Skip to content

Feature Availability

Not all features can be used in real-time policy rules. Some require aggregation over historical windows and must be computed by a feature store.

Real-time features

These fields are available in every live API request and can be referenced in data/active_policy.json rules.

Feature Type Validation
transaction_id string Required, non-empty
tx_type string Required
amount float Required, > 0.0
geo_velocity float Required, 0–5000 km/h
typing_entropy float Required, 0.0–6.0
device_is_emulator bool Optional, defaults to false

Pydantic enforces range constraints at the API boundary (src/api/router.py). Requests violating these constraints return HTTP 422 before evaluation begins.

Batch-only features

These signals require historical aggregation and are not available in the synchronous request payload. Rules referencing them will log a warning and be silently skipped at runtime.

Feature Description Requires
velocity_7d Number of transactions in the last 7 days Feature store / time-series aggregation
consortium_match_count Cross-client device/behavior matches External device graph integration
account_age_days Days since account creation Identity database query
chargeback_rate_30d Rolling chargeback rate for this user Aggregated label history

What happens when a rule references a missing field

If a rule's condition references a field not present in the payload:

  1. evaluate_policy() catches the KeyError
  2. Logs: WARNING: Missing field in payload during rule evaluation: <field> | rule=<rule>
  3. Skips that rule
  4. Continues evaluating remaining rules
  5. The request never fails due to a missing feature

This design allows risk managers to author rules for future signals — the rule will activate automatically once the caller starts sending that field.

Adding a new real-time feature

  1. Add the field to RiskPayload in src/api/router.py with appropriate Pydantic constraints
  2. Verify the field is passed through to evaluate_policy() in the data dict
  3. If using it for ML scoring, add it to FEATURE_COLS in src/models/train.py and retrain
  4. Reference the field in data/active_policy.json rules