# Model Development Document: [Model name] [version]

> Developer documentation for a quantitative risk model (for example, a
> probability of default model for the internal ratings-based approach,
> an IFRS 9 component, or a value-at-risk model): purpose and use,
> design choices, data, methodology, calibration, assumptions and
> limitations, developmental testing, implementation and the ongoing
> monitoring plan. It is written so that an independent validator can
> understand and reproduce the work, as expected by the US interagency
> Revised Guidance on Model Risk Management of 2026-04-17 (Federal
> Reserve SR 26-2, OCC Bulletin 2026-13, FDIC FIL-15-2026), PRA SS1/23
> Principle 3 (cited by number and name), the ECB guide to internal
> models (revised 28 July 2025) and, for EU IRB models, Regulation (EU)
> No 575/2013 as amended (CRR) with EBA/GL/2017/16. Replace every
> `[bracketed]` value and every row marked *Example*. The worked example
> is the fictional Fernhollow Bank retail mortgage PD model. This is a
> document format, not legal, regulatory or risk advice. Obtain the
> guidance or regulation for the normative text.

<!--
Guidance: How to use this template
- The developer writes this document during development, not after.
  Each section records a decision and the evidence for it while the
  reasoning is still fresh.
- Scale the depth to the model tier set under the Model Risk
  Management Policy (`finance/model-risk-management-policy`). A Tier 1
  model completes every section in full; a Tier 3 model may use short
  answers but must not leave sections empty.
- Write for the validator. The test of a good development document is
  that a competent person who was not on the team can rebuild the
  model from it, using the code and data references it gives.
- This document is not the validation report and not the model card.
  The Model Validation Report (`ai-data/model-validation-report`) is
  written by an independent validator; the Model Card
  (`ai-data/model-card`) is the two-page summary for users and
  committees; the Model Monitoring Plan
  (`ai-data/model-monitoring-plan`) may hold the detailed monitoring
  design that section 10 summarises.
- For market risk models, section 12 lists the adaptations; use the VaR
  Backtesting Report (`finance/var-backtesting-report`) for the
  back-testing evidence once the model is live.
- Related documents: ECL Methodology (`finance/ecl-methodology`),
  Credit Rating Rationale (`finance/credit-rating-rationale`), Stress
  Testing Programme (`finance/stress-testing-programme`), Market Risk
  Limit Framework (`finance/market-risk-limit-framework`), Data
  Dictionary (`architecture/data-dictionary`).
-->

## Document control

<!-- Guidance: Link the document to the model ID in the inventory and to the exact code and data versions, so that the document, the code and the validated model always match. -->

| Field | Value |
|---|---|
| Document ID | [MDD-MOD-NNN-vX] |
| Model ID and name | [MOD-NNN, model name] |
| Model version documented | [vX.Y] |
| Model tier | [Tier 1 / 2 / 3] |
| Model type | [PD / LGD / EAD / ECL component / VaR / pricing / other] |
| Regulatory use | [IRB own funds / IFRS 9 / IRRBB / internal limits / none] |
| Version | [1.0] |
| Status | [Draft / Submitted for validation / Approved / Superseded] |
| Owner | [@model-owner] |
| Lead developer | [@lead-developer] |
| Approvers/reviewers | [@head-of-modelling], [@model-owner], [@head-of-model-risk-management] |
| Code repository and commit | [Repository, tag or commit hash] |
| Data snapshot reference | [Dataset ID and extraction date] |
| Last updated | YYYY-MM-DD |
| Classification | [Confidential] |

### Approvals

| Role | Name | Decision | Date |
|---|---|---|---|
| Lead developer | [@name] | [Submit for validation] | YYYY-MM-DD |
| Head of modelling | [@name] | [Approve technical content] | YYYY-MM-DD |
| Model owner | [@name] | [Confirm purpose, use and limitations] | YYYY-MM-DD |
| Model risk management | [@name] | [Accepted for validation] | YYYY-MM-DD |

