---

name: statsforecast description: "Use Nixtla StatsForecast for statistical time-series forecasting, model selection, exogenous regressors, prediction intervals, cross-validation, feature engineering, and optional distributed execution." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

StatsForecast repo skill

Use this skill when the task is about the Python package statsforecast: fast statistical forecasting over many time series, ARIMA/ETS/CES/Theta/Croston/baseline models, exogenous regressors, prediction intervals, temporal cross-validation, MSTL feature workflows, or local/Dask/Ray/Spark-style distributed execution.

First checks

For a new runtime, install and verify the package before writing forecasting code:

python -m pip install statsforecast
python scripts/check_statsforecast_env.py --json
python scripts/statsforecast_quick_smoke.py --json

If the package comes from a source checkout, read installation and environment before editable installs because the compiled extension needs Eigen headers and a working C++ build path.

Route by user task

User asks for	Read
A normal pandas/polars forecast, forecast, fit/predict, fit_predict, fitted values, intervals, X_df, custom columns, cross-validation, persistence, plotting, or fallback behavior	core-forecasting
Which model class to use, constructor knobs, direct model.fit(y) / model.forecast(y, h) calls, aliases, AutoARIMA/ETS/CES/Theta, Croston, MSTL/MFLES/TBATS, GARCH/ARCH, or SklearnModel	model-selection
Synthetic panels, AirPassengers fixtures, static/exogenous features, mstl_decomposition, trend/seasonal future regressors, pandas-vs-polars feature validation	feature-engineering
n_jobs, ParallelBackend, MultiprocessBackend, Fugue, Dask, Ray, Spark, distributed materialization, local parallel parity, or optional backend dependency issues	distributed-execution
Import/build failures, optional dependencies, cross-skill error triage	troubleshooting and installation and environment
Staleness or whether this skill matches the source evidence	repo provenance

Core package facts

• Main import: from statsforecast import StatsForecast.
• Model import surface: from statsforecast.models import Naive, SeasonalNaive, AutoARIMA, ....
• Built-in toy data: statsforecast.utils.AirPassengersDF, AirPassengers, and generate_series.
• Standard panel schema: one row per observation with id, time, target columns (unique_id, ds, y by default) plus optional exogenous/static columns.
• Future exogenous dataframe: X_df must provide one future row per id and forecast horizon for models that use time-varying regressors.
• Fast one-shot forecasting: StatsForecast(...).forecast(df=panel_df, h=h, ...).
• Stateful workflow: fit(df=...), then predict(h=..., X_df=...), then optional forecast_fitted_values().
• Backtesting: cross_validation(df=..., h=..., n_windows=..., step_size=...).
• Intervals: pass level=[80, 95]; use ConformalIntervals when conformal intervals are needed and enough history exists.

Minimal examples

One-shot forecast

from statsforecast import StatsForecast
from statsforecast.models import AutoARIMA, SeasonalNaive
from statsforecast.utils import AirPassengersDF

sf = StatsForecast(
    models=[AutoARIMA(season_length=12), SeasonalNaive(season_length=12)],
    freq="ME",
    n_jobs=1,
)
forecast_df = sf.forecast(df=AirPassengersDF, h=12, level=[80, 95])

Cross-validation

from statsforecast import StatsForecast
from statsforecast.models import Naive
from statsforecast.utils import generate_series

panel = generate_series(n_series=5, freq="D", equal_ends=True)
sf = StatsForecast(models=[Naive()], freq="D")
cv_df = sf.cross_validation(df=panel, h=7, n_windows=3, step_size=7)

Bundled scripts

• scripts/check_statsforecast_env.py: check installed package, core modules, and optional backend imports without modifying data.
• scripts/statsforecast_quick_smoke.py: run a tiny deterministic package-level forecast and cross-validation smoke.
• Sub-skills also provide focused smoke/catalog scripts for core forecasting, model catalogs, MSTL features, and distributed execution.

Decision rules

• Start with core-forecasting for most user code, then route to model-selection only when choosing or configuring model classes becomes the blocker.
• Use forecast for memory-efficient production predictions; use fit/predict when later reuse, persistence, fitted values, or iterative inspection is needed.
• Keep model aliases unique whenever two models would share the same display name.
• Treat Dask/Ray/Spark/Prophet/sklearn/polars as dependency-gated surfaces. Do not claim an optional backend is available until the environment check or import proves it.
• Do not rely on original repository notebooks, tests, or scripts at runtime. Use the bundled references and scripts in this skill tree.