---

name: unit8co-darts description: "Use Darts for Python time-series construction, preprocessing, forecasting, PyTorch/foundation models, anomaly detection, metrics, and explainability." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Darts

Use this skill when a task names Darts, darts, u8darts, Unit8 Darts, or asks for Darts-specific time-series forecasting, anomaly detection, preprocessing/covariate, metric, or explainability workflows.

Install and import check

Darts 0.46.1 supports Python 3.10+. Choose the smallest install that matches the user's workflow:

pip install darts                    # core TimeSeries, preprocessing, statistical/regression models, anomaly, metrics
pip install "darts[torch]"           # add PyTorch neural/foundation wrapper infrastructure
pip install "darts[notorch]"         # add Prophet/LightGBM/XGBoost/CatBoost/StatsForecast families
pip install "darts[all]"             # broad install only when the user truly needs both torch and notorch extras
python - <<'PY'
import darts
from darts import TimeSeries
print(darts.__version__)
print(TimeSeries)
PY

For a quick environment diagnosis, run scripts/darts_doctor.py. For a tiny core forecast check without external data, run scripts/core_forecasting_smoke.py.

Route by task

• Build or inspect time-series data: use time-series-and-data for TimeSeries.from_dataframe(), from_group_dataframe(), date/frequency gaps, multivariate/stochastic shape (time, component, sample), static covariates, slicing, stacking, and export checks.
• Clean, scale, transform, or build covariates: use data-processing-and-covariates for MissingValuesFiller, Scaler, Pipeline, train-only fitting, inverse transforms, generated calendar covariates, and past/future covariate span validation.
• Choose and run non-neural forecasting models: use forecasting-workflows for naive/statistical/regression/global models, fit()/predict(), covariate-capable core models, probabilistic prediction, historical forecasts, and model-selection troubleshooting.
• Use PyTorch or foundation wrappers: use torch-and-foundation-models for darts[torch], TCNModel/N-BEATS/TFT-style models, chunk lengths, trainer kwargs, checkpoint directories, CPU/GPU verification, and foundation wrapper cache/download planning.
• Score or detect anomalies: use anomaly-detection for KMeansScorer, PyOD scorers, QuantileDetector, threshold/IQR detectors, aggregators, ForecastingAnomalyModel, and score-vs-binary-output semantics.
• Evaluate forecasts or explain models: use evaluation-and-explainability for deterministic metrics, probabilistic quantile/interval metrics, reductions, anomaly/classification metrics, ShapExplainer, and headless plotting boundaries.

Common starting routes

• New pandas data → time-series-and-data, then data-processing-and-covariates if values must be filled/scaled or covariates generated.
• Baseline forecast → forecasting-workflows first; add evaluation-and-explainability for metric/reporting details.
• Neural forecast request → verify darts[torch] with torch-and-foundation-models; do not claim CUDA/foundation execution without explicit backend/cache evidence.
• Anomaly detection request → anomaly-detection; route any residual forecasting model setup back to forecasting or torch.
• Installation/import or optional dependency errors → root references/installation-and-optional-dependencies.md and references/troubleshooting.md.

Shared references

• references/model-catalog.md maps model families to install extras and sub-skill owners.
• references/installation-and-optional-dependencies.md records install variants, optional dependency boundaries, and backend evidence.
• references/troubleshooting.md covers cross-cutting install/import, optional dependency, data, plotting, and backend failures.
• references/repo-provenance.md records the source commit and evidence baseline for refresh decisions.

Boundaries

This is a runtime package-use skill, not a Darts maintainer/release workflow. It does not cover publishing packages, building Docker images, editing CI, running long benchmarks, downloading private datasets, or executing large notebooks. Optional CUDA/GPU/TPU, foundation weight downloads, and heavy model-family extras are documented but not verified by the baseline CPU skill unless a later task proves them in the target environment.