### Revision history

| Version | Date | Author | Change |
|---|---|---|---|
| 0.1 | YYYY-MM-DD | [@author] | Purpose, scope and data sections drafted |
| 0.9 | YYYY-MM-DD | [@author] | Methodology, calibration and testing complete, internal peer review |
| 1.0 | YYYY-MM-DD | [@author] | Submitted for independent validation |
| 1.1 | 2026-10-25 | [@author] | Updated to the July 2025 ECB guide to internal models; margin of conservatism applied as a scaling of the grade PDs; excluded-record default rate and capital sensitivity corrected |

## Contents

- [1. Executive summary](#1-executive-summary)
- [2. Purpose, scope and intended use](#2-purpose-scope-and-intended-use)
- [3. Design choices and theory](#3-design-choices-and-theory)
- [4. Data](#4-data)
- [5. Methodology](#5-methodology)
- [6. Calibration](#6-calibration)
- [7. Developmental testing](#7-developmental-testing)
- [8. Assumptions, expert judgement and limitations](#8-assumptions-expert-judgement-and-limitations)
- [9. Implementation](#9-implementation)
- [10. Ongoing monitoring plan](#10-ongoing-monitoring-plan)
- [11. Regulatory requirements traceability](#11-regulatory-requirements-traceability)
- [12. Adaptations for market risk models](#12-adaptations-for-market-risk-models)
- [13. Glossary and references](#13-glossary-and-references)
- [Standard and licence note](#standard-and-licence-note)

## 1. Executive summary

<!-- Guidance: One page for the model owner and committee: what the model does, how it was built, how well it performs, what it cannot do, and what is being asked (validation, approval, supervisory application). Pitfall: a summary full of statistics with no statement of the main limitation. -->

[Summary in five to eight sentences.]

*Example:* This document describes version 3.0 of Fernhollow Bank's
retail mortgage probability of default model (MOD-001), used for IRB
own funds requirements, pricing and IFRS 9 staging inputs. Version 3.0
replaces a 2019 scorecard whose discrimination had declined and whose
data did not reflect the 2022 to 2024 interest rate rises. The new
model is a logistic regression scorecard on six risk drivers, mapped to
an eight-grade master scale and calibrated to a long-run average
default rate of 1.35% with a margin of conservatism of 0.15 percentage
points. Out-of-time testing on 2023 to 2024 data gives a Gini
coefficient of 0.62 and the 2025 confirmatory cohort 0.61, with no grade with a statistically significant calibration
deviation at the 95% level. The main limitations are the short history
of high-rate environments and thin data for buy-to-let loans, both
addressed by conservatism and monitoring. The model is submitted for
independent validation and, because it is a material change to an IRB
model, for the supervisory approval process.

## 2. Purpose, scope and intended use

<!-- Guidance: A clear statement of purpose is the starting point that SR 26-2 describes for sound development. Agree it with the model owner and users before choosing data or methods. Pitfall: "general purpose" models used for decisions their design never considered. -->

| Item | Description |
|---|---|
| Business problem | [What decision or estimate the model supports] |
| Business problem *(Example)* | Estimate the one-year probability of default of performing retail mortgage loans |
| Portfolio in scope | [Products, segments, geographies], *Example:* owner-occupied and buy-to-let residential mortgages, EU head office book |
| Portfolio out of scope | [Exclusions], *Example:* defaulted loans (covered by the LGD model), staff loans, loans under 3 months on book (assigned to a fixed grade) |
| Approved uses | [List], *Example:* IRB PD for own funds; risk-based pricing; input to IFRS 9 significant increase in credit risk assessment |
| Uses not permitted | [List], *Example:* lifetime PD without the IFRS 9 term-structure overlay; credit decisions on non-mortgage products |
| Users | [Units and roles] |
| Output | [Description], *Example:* PD per loan, rating grade 1 to 8, monthly |
| Model tier | [Tier and reason], *Example:* Tier 1, regulatory capital use |

Describe how users were involved in setting the purpose, which questions
they raised about the previous model, and how those questions shaped
this version. SR 26-2 notes that input from users strengthens both
model understanding and model quality. Record the meeting dates or
sign-offs as evidence.

*Example:* Retail credit underwriting and pricing teams took part in
three workshops in 2025. They reported that the old model ranked
recently originated high loan-to-value loans too favourably. This led
the team to test a loan-to-value variable updated with a house price
index and an interaction with time on book.

