Post-trade breaks: same-day resolution instead of a standing backlog

The problem

A mid-size asset manager reconciled positions and transactions daily across custodians, prime brokers, and its internal book. Breaks — where the records disagree — arrived every morning in the hundreds. Most were mundane: a fee booked on a different date, a corporate action applied differently, a settlement timing difference, a currency rounding convention. A minority were real errors with real money attached.

Operations analysts worked the list from the top, which meant genuine problems could sit for days behind trivial timing differences, and the backlog never reached zero. Aged breaks are an audit finding, a capital cost, and occasionally a loss.

The hypothesis

Break resolution is pattern recognition followed by evidence assembly. The same twenty or so causes explain the overwhelming majority of breaks, and they are recognizable from the surrounding data. Classify them, resolve the deterministic ones automatically, and assemble the evidence for the rest so an analyst starts with a case file rather than a discrepancy.

The build

• Multi-source normalization — custodian, broker, and internal records mapped to a common transaction and position representation, with the mapping treated as versioned, testable code rather than spreadsheet lore.
• Tolerance and timing rules first — deterministic matching with configurable tolerances handles known conventions: settlement lags, fee timing, rounding, and standard corporate action treatments. Ordinary code, owned by operations.
• Cause classification for the remainder — a model trained on historical breaks and their resolutions proposes the likely cause with its evidence, ranked by confidence and by financial exposure rather than by age.
• Document AI on the supporting paperwork — corporate action notices, contract notes, and custodian statements read and reconciled against the disputed values, which is where the answer usually is.
• Drafted counterparty queries — for genuine breaks, the agent drafts the query citing the specific trade, field, and values in dispute. An analyst reviews and sends; the system does not correspond with counterparties, and it never adjusts a book.

Rollout

We ran classification in advisory mode for eight weeks, comparing proposed causes against what analysts actually determined, and enabled automatic resolution one cause category at a time as each demonstrated accuracy. Anything touching the book of record required a human then and requires one now.

Results

Most breaks now clear automatically, median resolution moved from several days to same-day, and the aged-break population — the part auditors examine — shrank to a small, well-understood set. Analysts spend their time on genuine exceptions and counterparty relationships instead of re-reading settlement conventions.

What we'd tell you

• Do the deterministic matching properly before reaching for a model. Most breaks are convention, not mystery.
• Rank by exposure, not by age. It is a one-line change with an outsized effect on risk.
• Never let a system adjust the book of record. Draft the correction, have a human apply it.
• Treat source mapping as versioned code with tests. It is the part that silently rots.