sktime Cheat Sheet

sktime cheat sheet

sktime is the scikit-learn of time series: one consistent fit/predict API across the whole time-series lifecycle — forecasting, classification, regression, clustering, annotation/anomaly detection — plus pipelines, tuning and reduction to tabular ML. It wraps statsmodels, StatsForecast, Prophet, Darts, ARIMA and deep-learning backends behind a shared interface. This sheet targets sktime 0.3x and focuses on forecasting.

core forecasting forecasters & intervals pipelines & reduction evaluation & tuning other tasks gotcha ★ most common

Verified 2026-08-24 against the official docs at sktime.net (API reference + forecasting tutorial). scikit-learn-compatible; many estimators are adapters over other libraries (install their soft dependencies as needed).

Outline

The forecasting workflow is forecaster.fit(y, fh=...).predict(). Everything else — pipelines, tuning, other tasks — reuses that same estimator pattern.

1Install & data format0.3x

pip install "sktime[all_extras]"★
Base install is light; extras pull in soft deps (statsmodels, pmdarima, prophet, etc.). Install only what you use in prod.
y = pd.Series(values, index=pd.PeriodIndex(..., freq="M"))★
A univariate series with a proper time index (PeriodIndex preferred; DatetimeIndex/RangeIndex also work).
# panel/multi-series: pd.DataFrame with a MultiIndex (instance, time)
Hierarchical/multiple series use a MultiIndex; the same forecasters vectorize over instances.
from sktime.datasets import load_airline y = load_airline()
Built-in datasets for experiments & docs examples.

2fit / predict / fhthe workflow

from sktime.forecasting.naive import NaiveForecaster f = NaiveForecaster(strategy="last") f.fit(y, fh=[1,2,3]); f.predict()★
Fit, then predict the forecasting horizon. Always establish a naive baseline first.
from sktime.forecasting.base import ForecastingHorizon fh = ForecastingHorizon([1,2,3], is_relative=True)★
fh can be ints (steps ahead) or an absolute index. Pass it to fit or predict.
f.fit(y, X=X_train, fh=fh); f.predict(X=X_future)
Exogenous features via X — you must supply future X at predict time.
f.update(y_new) · f.predict()
Stream new observations without a full refit (where the forecaster supports it).

3Forecaster zoopick a model

from sktime.forecasting.arima import AutoARIMA★
Auto-tuned (S)ARIMA. Also StatsForecastAutoARIMA (the fast Nixtla backend).
from sktime.forecasting.exp_smoothing import ExponentialSmoothing from sktime.forecasting.ets import AutoETS
Holt-Winters / ETS state-space models. ThetaForecaster is a strong, simple benchmark.
from sktime.forecasting.trend import PolynomialTrendForecaster from sktime.forecasting.fbprophet import Prophet
Trend models and adapters to Prophet/Darts/statsforecast — all behind the same API.
f.get_params() · f.get_fitted_params()
Inspect hyperparameters and fitted coefficients (sklearn-style).

4Probabilistic forecastsuncertainty

f.predict_interval(coverage=0.9)★
Prediction intervals at a coverage level — returns lower/upper columns. Not every forecaster supports it.
f.predict_quantiles(alpha=[0.1, 0.5, 0.9])
Arbitrary quantiles; predict_var() for variance, predict_proba() for a full distribution.
f.get_tag("capability:pred_int")
Check a forecaster's capability tags before calling interval methods.

!Common gotchasread before shipping

never shuffle / KFold a time series
Use temporal_train_test_split and the window splitters — random splits leak the future into training.
index must carry a frequency
A PeriodIndex/DatetimeIndex with a set freq avoids ambiguous-horizon errors; irregular gaps confuse many models.
exogenous X needs the future
If you fit with X, you must provide X for the forecast horizon at predict time — you can't forecast without it.
soft dependencies
Many estimators wrap other libraries; a missing one raises at import/fit. Install the specific extra, not always [all_extras].
not every forecaster does intervals
Check get_tag("capability:pred_int") before calling predict_interval.

Worth memorizing

scikit-learn for time seriesone fit/predict API across forecasting, classification, detection

fit(y, fh=...).predict()fh = ForecastingHorizon (steps ahead or an absolute index)

always beat NaiveForecasterestablish the naive baseline before anything fancy

make_reductionturns any sklearn regressor into a windowed forecaster

TransformedTargetForecasterdetrend/deseasonalize the target and auto-invert

window splitters, never KFoldExpandingWindowSplitter + evaluate() for honest backtests

predict_interval / predict_quantilesuncertainty when the forecaster supports it

all_estimators + tagsdiscover models and filter by capability

exogenous X needs future valuessupply X for the horizon at predict time

sktime cheat sheet

Outline

Forecast

Compose

Evaluate

Beyond forecasting

Forecast

Compose

Evaluate & Tune

Beyond Forecasting

Gotchas

Worth memorizing