Sktime Forecasting

sktime Forecasting

Use this skill after forecasting-data-prep. sktime is best when you need a scikit-learn-like forecasting interface across many model families, pipelines, hyperparameter search, temporal backtesting, probabilistic forecasts, or panel/hierarchical workflows on single-machine pandas/NumPy-scale data.

Minimum Install

python -m pip install sktime
python -m pip install "sktime[forecasting]"
python -m pip install "sktime[all_extras]"
conda install -c conda-forge sktime

Install estimator-specific soft dependencies only when needed. all_extras is not required for most workflows and may fail on some platforms.

Data Contract

• Start from a forecasting-data-prep contract: sorted time index, no duplicate time keys per series, documented frequency, horizon, cutoffs, known-future covariates, and leakage notes.
• Use y as the endogenous target: pd.Series for one univariate series, pd.DataFrame for uni/multivariate series, or a pd.DataFrame with MultiIndex for panel/hierarchical data.
• For panel/hierarchical data, make the last index level the time-like index; earlier levels identify instances or hierarchy nodes.
• Use X only for exogenous variables. X is a pd.DataFrame; if passed to fit, pass aligned future X to predict unless using a documented composition such as ForecastX.
• Do not assume every forecaster supports multivariate y, panel/hierarchical data, probabilistic output, or X. Inspect tags with sktime.registry.all_estimators and all_tags.
• Use ForecastingHorizon for absolute horizons; use relative integer horizons for regular next-step forecasts.

Model Selection

• Always benchmark with NaiveForecaster or another simple baseline before using complex models.
• Use classical/statistical forecasters for short-to-medium univariate series with trend/seasonality: ThetaForecaster, ExponentialSmoothing, AutoETS, ARIMA, AutoARIMA, SARIMAX, BATS, TBATS.
• Use VAR, VARMAX, VECM, or DynamicFactor only when multiple targets are jointly endogenous and the chosen class supports that shape.
• Use reduction forecasters or make_reduction when you want scikit-learn regressors with lag windows, direct/recursive/dirrec strategies, and optional X.
• Use panel/global/hierarchical workflows when the data has multiple related series; expect automatic vectorization for forecasters that are not genuinely global/hierarchical.
• Use deep, foundation, AutoML, and third-party adapters only after checking soft dependencies, documented input shape, and installed estimator tags.

Read references/sktime-model-map.md before selecting among model families or claiming model support.

Workflow

1. Prepare data with forecasting-data-prep; preserve frequency, horizon, cutoffs, gap, panel keys, and known-future covariate classification.
2. Split by time with temporal_train_test_split for a holdout, or use ExpandingWindowSplitter, SlidingWindowSplitter, or related splitters for backtesting.
3. Build y_train, y_test, optional X_train, X_test; never fit transformations before the temporal split.
4. Select a forecaster by data shape and tags. For transformations, use TransformedTargetForecaster for y and ForecastingPipeline for X.
5. Fit with forecaster.fit(y_train, X=X_train, fh=fh) if the forecaster requires or benefits from fh at fit; otherwise pass fh at predict.
6. Predict point forecasts with predict(fh=fh, X=X_future).
7. For probabilistic forecasts, call documented methods only when supported: predict_interval, predict_quantiles, predict_var, or predict_proba.
8. Evaluate with temporal cutoffs/backtesting and keep the test period for final reporting.

Python Pattern

from sktime.forecasting.base import ForecastingHorizon
from sktime.forecasting.naive import NaiveForecaster
from sktime.performance_metrics.forecasting import mean_absolute_scaled_error
from sktime.split import ExpandingWindowSplitter, temporal_train_test_split
from sktime.utils.plotting import plot_series

y_train, y_test, X_train, X_test = temporal_train_test_split(
    y, X=X, test_size=horizon
)
fh = ForecastingHorizon(y_test.index, is_relative=False)

forecaster = NaiveForecaster(strategy="last", sp=seasonal_period)
forecaster.fit(y_train, X=X_train)
y_pred = forecaster.predict(fh=fh, X=X_test)

score = mean_absolute_scaled_error(y_test, y_pred, y_train=y_train)
plot_series(y_train, y_test, y_pred, labels=["train", "test", "forecast"])

cv = ExpandingWindowSplitter(
    fh=fh.to_relative(cutoff=y_train.index[-1]),
    initial_window=initial_window,
    step_length=step_length,
)

Use evaluate(forecaster, cv, y, X=...) or ForecastingGridSearchCV with a temporal splitter for backtesting and tuning.

Validation, Metrics, and Plotting

• Prefer rolling-origin or expanding-window backtesting over one split for model selection.
• Recommended point metrics: MAE, RMSE, MASE/RMSSE, WAPE or relative loss against a baseline; avoid MAPE/sMAPE when actuals can be zero or near zero.
• Recommended probabilistic metrics: pinball loss for quantiles, empirical coverage and interval width for intervals, CRPS or log loss for distribution forecasts.
• Plot y_train, y_test, y_pred with sktime.utils.plotting.plot_series; pass pred_interval for interval plots.
• Inspect get_fitted_params() and documented wrapped-model diagnostics where available. sktime does not provide one universal residual-diagnostics interface for every forecaster.

Read references/sktime-validation-pipelines.md before building pipelines, tuning, probabilistic evaluation, or panel/hierarchical validation.

Anti-Leakage Rules

• Never random split forecasting rows for validation.
• Fit scalers, imputers, encoders, detrenders, deseasonalizers, feature selectors, and target transforms on train only; use sktime pipelines or refit inside each backtest fold.
• Create lag, rolling, expanding, and window features using only data available before each forecast cutoff.
• Use future X only when the covariate is genuinely known at prediction time for every horizon step; otherwise forecast it separately and account for that uncertainty.
• Respect forecast horizon, gap, frequency, cutoff timestamps, panel instance boundaries, and valid timestamp windows.
• During backtesting, rebuild features and future covariates separately for every cutoff.

Common Errors

• Installing only core sktime and then using a forecaster that needs an uninstalled soft dependency.
• Passing X to fit but omitting future X at predict.
• Treating unrelated independent series as multivariate endogenous variables instead of panel instances.
• Assuming all forecasters support intervals, multivariate y, panel/global fitting, or exogenous variables.
• Passing an absolute horizon whose dates do not match the prepared frequency.
• Tuning on transformed full data instead of wrapping transformations inside TransformedTargetForecaster or ForecastingPipeline.

References

• Read references/sktime-model-map.md for official forecaster categories, supported model names, tags, and selection guidance.
• Read references/sktime-validation-pipelines.md for data formats, X, ForecastingHorizon, temporal validation, metrics, probabilistic forecasts, plotting, and diagnostics.
• Read references/official-sources.md for official sources consulted.

Ready Checklist

• forecasting-data-prep contract is complete and leakage risks are resolved or documented.
• y, optional X, frequency, horizon, panel/hierarchy index levels, and future covariate availability are explicit.
• Chosen forecaster supports the required y scitype, X, probabilistic output, and panel/global behavior according to tags or API docs.
• Train/validation/test or backtest folds use temporal cutoffs, not random splits.
• Transformations and model selection are fit inside the training data of each fold.
• Forecasts are compared with a simple baseline using horizon-aware metrics and plots.