Feature Engineering & Offline/Online Parity¶

Feature design principles¶

• features must be computable before match start,
• features must be stable across environments,
• feature definitions are deterministic and versioned.

Feature categories¶

Examples include: - team historical performance, - recent form statistics, - head-to-head aggregates, - contextual signals (home/away, rest days).

Offline feature pipeline¶

Offline features are: - computed in DVC pipelines, - stored as versioned feature tables, - validated via data contracts.

Online feature parity¶

The system enforces train/serve parity: - feature logic is shared between offline and online code paths, - no ad-hoc transformations at inference time.

If parity cannot be guaranteed, the feature is excluded from the model.

Anti-patterns avoided¶

• manual feature hacks in notebooks,
• inference-only features,
• implicit feature ordering.