Data Quality Dimension

The degree to which data correctly describes the real-world object or event it represents. In financial fraud, this means a transaction amount, timestamp, and merchant identifier must reflect the objective truth.

• Syntactic Accuracy: The value is in the correct format (e.g., a valid ISO 8601 date).
• Semantic Accuracy: The value is true to the real-world entity (e.g., the merchant category code actually matches the business type).
• Measurement: Often requires reconciliation against a verified source of truth, such as a general ledger or a central bank reference table.

The proportion of required data that is actually present in a dataset. A record is incomplete if a mandatory attribute, such as a transaction's originating IP address or device fingerprint, is null or missing.

• Record Completeness: The percentage of expected records that exist (e.g., no missing transaction logs for a given hour).
• Attribute Completeness: The percentage of non-null values for a critical column (e.g., 99.99% of transactions have a valid BIN number).
• Impact: Incomplete features force models to rely on default imputation, which can mask genuine fraud signals.

The absence of logical contradiction between two or more data elements. A transaction is inconsistent if its status is 'SETTLED' but its settlement timestamp is null, or if a single account holder has conflicting KYC statuses across different system silos.

• Cross-Field Consistency: Validating that related fields obey business rules (e.g., refund amount ≤ original purchase amount).
• Cross-Record Consistency: Ensuring a unique entity, like a customer ID, has identical core attributes across all databases.
• Violation: Inconsistencies break feature engineering logic, creating null or infinite values in derived ratios used for anomaly detection.

The degree to which data represents reality at the required point in time. For real-time fraud scoring, this is measured as the latency between the event occurrence and its availability for model inference.

• Freshness: The age of the data. A velocity check is useless if the transaction count feature is 15 minutes stale.
• Temporal Consistency: The chronological ordering of events must be preserved. A login event timestamped after a transaction it supposedly authorized indicates a data pipeline error.
• Volatility: The period of time during which data remains valid. A real-time risk score has a volatility measured in milliseconds.

The conformity of data to its defined format, type, range, and business rules. Validity is a syntactic gate that catches corrupted or malformed records before they enter a model.

• Format Validity: A transaction amount must be a decimal, not a string. A SWIFT code must match the 8 or 11 character pattern.
• Range Validity: A transaction amount cannot be negative for a purchase, and an account age cannot be a future date.
• Domain Validity: A country code must exist in the ISO 3166-1 reference set. Invalid categorical values cause one-hot encoding failures.

The requirement that no entity or event is represented more than once in a dataset. Duplicate transaction records artificially inflate monetary velocity features and can cause a single legitimate purchase to trigger a fraud alert.

• Entity Uniqueness: A single customer master record must map to one real-world individual. Duplicate identities are a primary vector for synthetic identity fraud.
• Event Uniqueness: A transaction must have a single, unique identifier. Idempotency keys in payment gateways prevent network retries from creating duplicate financial events.
• Measurement: Deduplication is performed by hashing a composite key of deterministic attributes.

The systematic process of statistically examining a dataset's structure, content, and interrelationships to discover metadata and assess its alignment with specific quality dimensions. Profiling generates summary statistics—null counts, cardinality, value distributions, and data type conformance—that serve as the empirical foundation for a data quality assessment. In fraud analytics, profiling reveals whether a transaction feed's merchant_category_code field is 40% null, directly quantifying the completeness dimension.

The end-to-end mapping of data's origin, transformations, and movement across systems. Lineage tracking is essential for diagnosing the root cause of a quality dimension failure. If a model's input feature avg_txn_7d suddenly exhibits a distribution shift, lineage allows engineers to trace backward through the ETL pipeline to identify whether the break occurred at the source system, during a join operation, or within an aggregation step. This capability directly supports the auditability requirements of SR 11-7.

A symmetric metric quantifying the shift in a variable's distribution between a development sample and a production sample. PSI is the primary leading indicator for the consistency and stability dimensions of data quality. A PSI value exceeding 0.25 signals a significant shift requiring investigation. For a fraud model, a high PSI on transaction_amount may indicate a new merchant segment entering the portfolio, necessitating a model recalibration or a data quality intervention at the ingestion layer.

A formal, measurable contract between data producers and data consumers that defines acceptable thresholds for each quality dimension. An SLA transforms abstract dimensions into operational targets. For example: 'The timeliness of the real-time transaction stream shall not exceed 500ms end-to-end latency for 99.9% of events, and the completeness of mandatory fields shall be 100%.' Automated monitoring dashboards track SLA compliance, triggering alerts when a dimension breaches its defined threshold.

An architectural capability that applies telemetry principles to data pipelines, providing continuous, automated visibility into the five pillars of data health: freshness, distribution, volume, schema, and lineage. While a data quality dimension defines what to measure, data observability provides the real-time instrumentation to detect anomalies. It surfaces silent failures—such as a gradual drift in the currency_code field's cardinality—before they corrupt downstream fraud model scoring.

A change in the underlying statistical relationship between model inputs and the target variable. While distinct from a data quality issue, concept drift is often detected through the same monitoring infrastructure. A sudden degradation in a model's precision metric may be misdiagnosed as a data pipeline failure when the root cause is a genuine shift in fraudster behavior. Distinguishing between a data quality dimension breach and true concept drift is a core competency of model risk management teams.