## 3. Design choices and theory

<!-- Guidance: Explain why this approach suits the problem, and what alternatives were considered and rejected. Validators assess conceptual soundness largely from this section. Pitfall: describing only the chosen method, which leaves the validator unable to judge whether it was the right choice. -->

State the economic or statistical theory that links the inputs to the
output and the reason the chosen model class fits the data and the use.
For a PD model, explain the rating philosophy (point-in-time,
through-the-cycle or hybrid), because it determines how grades migrate
over the cycle and how the model interacts with capital and IFRS 9.

| Option | Description | Strengths | Weaknesses | Decision |
|---|---|---|---|---|
| A *(Example)* | Logistic regression scorecard with grade mapping | Transparent, stable, accepted by users and supervisors, easy to implement in the rating engine | Limited non-linearity | Selected |
| B *(Example)* | Gradient boosted trees | Higher development Gini (0.67) | Lower transparency; monotonicity constraints needed; out-of-time Gini fell to 0.60, below option A | Rejected; kept as challenger benchmark |
| C *(Example)* | Recalibrate the 2019 scorecard | Lowest effort | Discrimination already below the monitoring threshold | Rejected |

*Example rating philosophy:* The scorecard ranks borrowers on mostly
point-in-time drivers, but grade PDs are calibrated to a long-run
average default rate. The result is a hybrid philosophy: borrowers
migrate between grades as their circumstances change, while grade PDs
stay stable across the cycle. The IFRS 9 team adds a forward-looking
adjustment in the ECL Methodology (`finance/ecl-methodology`), not in
this model.

## 4. Data

<!-- Guidance: Data decisions often drive model risk more than method choices. Record sources, definitions, cleaning steps and every exclusion with its effect on counts and default rates. Pitfall: exclusions applied in code but not documented, so the validator cannot reproduce the sample. -->

### 4.1 Sources and lineage

| Dataset ID | Source system | Content | Period | Owner |
|---|---|---|---|---|
| DS-01 *(Example)* | Mortgage servicing platform | Loan attributes, balances, arrears | 2014-01 to 2025-12 | Head of retail operations |
| DS-02 *(Example)* | Origination system | Application data: income, debt service, employment | 2008-01 to 2025-12 | Head of retail credit |
| DS-03 *(Example)* | Default database | Default events and dates under the bank's default definition | 2010-01 to 2025-12 | Head of credit risk control |
| DS-04 *(Example)* | Public house price index | Regional house prices for indexed loan-to-value | 2008-01 to 2025-12 | Market data team |

Describe the lineage from source to modelling dataset, including joins,
keys and the extraction date. Reference the data dictionary entries
(`architecture/data-dictionary`) for each variable used.

### 4.2 Default definition

State the default definition and how it is applied in the data. For EU
IRB models, the definition must follow CRR Article 178 and the EBA
Guidelines on the application of the definition of default
(EBA/GL/2016/07, amended by EBA/GL/2026/05, published 2026-05-07, applicable from 2026-10-19; check
the consolidated version). Describe how historical defaults recorded
under earlier definitions were adjusted, and the effect of the
adjustment on default rates.

*Example:* Defaults are identified when a loan is more than 90 days past
due on a material amount or when the bank considers the borrower
unlikely to pay, including distressed restructurings. Defaults before
2021 were re-derived from the arrears history under the current
definition; this raised the 2014 to 2020 average default rate by 0.06
percentage points.

