Topic 14 of 20 · Part IV : Data and Learning Operations

Data quality

A sophisticated model cannot reliably repair a pipeline that confuses millions with billions or maps a ticker to the wrong legal entity.

The control pattern

Data asset

Source records and canonical data models

Knobs

Completeness, Validity, Consistency, Identity resolution, Units and adjustments, Lineage and deduplication

Outcome

Dependable downstream analysis

Measurement

See how to measure this below

What it is

Data quality is the foundation beneath every model and score. Source records are the data asset; schema rules and validation properties are the knobs. Completeness asks whether required fields exist, validity checks permitted values, consistency compares related records, and deduplication prevents the same event appearing multiple times.

Financial systems also need careful identity resolution, unit normalisation, currency handling and treatment of corporate actions. A sophisticated model cannot reliably repair a pipeline that confuses millions with billions or maps a ticker to the wrong legal entity.

In practice, improving these knobs often produces larger gains than switching to a more capable language model.

The knobs in detail

Each row is one adjustable property of the data asset, and what moving it tends to do.

KnobWhat you adjustLikely effect
CompletenessWhether required fields existDetermines what can be computed at all
ValidityWhether values are permittedCatches malformed records at the door
ConsistencyWhether related records agreeSurfaces reconciliation failures
Identity resolutionMapping symbols to legal entitiesPrevents the wrong company entirely
Units and adjustmentsCurrency, scale, splits, restatementsStops silent order-of-magnitude errors
Lineage and deduplicationWhere a value came from, and how oftenMakes every score traceable

Applied: Stocks Assistant

For Stocks Assistant, create canonical entities for company, security, filing, fiscal period, market observation, analyst estimate, transcript segment, options contract and portfolio position. Preserve both reported and normalised values along with the transformation used. Distinguish adjusted and unadjusted prices; record splits, dividends, symbol changes and restatements; and use nulls rather than silently converting missing data to zero. Each derived score should be traceable to source records and a versioned transformation.

How to measure it

Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.

  • Feed latency and stale-record counts
  • Missing fields and duplicate rate
  • Reconciliation differences between related sources
  • Unit anomalies and lineage gaps
  • Quality gates passed before data enters RAG indexes, training, scoring or reports

Common mistakes

Serving plausible-looking values from an unknown state when a source fails. Degrade visibly and selectively instead.

Converting missing data to zero, which turns absence into a false fact.