---

name: "data-and-exogenous" description: "Routes NeuralForecast panel schema, exogenous-variable, categorical, sample-weight, and scaler workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

data-and-exogenous

Use this sub-skill when the user is preparing or debugging the panel dataframe that NeuralForecast consumes: unique_id, ds, y, static features, future features, categorical variables, sample weights, availability masks, and local scalers.

What this route covers

• Panel dataframe layout and TimeSeriesDataset.from_df conversion.
• static_df, futr_df, available_mask, and sample_weight handling.
• Categorical feature declarations and cardinality checks.
• Local scaling for temporal and static features.
• pandas and Polars interoperability.

What this route does not cover

• The forecast model choice itself.
• Auto* search or distributed Spark execution.
• Probabilistic loss construction.
• Save/load or serialization details except where they depend on the panel shape.

If the request is about which model family needs which feature set, open ../model-selection/SKILL.md. If the request is about fit / predict behavior after the data is already correct, open ../core-forecasting/SKILL.md.

Read these bundled references

• ../../references/data-formats.md for the long-format panel contract.
• ../../references/api-reference.md for TimeSeriesDataset and data-module signatures.
• references/troubleshooting.md for schema, null, categorical, and scaler failures.

Run this bundled script

• ../../scripts/validate_panel.py for a safe dataframe-schema check.

Common triggers

• "What columns does NeuralForecast need?"
• "My future dataframe is missing rows"
• "How do I use static exogenous variables?"
• "Why is a categorical column failing?"
• "Why do I get a scaler or mask error?"
• "Can I use Polars with NeuralForecast?"

Minimal route

1. Read ../../references/data-formats.md to confirm the dataframe contract.
2. Use ../../scripts/validate_panel.py on a small sample or problematic file.
3. Read references/troubleshooting.md when the dataframe fails validation.

Decision cues

• If the issue is model-specific exogenous support, route to model-selection.
• If the issue is a downstream fit or predict error after the data is valid, route to core-forecasting.

Safe default

When in doubt, validate the panel first. Most NeuralForecast errors become obvious once the unique_id / ds / y contract is checked explicitly.