---

name: transforms-features description: "Use GluonTS transformations, time features, samplers, splitters, and missing-value indicators for training and prediction data preparation." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

transforms-features

Use this sub-skill when a task needs to prepare GluonTS DataEntry streams for model training or prediction with transformation chains, observed-value indicators, time features, lag selection, or instance sampling/splitting.

Route here for

• Building gluonts.transform.Chain pipelines that operate on iterables of dictionaries.
• Adding missing-value observed indicators and deterministic target imputations.
• Generating calendar/time features from a frequency string and choosing model lag sequences.
• Creating training or prediction instances with InstanceSampler and InstanceSplitter.
• Diagnosing shape, length, frequency, or is_train mistakes in transformed datasets.

Do not use this sub-skill for

• Constructing the original dataset from pandas/files/JSON Lines: use the data-pipelines sub-skill first.
• Choosing, training, serializing, or serving estimators/predictors: use the forecasting-models or deployment-extensions sub-skill.
• MXNet-specific transformation recipes: those legacy workflows were not selected as verified required coverage.

Read map

1. Start with references/api-reference.md for API contracts, constructor signatures, field names, and shape conventions.
2. Use references/workflows.md for train/prediction transformation recipes and assertion checklists.
3. Use references/troubleshooting.md when a chain yields zero instances, wrong feature lengths, unsupported frequencies, or NaN/shape failures.
4. Run scripts/transform_feature_smoke.py --help and then scripts/transform_feature_smoke.py to check that the installed gluonts package can execute the core transform/time-feature path without the source checkout.

Minimal operating pattern

• Convert raw data to a GluonTS dataset first, so entries contain start as a period-like value with a frequency and target as a numeric array.
• Build a Chain in feature order: target conversion/observed indicator, time/age/constant features, feature stacking if needed, then an instance splitter if the model expects windows.
• Pass is_train=True when creating training windows with available future target values; pass is_train=False for prediction-time windows and ensure known-future features have len(target) + prediction_length values.
• Assert field presence and shapes after transformation before handing the result to an estimator or predictor.