Feature Engineering Specification Template

Feature engineering creates new features from raw data. Feature selection chooses which features to include in a model from the available set. This template covers engineering. After building features, use techniques like correlation analysis, mutual information, or model-based importance scores to select which features to keep.

There is no universal answer. Start with 10-20 well-reasoned features and add more only if evaluation metrics improve. More features increase training time, serving latency, and maintenance burden. Each feature you add is a dependency you must maintain. The [AI PM Handbook](/ai-guide) covers feature selection trade-offs in its model development chapters.

Use point-in-time joins. When computing features for a training example dated January 15, only use data available before January 15. Never include the target variable or its proxies as features. Document the temporal boundary for each feature in the Aggregation Window column. The [AI Eval Scorecard](/tools/ai-eval-scorecard) includes leakage checks in its evaluation criteria.

It depends on the model. Tree-based models (XGBoost, Random Forest) do not require normalization. Neural networks and linear models benefit from standardization (zero mean, unit variance). Document the expected distribution in the Validation Criteria section so the normalization strategy is explicit.

Review quarterly or whenever model performance degrades. Features based on business logic (like plan_tier categories) must be updated when the business changes. Features based on behavioral data should be monitored for [model drift](/glossary/model-drift). If a feature's distribution shifts significantly, investigate whether the underlying user behavior changed or the data pipeline broke. ---