Economic uncertainty doesn't just alter your loss forecasts. It reveals every shortcut you took when building your Current Expected Credit Losses (CECL) model three years ago during stable conditions.
This checklist helps your model risk and assurance teams validate CECL models when forward-looking scenarios are crucial. It covers what SR 11-7 requires, what your auditors will ask, and what breaks first when recession indicators start flashing.
What This Checklist Covers
You'll find actionable items for CECL model validation organized around three phases: pre-implementation review, ongoing monitoring, and stress scenario readiness. Each item references specific SR 11-7 requirements and includes a clear pass/fail state.
This isn't about initial CECL adoption compliance. It's about proving your model works when economic assumptions shift faster than your quarterly review cycle.
Prerequisites
Before starting this checklist, confirm you have:
- Model documentation identifying all data sources, modeling techniques, and key assumptions (SR 11-7 requires this for effective challenge).
- Access to model code and version control showing changes since the last validation.
- Historical performance data covering at least one full credit cycle, if available.
- Current economic scenarios your model uses for reasonable and supportable forecast periods.
- Defined roles for who approves scenario changes, model overrides, and materiality thresholds.
If you're missing any of these, stop. Your validation will produce findings, not assurance.
Checklist Items
Data Integrity and Lineage
1. Verify charge-off definitions match accounting policy
Your model's definition of "default" must align with how your finance team books charge-offs. Check the data dictionary against GL account mappings.
Good looks like: A documented reconciliation showing model training data charge-off rates match audited financial statement rates within your materiality threshold for the same period.
2. Confirm loan-level data completeness for past 12 months
Missing FICO scores, origination dates, or payment histories create silent errors in loss curves. Run completeness checks on every field your model consumes.
Good looks like: Completeness reports showing >99% population coverage for required fields, with documented exclusion rules for the remainder that don't introduce selection bias.
3. Document data transformations between source and model
Every filter, join, and calculated field between your loan servicing system and model input creates validation risk. Map each step.
Good looks like: A data lineage diagram your IT auditor can follow, with SQL or transformation logic attached, and a named owner for each pipeline stage.
Model Logic and Assumptions
4. Test scenario weightings under stress conditions
Your probability-weighted scenarios might assign 60% weight to baseline, 20% to upside, 20% to downside. Rerun that weighting logic using 2008 or 2020 economic indicators.
Good looks like: Documented results showing how scenario weights would shift if unemployment spiked 3 percentage points in one quarter, with evidence the shift is reasonable given historical precedent.
5. Validate reasonable and supportable period selection
You chose 12 months, 24 months, or 36 months for your forecast horizon. Prove that choice using empirical evidence, not convenience.
Good looks like: Analysis comparing your model's forecast accuracy at different horizons against actual performance, with a written justification for why your selected period minimizes error without introducing pro-cyclical bias.
6. Review reversion methodology to historical loss rates
After your reasonable and supportable period, you revert to long-run averages. Which averages? Calculated how? Updated when?
Good looks like: A policy stating the lookback window (e.g., 10 years, full cycle), calculation method (mean, median, loss curve fitting), and refresh frequency (quarterly, annually), with evidence of the last refresh.
7. Challenge qualitative adjustments and management overlays
If your team added basis points to the model output "because it feels low," document the rationale with supporting evidence. SR 11-7 requires effective challenge of expert judgment.
Good looks like: A memo explaining each overlay, the business condition that triggered it, the quantitative support (even if imperfect), and the sunset date when you'll revisit whether it's still needed.
Ongoing Performance Monitoring
8. Compare model predictions to actual losses quarterly
Your model predicted X% loss rate for Q1 vintages. Actual losses are now observable. Calculate the error.
Good looks like: A tracking report showing predicted vs. actual by vintage, product, and risk segment, with variance explanations for anything exceeding your monitoring threshold (commonly ±10-20% for allowance models).
9. Test model stability across economic scenarios
Run your model using the past quarter's actual economic data, then rerun it with a 200 basis point unemployment increase. Does the output change proportionally?
Good looks like: Sensitivity analysis showing loss estimates respond to economic inputs in the expected direction and magnitude, without implausible jumps or threshold effects that suggest model instability.
10. Verify model inputs refresh on schedule
Economic forecasts, prepayment speeds, and recovery rate assumptions should update per your model documentation. Confirm they actually do.
Good looks like: Audit logs or change tickets showing input data refreshed on the documented schedule, with version control indicating what changed and who approved it.
Common Mistakes
Treating CECL as a finance problem, not a model risk problem. Your accounting team owns the reserve number. Your model risk team owns whether the model producing that number is sound. These are different accountability chains with different validation standards.
Validating the model once at implementation, then only monitoring outputs. SR 11-7 requires ongoing validation, not just performance monitoring. You need to re-validate assumptions, logic, and data when economic conditions change materially.
Using vendor models without independent validation. Outsourced models still require the same validation rigor. You can't delegate accountability to your vendor, even if they provided the CECL engine.
Skipping documentation for "temporary" overrides. Every override that persists more than one reporting cycle needs documentation explaining why the model itself wasn't updated. Undocumented overrides are model risk findings waiting to happen.
Next Steps
If you found gaps in this checklist, prioritize them by regulatory exposure and model materiality. Items 1-3 (data integrity) typically surface in audits first. Items 7 and 9 (overlays and scenario testing) matter most when economic conditions deteriorate.
Schedule your next model validation to include stress scenario testing, not just business-as-usual performance review. The goal isn't to predict the next recession perfectly. It's to prove your model won't produce nonsense when forward-looking indicators start moving in directions your training data never saw.
If you're using the same validation approach you used in 2019, you're not ready for 2024's economic volatility.