### 4.3 Sample construction and exclusions

| Step | Rule | Observations remaining | Defaults remaining |
|---|---|---|---|
| Start *(Example)* | All performing loans at each annual snapshot, 2014 to 2024 | 1,912,400 | 22,950 |
| Exclusion 1 *(Example)* | Loans under 3 months on book | 1,884,100 | 22,610 |
| Exclusion 2 *(Example)* | Staff loans | 1,871,300 | 22,470 |
| Exclusion 3 *(Example)* | Records with missing income that cannot be recovered | 1,860,000 | 22,320 |
| Development sample *(Example)* | Snapshots 2014 to 2022 | 1,512,000 | 18,450 |
| Out-of-sample holdout *(Example)* | Random 20% of loans in 2014 to 2022, held out from estimation | Included in the line above, separated by loan ID | |
| Out-of-time sample *(Example)* | Snapshots 2023 and 2024 | 348,000 | 3,870 |

The final modelling dataset holds 1,860,000 loan-year observations with
22,320 defaults, a pooled default rate of 1.20%. The development and
out-of-time samples add up to that total (1,512,000 plus 348,000
observations; 18,450 plus 3,870 defaults). A separate 2025 cohort of
172,000 loans, which became available after development, is used for
the confirmatory test in section 7.3.

### 4.4 Data quality

Report missing values, outliers, inconsistencies and how each was
treated. Show the effect of treatment on the default rate where it is
material. Record known data limitations here and carry them to
section 8.

| Variable | Issue | Treatment | Effect |
|---|---|---|---|
| Income at origination *(Example)* | 2.1% missing before 2016 | Recovered from scanned files where possible; remaining cases excluded (exclusion 3) | Default rate of excluded records 1.33% (150 of 11,300), close to the 1.20% sample average, so the exclusion is unlikely to bias the model |
| Indexed LTV *(Example)* | Regional index starts 2008 | Loans originated before 2008 indexed from the 2008 level | Affects 3% of observations |
| Employment type *(Example)* | Inconsistent codes before 2018 | Mapped to four categories | None measurable |

## 5. Methodology

<!-- Guidance: Describe the method in enough detail to reproduce it: variable selection, transformations, estimation, segmentation and grade mapping. Put formulas in display maths with every symbol defined. Pitfall: listing the final variables without explaining why others were dropped. -->

### 5.1 Model form

The model estimates the probability of default over the next twelve
months with a logistic regression:

$$
PD_i = \frac{1}{1 + e^{-(\beta_0 + \sum_{k=1}^{K} \beta_k x_{ik})}}
$$

where $x_{ik}$ is the transformed value of risk driver $k$ for loan $i$,
$\beta_k$ are the estimated coefficients and $K$ is the number of risk
drivers. The linear part is converted to a score and the score is
mapped to rating grades in section 5.4.

### 5.2 Variable selection

Describe the long list, the univariate analysis, correlation checks,
the multivariate selection procedure and the business review. Explain
each variable's expected direction and confirm that the estimated sign
matches it.

| Risk driver | Transformation | Expected sign on PD | Estimated sign | Reason for inclusion |
|---|---|---|---|---|
| Indexed loan-to-value *(Example)* | Weight of evidence, 6 bins | Positive | Positive | Lower equity raises default incentive and loss |
| Debt service to income at origination *(Example)* | Weight of evidence, 5 bins | Positive | Positive | Affordability |
| Months on book *(Example)* | Weight of evidence, 4 bins | Negative | Negative | Seasoning |
| Worst arrears in the last 12 months *(Example)* | Categories 0, 1 to 29, 30 to 59, 60 to 89 days | Positive | Positive | Behavioural signal |
| Employment type *(Example)* | Four categories | Varies | As expected | Income stability |
| Buy-to-let flag *(Example)* | Binary | Positive | Positive | Different repayment source |

Variables considered and rejected, with reasons: [list].

*Example:* Region was tested but dropped because its effect disappeared
once indexed loan-to-value was included, and it raised concerns about
fair treatment. Current account balance trends were dropped because
data covered only 40% of borrowers.

### 5.3 Estimation

Record the software, version, estimation settings, treatment of class
imbalance and the coefficient table with standard errors. Attach the
full output as an annex, and reference the code path that produces it.

