduals([forecaster, ...])

Compute the prediction variance based on a separate forecaster.

NaiveVariance(forecaster[, initial_window, ...])

Compute the prediction variance based on a naive strategy.

ConformalIntervals(forecaster[, method, ...])

Empirical and conformal prediction intervals.

BaggingForecaster([bootstrap_transformer, ...])

Forecast a time series by aggregating forecasts from its bootstraps.

EnbPIForecaster([forecaster, ...])

Ensemble Bootstrap Prediction Interval Forecaster.

Trend forecasters

TrendForecaster([regressor])	Trend based forecasts of time series data, regressing values on index.
PolynomialTrendForecaster([regressor, ...])	Forecast time series data with a polynomial trend.
STLForecaster([sp, seasonal, trend, ...])	Implements STLForecaster based on statsmodels.tsa.seasonal.STL implementation.
CurveFitForecaster(function[, ...])	The CurveFitForecaster takes a function and fits it by using scipy curve_fit.
ProphetPiecewiseLinearTrendForecaster([...])	Forecast time series data with a piecewise linear trend, fitted via prophet.

StatsForecastMSTL(season_length[, ...])

StatsForecast Multiple Seasonal-Trend decomposition using LOESS model.

Exponential smoothing based forecasters

ExponentialSmoothing([trend, damped_trend, ...])

Holt-Winters exponential smoothing forecaster.

AutoETS([error, trend, damped_trend, ...])

ETS models with both manual and automatic fitting capabilities.

StatsForecastAutoETS([season_length, model, ...])	StatsForecast Automatic Exponential Smoothing model.
StatsForecastAutoCES([season_length, model])	StatsForecast Complex Exponential Smoothing model.

ThetaForecaster([initial_level, ...])

Theta method for forecasting.

StatsForecastAutoTheta([season_length, ...])

Statsforecast AutoTheta estimator.

AR/MA type forecasters

Forecasters with AR or MA component.

All “ARIMA” and “Auto-ARIMA” models below include SARIMAX capability.

(V)AR(I)MAX models

AutoREG([lags, trend, seasonal, hold_back, ...])

Autoregressive AR-X(p) model.

ARIMA([order, seasonal_order, start_params, ...])	(S)ARIMA(X) forecaster, from pmdarima package.
StatsModelsARIMA([order, seasonal_order, ...])	(S)ARIMA(X) forecaster, from statsmodels, tsa.arima module.

SARIMAX([order, seasonal_order, trend, ...])

(S)ARIMA(X) forecaster, from statsmodels, tsa.statespace module.

VAR([maxlags, method, verbose, trend, ...])

VAR model from statsmodels.

VARReduce([lags, regressor])

A flexible VAR-like forecaster that combines tabularization with regression.

VARMAX([order, trend, error_cov_type, ...])

VARMAX forecasting model from statsmodels.

Auto-ARIMA models

AutoARIMA([start_p, d, start_q, max_p, ...])

Auto-(S)ARIMA(X) forecaster, from pmdarima package.

StatsForecastAutoARIMA([start_p, d, ...])

StatsForecast AutoARIMA estimator.

ARCH models

StatsForecastARCH([p])	StatsForecast ARCH estimator.
StatsForecastGARCH([p, q])	StatsForecast GARCH estimator.
ARCH([mean, lags, vol, p, o, q, power, ...])	Directly interfaces ARCH models from python package arch.

Structural time series models

ARDL([lags, order, fixed, causal, trend, ...])

Autoregressive Distributed Lag (ARDL) Model.

BATS([use_box_cox, box_cox_bounds, ...])

BATS forecaster for time series with multiple seasonality.

TBATS([use_box_cox, box_cox_bounds, ...])

TBATS forecaster for time series with multiple seasonality.

StatsForecastAutoTBATS(seasonal_periods[, ...])

StatsForecast TBATS model.

Prophet([freq, add_seasonality, ...])

Prophet forecaster by wrapping Facebook's prophet algorithm [R995275cbd543-1].

Prophetverse([changepoint_interval, ...])	Univariate prophetverse forecaster - prophet model implemented in numpyro.
HierarchicalProphet([trend, ...])	A Bayesian hierarchical time series forecasting model based on Meta's Prophet.

