> ## Documentation Index
> Fetch the complete documentation index at: https://docs.creditbenchmark.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Data Validation

> Credit Benchmark validates contributed PD data through layered controls including outlier detection, cross-bank checks, and trend analysis before consensus.

Credit Benchmark validates contributed data to support reliable consensus outputs. Before PD validation begins, contributions are matched to the correct legal entity — see [Data Submission](/methodology/data-processing/data-submission) for how entity resolution works.

## Observation-Level PD Validation

Each monthly submission passes through layered controls at the observation level before it can contribute to the published consensus. The purpose is to identify data inconsistencies or outliers where a PD value is materially different from:

* **Other contributions from different banks** — a single extreme outlier can skew the consensus average. Cross-contributor comparison identifies values that are anomalous relative to the broader market view for that entity.

* **Previous contributions from the same bank** — banks can and do make large downgrades legitimately. The threshold here is set high to catch only extreme shifts — the kind more likely to be a submission or processing error than a genuine credit reassessment. Moves that cross the threshold are reviewed and cleared before publication.

<Note title="What is quarantine?">
  An observation that fails a control is temporarily excluded from consensus eligibility until it is cleared or otherwise handled per policy.
</Note>

## Quarantine Rules

Credit Benchmark applies several classes of checks to validate PD data and identify outliers. The published methodology summarises the control framework rather than listing every internal threshold or operating rule.

* **Relative Standard Deviation (RSD)** — standard deviation divided by mean across contributed PDs for an entity; higher values indicate wider dispersion
* **Logarithmic PD distance** — the log-scale distance of an observation from the entity median; large negative values indicate unusually low PDs, large positive values indicate unusually high PDs
* All checks are performed on **1-year TTC PDs**

| Stage                       | What we review                                                                   | Typical outcome                                            |
| --------------------------- | -------------------------------------------------------------------------------- | ---------------------------------------------------------- |
| **Duplicates**              | More than one current observation submitted by the same bank for the same entity | One record retained or held for manual review              |
| **Consensus spread**        | Observation sits materially away from the current contributor set                | Unusual values quarantined pending review                  |
| **Month-on-month movement** | PD has moved unusually far versus the same bank's prior view                     | Large moves reviewed before publication                    |
| **New entrant screening**   | First-time or returning contribution inconsistent with the existing consensus    | Additional scrutiny applied before entering the pool       |
| **Default classification**  | How 10,000 bps observations should be treated under publication policy           | Technical and potential default cases handled consistently |

## Persistent Outliers

Any PD that remains an outlier for multiple cycles is treated as a **persistent outlier** and stays quarantined until either:

* **Client confirmation** — the bank confirms the view is intentional
* **PD change** — the PD changes for that entity and the observation is re-evaluated

This ensures persistent data-quality issues and genuine changes in credit view are both handled consistently.