8 patterns for this goal
Data quality failures in financial systems occur when agents ingest, cleanse, or propagate data containing misattributed entities, stale hierarchies, fabricated values, or unresolved ambiguities—and pass that corrupted data downstream to pricing, risk, or compliance calculations without flagging the corruption. Most data quality failures are silent: a misread field, a merged entity, or a cached value presents itself as clean data until a downstream reconciliation or a compliance audit discovers the error months later. The core failure is structural—multi-stage cleansing and consuming agents operate on fixed schemas that drop critical provenance and confidence signals at handoff boundaries, leaving downstream systems blind to data-quality warnings that upstream systems actually surfaced.
All 5 data-quality patterns share a common root mechanism: data flows through a pipeline where intermediate agents surface findings in free text (confidence notes, discrepancy flags, imputation methods, temporal concerns) but the structured schemas passed between agents omit fields to carry those findings downstream. The fix is architectural: add structured fields for cleansing confidence, hierarchy versioning, imputation method, and point-in-time validity to every handoff schema, backed by mandatory gating that any data failing these checks is flagged for human review before use. Entity resolution via embedding similarity is fixable by indexing on unique identifiers first, then similarity only on the already-matched cohort.
Only if the embedding is explicitly weighted toward unique-identifier fields (LEI, CUSIP, ID) and the retrieval is constrained to a set of entities that have already been confirmed to share a common parent or identifier. Name-based embedding matching across a full population of potentially unrelated entities will continue to surface lexically similar but genuinely distinct entities. Unique-identifier matching first, similarity only within the already-matched set, is the only reliable approach.
Missing-data mishandling is local to a single time series: how to fill a gap when no data point exists (mean imputation, forward-fill, deletion). Point-in-time violations are forward-looking: using data that exists but was not yet known at the decision timestamp (restatements, later-released earnings, retroactive index changes). Both corrupt backtests, but they require different fixes — missing-data needs domain-aware imputation; point-in-time needs temporal labeling and versioning.
Require every handoff schema to include a required data_quality_flags field (an array of issues the upstream agent surfaces), with a mandatory post-handoff validation that any flag in that field triggers downstream review or gating rather than allowing downstream agents to proceed with unreviewed data. Log the field’s contents alongside every calculation, so audit can trace missing flags back to the handoff boundary.
Embedding similarity is dominated by shared vocabulary (industry terms, company-name roots), and regional naming conventions (e.g., “Investment Holdings” + region name repeated across unrelated entities) produce high embedding overlap despite distinct identifiers and no corporate relationship. Embedding is a signal for topical relevance, not identity confirmation. Use embedding only after confirming identifier match or documented corporate relationship.
| Pattern | Mechanism |
|---|---|
| Corporate Hierarchy Misattribution | Risk/exposure attributed to legal entity instead of ultimate parent; hierarchy mapping stale post-M&A |
| Embedding Retrieval Merges Similarly-Named Issuers | Two distinct issuers merged by name similarity without identifier confirmation; exposure concentrates incorrectly |
| Missing Data Mishandling | Mean imputation reduces volatility; forward-fill introduces lookahead bias; deletion creates survivor bias |
| Multi-Agent Handoff Drops Data-Quality Flag | Cleansing confidence/method in free text not carried to downstream agent’s structured input; field treated as reliable |
| Point-in-Time Data Violations | Backtest uses restated/later-released data unavailable at decision time; inflates backtest performance |
Total: 5 patterns
When a Security's ISIN/CUSIP Is Absent or Malformed in the Reference-Data Feed, an Agent Tasked With Enriching Trade or Position Records Generates a Plausible-Looking Identifier by Pattern-Completing From the Security's Name and Exchange, Instead of Flagging the Record as Unresolved
When an Agent Is Used to Canonicalize Free-Text Counterparty or Issuer Names Into a Standard Form, Its Generative Rewriting Silently Drops or Alters a Legal-Entity Suffix (Ltd, LLC, GmbH, Inc, N.V.) That Materially Changes Which Legal Entity — and Therefore Which Counterparty-Risk Profile — the Record Refers To
An Agent Computing Net Asset Value Through a Multi-Hop Sequence of Tool Calls Re-Types the Intermediate Numeric Result as Text Between Steps Instead of Passing the Exact Value Forward, Introducing Rounding and Truncation Drift That Compounds Across the Chain
Agent Attributes Risk, Exposure, or Performance Data to the Wrong Entity Within a Corporate Family Due to Incomplete Parent-Subsidiary Mapping
A Data-Quality Agent Deduplicating Issuer Records Across Multiple Source Feeds Using Embedding Similarity Over Issuer Names, Rather Than Matching on a Unique Identifier Such as LEI or CUSIP Issuer Code, Merges Two Distinct Issuer Entities With Coincidentally Similar Names Into a Single Record, Corrupting Downstream Holdings and Exposure Calculations
Model Handles Missing Data Incorrectly (Mean Imputation, Deletion); Introduces Bias or Information Loss
A Data-Cleansing Agent Notes in Free Text That a Field It Cleaned Was Ambiguous or Low-Confidence, but the Structured Cleansed Record Handed Off to a Downstream Risk- or Pricing-Consuming Agent Has No Field for Cleansing Confidence, So the Downstream Agent Treats the Value as Fully Reliable
Backtests and Live Recommendations Use Data That Was Not Actually Available at the Decision Timestamp (Look-Ahead via Restated/Revised Data)