What is model risk management (MRM)?

Model risk management (MRM) is the discipline of identifying, measuring, and controlling the risks arising from using mathematical and statistical models in financial decision-making. The Federal Reserve's SR 11-7 guidance defines model risk as the potential for adverse consequences from decisions based on incorrect or misused model outputs. MRM covers the full model lifecycle: development, validation, implementation, ongoing monitoring, and retirement.

What does SR 11-7 require for model validation?

SR 11-7 (Guidance on Model Risk Management) requires: independent validation by a party separate from model developers; conceptual soundness review (review of model theory, assumptions, and methodology); testing of outcomes (backtesting, benchmarking to alternative models); assessment of ongoing model performance through monitoring; and documentation of findings and any model limitations or compensating controls.

What are the most common model risk findings regulators cite?

Common model risk regulatory findings: (1) inadequate documentation — model assumptions and limitations not clearly stated; (2) insufficient independent validation — validators too close to development team; (3) weak ongoing monitoring — no performance metrics tracked post-implementation; (4) missing model inventory coverage — models not captured in the inventory; (5) CECL and DFAST model validation deficiencies are most cited by OCC and Fed examiners in 2024-2026.

AI for Model Risk Management: Claude Tools for Model Validation and SR 11-7 (2026)

How model risk teams use Claude AI for SR 11-7 model validation: model inventory management, conceptual soundness review, outcome testing and backtesting, ongoing monitoring scorecards, and regulatory examination response documentation.

Model Risk Management and AI

SR 11-7 model risk management is document-intensive and analytically demanding. A single model validation requires reviewing conceptual soundness, running sensitivity analysis, backtesting outcomes, and producing a validation report that satisfies both internal governance and regulatory examiner expectations. Claude with ClaudeFinLab accelerates the documentation, testing design, and analytical writing across the model lifecycle.

Model Inventory Management

"Design the model inventory classification framework: our bank has 147 models. Classify by: (1) materiality tier (Tier 1 = high impact, Tier 2 = moderate, Tier 3 = low/administrative); (2) model type (statistical/econometric, financial pricing, regulatory capital, behavioral, vendor); (3) validation status (validated, in-validation, overdue, retired). Our Tier 1 models include CECL ALLL model, DFAST stress testing model, LIBOR fallback model (replaced SOFR), and AML transaction monitoring. For each Tier 1 model, specify: validation frequency, owner, validator, last validation date, next due date, current findings status."
"Identify models not currently in the inventory: our lending process uses 8 scoring models from 3 vendors (FICO, TransUnion, internal). Our loan pricing model is embedded in the LOS (Loan Origination System) — is it in scope under SR 11-7? Per the guidance, vendor models and models embedded in systems must be inventoried and validated if they inform significant decisions. Draft the memo recommending which of these to add to the inventory."

Model Validation: Conceptual Soundness

"Review the conceptual soundness of our CECL PD model: the model uses logistic regression with 8 variables (DTI, LTV, FICO, employment tenure, property type, loan age, macro GDP growth, unemployment rate). Assumptions: linearity in log-odds, no multicollinearity, no time-series non-stationarity. Conceptual concerns: (1) logistic regression may not capture non-linear default behavior at extreme stress; (2) GDP and unemployment are correlated (VIF > 5); (3) model trained on 2010-2019 data may underfit 2020 COVID shock behavior. Document findings and recommend enhancements."
"Assess the sensitivity of the CECL model to key assumptions: base case lifetime loss rate 2.4%. Test: (1) +1pp unemployment → loss rate +0.42pp; (2) GDP -3% → loss rate +0.38pp; (3) FICO threshold shift (all borrowers -20 FICO points) → loss rate +0.65pp. Are these sensitivities reasonable vs historical analogues? During 2020, unemployment rose 11pp, GDP fell 9%, and CECL models increased loss estimates by 3-4x. Is our model's sensitivity consistent?"

Outcome Testing and Backtesting

"Run backtesting for the DFAST PD/LGD model: compare model-projected default rates vs actual default rates for 2019-2024 vintages. For each vintage: model PD at origination, actual 3-year cumulative default rate. Compute: Gini coefficient (discriminatory power), PSI (population stability index — is the population shifting?), Kolmogorov-Smirnov statistic. Flag vintages where actual defaults exceeded model projections by >20%. Identify if the model is optimistic, conservative, or well-calibrated."
"Design the ongoing model monitoring scorecard for CECL: monthly metrics — (1) PSI of key input variables vs development sample; (2) actual loss rate vs model projection (deviation > 15% triggers model review); (3) concentration drift (portfolio composition changes that were not in training data); (4) macro variable alignment (compare model's macro scenarios to consensus forecast). Define green/amber/red thresholds and escalation path. Format as a risk committee monitoring dashboard."

Validation Documentation and Reporting

"Draft the model validation summary for the CECL ACL model: finding classification — (1) Significant Weakness: FICO variable treatment — coefficient is statistically significant but the model uses binned FICO rather than continuous, reducing predictive power (recommend continuous treatment); (2) Acceptable Finding: training sample excludes 2020 COVID period — mitigated by qualitative overlays and management judgment adjustments; (3) Observation: macro scenario generation methodology should be documented more explicitly. Recommendations and management response per SR 11-7 validation report format."
"Draft the regulatory exam response letter: the OCC exam team identified a finding — 'model validation for [Model X] did not include challenger model benchmarking as required by SR 11-7.' Write a management response: acknowledge the finding, explain the root cause (resource constraints during COVID), describe the remediation plan (challenger model by Q2 2026, validation report by Q3 2026), and identify any compensating controls in place during the gap period."

Where to Start

Start with model inventory — you can't manage what you haven't catalogued. Ask Claude to help structure the inventory spreadsheet with the right fields (model ID, description, tier, owner, validator, validation due date, findings). For validation work, describe the model methodology and ask Claude to identify the key SR 11-7 conceptual soundness questions. Documentation drafting is where AI adds the most immediate value — paste your analysis and ask Claude to write it in SR 11-7 report format.

Get your free API key