UnobservedComponents([level, trend, ...])

UnobservedComponents forecasting model from statsmodels.

DynamicFactor([k_factors, factor_order, ...])

Dynamic Factor Forecaster.

Deep learning based forecasters

LTSFLinearForecaster(seq_len, pred_len, *[, ...])	LTSF-Linear Forecaster.
LTSFDLinearForecaster(seq_len, pred_len, *)	LTSF-DLinear Forecaster.
LTSFNLinearForecaster(seq_len, pred_len, *)	LTSF-NLinear Forecaster.
LTSFTransformerForecaster(seq_len, ...[, ...])	LTSF-Transformer Forecaster.

CINNForecaster([n_coupling_layers, ...])

Conditional Invertible Neural Network (cINN) Forecaster.

NeuralForecastRNN([freq, local_scaler_type, ...])	NeuralForecast RNN model.
NeuralForecastLSTM([freq, ...])	NeuralForecast LSTM model.

PytorchForecastingTFT([model_params, ...])	pytorch-forecasting Temporal Fusion Transformer model.
PytorchForecastingDeepAR([model_params, ...])	pytorch-forecasting DeepAR model.
PytorchForecastingNHiTS([model_params, ...])	pytorch-forecasting NHiTS model.
PytorchForecastingNBeats([model_params, ...])	pytorch-forecasting NBeats model.

PyKANForecaster([hidden_layers, ...])

PyKANForecaster uses Kolmogorov Arnold Network [1] to forecast time series data.

Pre-trained and foundation models

HFTransformersForecaster(model_path[, ...])

Forecaster that uses a huggingface model for forecasting.

ChronosForecaster(model_path[, config, ...])

Interface to the Chronos Zero-Shot Forecaster by Amazon Research.

MOIRAIForecaster(checkpoint_path[, ...])

Adapter for using MOIRAI Forecasters.

TimesFMForecaster(context_len, horizon_len)

Implementation of TimesFM (Time Series Foundation Model) for Zero-Shot Forecasting.

TinyTimeMixerForecaster([model_path, ...])

TinyTimeMixer Forecaster for Zero-Shot Forecasting of Multivariate Time Series.

Intermittent time series forecasters

Croston([smoothing])

Croston's method for forecasting intermittent time series.

Ensembles and stacking

EnsembleForecaster(forecasters[, n_jobs, ...])	Ensemble of forecasters.
AutoEnsembleForecaster(forecasters[, ...])	Automatically find best weights for the ensembled forecasters.
StackingForecaster(forecasters[, regressor, ...])	StackingForecaster.

Hierarchical reconciliation

ReconcilerForecaster(forecaster[, method, ...])

Hierarchical reconciliation forecaster.

Online and stream forecasting

OnlineEnsembleForecaster(forecasters[, ...])	Online Updating Ensemble of forecasters.
NormalHedgeEnsemble([n_estimators, a, loss_func])	Parameter free hedging algorithm.
NNLSEnsemble([n_estimators, loss_func])	Ensemble forecasts with Non-negative least squares based weighting.

UpdateEvery(forecaster[, update_interval])	Update only periodically when update is called.
UpdateRefitsEvery(forecaster[, ...])	Refits periodically when update is called.
DontUpdate(forecaster)	Turns off updates, i.e., ensures that forecaster is only fit and never updated.

Adapters to other forecasting framework packages

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

HCrystalBallAdapter(model)

Adapter for using hcrystalball forecasters in sktime.

Model selection and tuning

ForecastingGridSearchCV(forecaster, cv, ...)	Perform grid-search cross-validation to find optimal model parameters.
ForecastingRandomizedSearchCV(forecaster, ...)	Perform randomized-search cross-validation to find optimal model parameters.
ForecastingSkoptSearchCV(forecaster, cv, ...)	Bayesian search over hyperparameters for a forecaster.
ForecastingOptunaSearchCV(forecaster, cv, ...)	Perform Optuna search cross-validation to find optimal model hyperparameters.

Model Evaluation (Backtesting)

evaluate(forecaster, cv, y[, X, strategy, ...])

Evaluate forecaster using timeseries cross-validation.

Time index splitters

Evaluation and tuning can be customized using time index based splitters, for a list of these consult the splitter API