---

name: data-processing-and-covariates description: "Use Darts preprocessing transformers, Pipeline, missing-value filling, scaling, inverse transforms, and covariate span checks." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Data processing and covariates

Use this sub-skill when the user needs to clean/transform Darts TimeSeries, fit preprocessing only on training data, inverse-transform forecasts, or build/validate past and future covariates.

Read first

• references/workflows.md for train-only preprocessing, generated covariates, stacking, and span validation patterns.
• references/api-reference.md for verified constructor signatures and common transformer/covariate APIs.
• references/troubleshooting.md for leakage, inverse-transform, missing-value, covariate frequency, and horizon coverage errors.
• scripts/transform_pipeline_smoke.py for a tiny generated pipeline/covariate smoke.

Route by task

• Fill missing target values: use MissingValuesFiller, usually inside a Pipeline.
• Scale data: fit Scaler or a Pipeline on train only, then transform validation/test; never fit on held-out data.
• Invert forecast scale: keep the fitted transformer and call inverse_transform() on the forecast with matching components.
• Generate calendar covariates: use Darts time-series generation helpers such as datetime/day-of-week attributes or caller-provided covariate tables, then stack components.
• Fix covariate span errors: validate target end time, forecast horizon, covariate start/end, frequency, and one covariate series per target series.
• Choose a model that consumes covariates: route to ../forecasting-workflows/ for core/regression models or ../torch-and-foundation-models/ for neural models.

Safe check

python scripts/transform_pipeline_smoke.py --quiet

The script generates a small daily series, fills/scales train data, transforms validation data, inverse-transforms a toy forecast, stacks two future covariates, and asserts horizon coverage.

Boundaries

This sub-skill prepares data and covariates. It does not decide all model-family details or interpret final metrics. Do not solve future covariate span errors by extending the target series with fake values; extend/regenerate the covariate series instead.