Reconcile the denominator and observation window behind a settlement-failure rate
Compare temporary settlement-failure changes only with explicit due-transaction populations, asset classes, observation windows and as-of states.
Ask about the denominator first
A report of increased settlement failures should prompt a question about what was counted out of which population. Equal rates can describe different operations when due transactions or observation times differ. Truthound can check quantity relationships in a prepared aggregate table. It does not forecast market risk or settle an outstanding transaction.
A September report about the first half
ESMA published its second risk-monitoring report on September 10, 2026, describing conditions during the first half of the year. Its release notes a brief increase in settlement failures across asset classes in early April 2026. That statement does not provide a particular institution's failure rate or establish a persistent deterioration. Clearing-house resilience and settlement-failure observations also concern different operating scopes.
Separate due date and as-of state
With authorised operating data, record settlement due date, asset class, counting basis, due quantity, unsettled quantity and the state's as-of time. Keep transaction counts separate from value. A trade unsettled at a daily cutoff and the same trade recovered later may share a due date without sharing an observation state.
Distinguish denominator changes from recovery
Start by distinguishing a changed denominator from a changed failure count. Using an older denominator after the due population expands can distort a comparison. So can overwriting an earlier cutoff with the latest recovery state. Do not average different-length stress and ordinary windows without a stated method, or fill unpublished internal quantities with invented figures.
Checks on the prepared aggregate
On a prepared Polars table, Truthound Core Python SDK 3.3.4 provides ColumnPairValidator for a declared relationship between non-null quantity fields, such as unsettled counts not exceeding due counts within one scope. QueryGroupAggregateValidator requires caller-supplied SQL, expected asset groups and bounds. Operators first define missingness, treatment of a denominator of 0 and which states count as failures.
The boundary of an existing pandas route
Retain a tested Pandera 0.35.0 pandas reconciliation function if it already maintains this register. A pandas predicate is not automatically portable to a Polars path: examine windows with a denominator of 0 and incompatible as-of rows separately on each. Compare how an undefined rate returns to its reviewer. Neither product's documentation supplies an execution result on these settlement records or a market-risk judgment.
Reconcile one asset class and window
Choose one asset class and window for the first review. Preserve original and revised aggregates as separate vintages, explaining whether a difference reflects additional transactions, later recovery or a changed denominator. If the necessary state history is unavailable, withhold that trend comparison. A valid quantity relationship does not identify an individual transaction's failure cause.
Turn a temporary observation into a precise question
A temporary increase need not become an immediate risk ranking. An account of which transactions were counted at which time makes the next operational question more precise. Hand the reconciled table to its responsible operating review, and establish causes and actual processing outcomes from the relevant system evidence.
Conditions
An authorised owner supplies a prepared Polars extract with consistent asset, period, counting/value basis, state definitions and null policy.
Alternatives
Keep a tested Pandera 0.35.0 pandas reconciliation; a Polars migration needs separate denominator-of-0, missingness and as-of counterexamples.
Limitations
Input checks do not process settlements, establish failure causes, rate an institution's risk or make an investment judgment.
Evidence
- Event checked
- Subject checked
- Modified
