Practice area

Model validation.

Model validation that satisfies regulatory requirements and delivers more than a passing opinion. Technical findings on model performance and the development lifecycle — giving the second line of defense something to build on.

Independent validation, by someone who has built models.

A validation report should hold up to an examiner — and to the modeler whose model it covers.

The SR 11-7 framework asks for independent validation across three dimensions: conceptual soundness, ongoing monitoring, and outcomes analysis. Done well, a validation gives the model risk committee real comfort. Done poorly, it gives them a 60-page document that says 'no material issues identified' and an examiner who finds three.

Our practice covers credit risk models (PD/LGD/EAD, CECL, CCAR loss forecasting), market risk models (VaR, sensitivity, stress), AML transaction monitoring models (rule sets and machine-learning hybrids), and the increasing class of AI/ML models used in underwriting, fraud detection, and customer-facing applications. For each, we validate against the data, the assumptions, the implementation, and the use.

The validator has to understand the model. That sounds obvious; it is what most validation reports fail at. Edgar built his career as a quantitative practitioner before he became a validator. The reports our practice produces show the math; they do not paper over it.

The work in this practice, named.

  1. Credit risk models PD/LGD/EAD, CECL allowance models, CCAR/DFAST loss forecasting, scorecard models.
  2. Market risk models VaR, expected shortfall, sensitivity, scenario and reverse-stress models.
  3. AML / TM models Rule-set calibration, threshold tuning, hybrid ML-based detection systems.
  4. AI / ML models Underwriting, fraud, churn, and customer-facing models — including fairness, explainability, and drift monitoring.
  5. Conceptual soundness Theory, assumptions, choice of methodology, alternatives considered, data appropriateness.
  6. Ongoing monitoring & outcomes Backtesting, benchmarking, sensitivity, monitoring plan, threshold setting.

A model validation, beginning to end.

Phase Deliverable
Intake Model documentation reviewed, data dictionary received, scope confirmed with model risk management.
Replication Independent replication on the same data; alternative specifications considered.
Testing Sensitivity, stability, fairness (where applicable), backtesting, benchmarking.
Reporting Findings rated, validation report drafted, MRMC presentation prepared.