A practical monitoring framework for lending models at small institutions: CDFIs, community development credit unions, community banks, and nonprofit loan funds.

It covers what to check, how often, what level should trigger action, who acts, and what to write down. It is sized for a lender that scores a few dozen applications a month and has no dedicated model-risk or data-science team.

You make lending decisions with a scorecard, a vendor score, or a rules engine, and you need to know it is still working. You may have one analyst, or none. Your loan volumes are small enough that techniques built for large banks give unreliable answers.

1. It tells you what to do with the numbers. Open-source libraries compute drift and fairness metrics. Supervisory guidance sets principles. Neither tells a small lender which checks to run each month, what level should trigger an alert, who acts, and what to write down. This guide does, in chapter 6 and the one-page checklist.

2. It is built for small portfolios and slow outcomes.

• Standard PSI thresholds (0.10 / 0.25) fire on pure noise at small volumes. With 60 applications and 10 bins, the median PSI from noise alone is about 0.15. The guide uses a sample-size-aware threshold instead (chapter 2).
• Defaults take a year to show up, so the guide gives you early-warning checks you can run at six months on book (chapter 3).
• Performance metrics are always reported with confidence intervals, because a few dozen defaults cannot support a point estimate on their own (chapter 4).
• Everything can be run in a spreadsheet (spreadsheet/monitoring-calculator.xlsx) or in a single Python notebook.

3. It goes from fair-lending outcomes to inputs. Beyond testing whether approval rates differ across groups, it shows how to find which inputs drive a difference, whether each has a business justification, and whether a less discriminatory alternative exists without meaningful loss of predictive power (chapter 5). It reflects the 2026 changes to Regulation B.

flowchart TB
    subgraph DATA["Your data"]
        A["Application data<br/><i>available day one</i>"]
        S["Scores & decisions<br/><i>available immediately</i>"]
        O["Loan outcomes<br/><i>mature in 6–12 months</i>"]
    end
    subgraph CHECKS["Checks, by cadence"]
        direction LR
        subgraph M["Monthly · ch 2–3"]
            M1[Data quality] --- M2[Score PSI] --- M3[Decision rates]
        end
        subgraph Q["Quarterly · ch 2, 3, 5"]
            Q1[Input CSI] --- Q2[Early delinquency] --- Q3[Approval AIR]
        end
        subgraph Y["Annually · ch 4–5"]
            Y1[AUC & calibration] --- Y2[Input proxy review] --- Y3[LDA search]
        end
    end
    DATA --> CHECKS
    CHECKS --> T["Alert thresholds & owners<br/><i>ch 6</i>"]
    T --> I["Investigate<br/><i>data, mix, or model?</i>"]
    I --> D["Decide<br/><i>keep, recalibrate, rebuild, restrict</i>"]
    D --> R1[Monitoring report · ch 8]
    D --> R2[Fair-lending review · ch 5]
    D --> R3[Model inventory · ch 1]

Everything runs from a one-page checklist, a spreadsheet, or a Python notebook.

No code: read the checklist, open the spreadsheet calculator, and replace the yellow example values with your own counts.

Python:

git clone https://github.com/hannahxuhang-hue/lending-model-monitoring-guide.git
cd lending-model-monitoring-guide
pip install -r requirements.txt
cd examples && jupyter notebook walkthrough.ipynb

monitoring_checks.py is a single file of plain functions (pandas, numpy, scipy) that you can copy into your own work.

All data in this repository is synthetic, generated by examples/synthetic_data.py. It does not describe any real lender, applicant, or model. The group-membership probabilities are simulated, not computed from names or addresses.

Causal Uplift Targeting Guide: measuring and targeting the incremental effect of outreach at small lenders. Any targeting model it produces belongs in the monitoring cycle described here.

This is a personal project, written on personal time. It draws only on published methods and public regulatory sources, cited in each chapter and below. It contains no data, code, thresholds, or documentation from any employer, and it is not affiliated with or endorsed by any employer or financial institution.

This guide is general methodology. It is not legal, regulatory, or compliance advice. Thresholds are illustrative starting points. Confirm your obligations with counsel and your examiner.

• Board of Governors of the Federal Reserve System (2026). SR 26-2: Revised Guidance on Model Risk Management.
• OCC (2025). Bulletin 2025-26: Model Risk Management: Clarification for Community Banks.
• Consumer Financial Protection Bureau (2026). Equal Credit Opportunity Act (Regulation B), final rule, April 22, 2026.
• Consumer Financial Protection Bureau (2014). Using Publicly Available Information to Proxy for Unidentified Race and Ethnicity.
• Elliott, M. N., et al. (2009). Using the Census Bureau's surname list to improve estimates of race/ethnicity and associated disparities. Health Services and Outcomes Research Methodology, 9(2), 69–83.
• Gillis, T., Meursault, V., & Ustun, B. (2024). Operationalizing the search for less discriminatory alternatives in fair lending. Proceedings of ACM FAccT 2024.
• Hanley, J. A., & McNeil, B. J. (1982). The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology, 143(1), 29–36.
• Yurdakul, B. (2018). Statistical Properties of Population Stability Index. Dissertation, Western Michigan University.
• Federal Reserve Bank of Richmond (2026). 2025 CDFI Survey: Key Findings.

If you work at a lender or network and use or adapt this material, or find it doesn't fit how you work, please share how you used it or open an issue. See CONTRIBUTING.md for other ways to help. Changes made in response to feedback are recorded in CHANGELOG.md.

See CITATION.cff, or: Xu, H. (2026). Lending Model Monitoring Guide (Version 1.0). https://github.com/hannahxuhang-hue/lending-model-monitoring-guide

Guide text, checklist, templates, and spreadsheet: CC BY 4.0. Code: MIT. You may use, adapt, and redistribute both, including inside your institution's own policies, with attribution.

Author: Hang (Hannah) Xu