### 5.4 Rating scale

Map scores to grades on the bank's master scale. Each grade has a
calibrated PD (section 6). Explain the number of grades and the
boundaries, and show that no grade holds an excessive concentration of
exposures.

| Grade | Calibrated PD *(Example)* |
|---|---|
| 1 | 0.10% |
| 2 | 0.25% |
| 3 | 0.50% |
| 4 | 1.00% |
| 5 | 2.00% |
| 6 | 4.00% |
| 7 | 8.00% |
| 8 | 20.00% |

## 6. Calibration

<!-- Guidance: Calibration sets the level of PD; discrimination only sets the order. Explain the target, the period it covers, and every adjustment for conservatism. Pitfall: calibrating to the development sample average when that period was unusually benign. -->

Explain the calibration target, the period used to compute it and why
the period is representative of the likely range of variability of
default rates. For EU IRB models, the target is a long-run average
default rate, and EBA/GL/2017/16 (applicable from 2021-01-01) describes
how to estimate it and how to quantify a margin of conservatism by
category of deficiency. Obtain the guidelines for the normative text.

*Example:* The long-run average default rate is the average of annual
default rates from 2010 to 2024, which includes the tail of the
euro-area debt crisis and the 2022 to 2024 rate rises. It equals
1.35%, higher than the 1.20% pooled default rate of the modelling
sample because the sample starts in 2014. Loan-level servicing data
(DS-01) start in 2014, so the 2010 to 2013 annual default rates use
default counts from the default database (DS-03) and portfolio counts
reconciled to the finance ledger; the annex documents the denominators
and the reconciliation. The calibration scales the
score-to-PD relationship so that the portfolio's average predicted PD
equals 1.35% on the reference portfolio, then adds a margin of
conservatism.

| Margin of conservatism component | Reason | Add-on (percentage points) |
|---|---|---|
| Category A, data and methodology deficiencies *(Example)* | Default re-derivation for pre-2021 data | 0.06 |
| Category B, relevant changes *(Example)* | Lending policy change in 2023 on debt service limits | 0.04 |
| Category C, general estimation error *(Example)* | Statistical uncertainty of the long-run average | 0.05 |
| **Total** *(Example)* | | **0.15** |

The calibrated portfolio-level PD including the margin is 1.35% plus
0.15 percentage points, which is 1.50%. The grade PDs in section 5.4
are the calibrated PDs before the margin of conservatism: on the
reference portfolio (the grade mix of the 2010 to 2024 calibration
period) they average 1.35%. For own funds each grade PD is scaled by
1.50 / 1.35 = 1.111 to include the margin (for example, grade 4 moves
from 1.00% to 1.11% and grade 6 from 4.00% to 4.44%), subject to the
regulatory PD input floor. IFRS 9 staging and pricing use the grade PDs
without the margin, because the margin is a regulatory conservatism
add-on, not a best estimate.

## 7. Developmental testing

<!-- Guidance: Show that the model works on data it did not see, across segments and over time. SR 26-2 lists out-of-sample and out-of-time testing, comparison with alternatives and assessment of data quality among typical testing activities. Pitfall: reporting only in-sample statistics. -->

### 7.1 Discrimination

The Gini coefficient measures how well the model ranks defaulters above
non-defaulters. It relates to the area under the ROC curve (AUC) as:

$$
\text{Gini} = 2 \times \text{AUC} - 1
$$

| Sample | AUC *(Example)* | Gini *(Example)* |
|---|---|---|
| Development | 0.820 | 0.640 |
| Out-of-sample holdout | 0.815 | 0.630 |
| Out-of-time 2023 to 2024 | 0.808 | 0.616 |
| Confirmatory 2025 cohort | 0.805 | 0.610 |
| Previous model on 2025 cohort (benchmark) | 0.760 | 0.520 |
| Challenger gradient boosted trees on 2025 cohort | 0.800 | 0.600 |

Each Gini equals twice the AUC minus one (for example, 2 times 0.805
minus 1 equals 0.610).

### 7.2 Stability

The population stability index (PSI) compares the distribution of a
recent population across score bands with the development
distribution:

$$
PSI = \sum_{j} (a_j - e_j) \ln\left(\frac{a_j}{e_j}\right)
$$

where $e_j$ is the share of the development population in band $j$ and
$a_j$ is the share of the recent population.

| Band | Development share | 2025 share | Contribution |
|---|---|---|---|
| Low risk *(Example)* | 30% | 25% | 0.0091 |
| Medium risk *(Example)* | 50% | 52% | 0.0008 |
| High risk *(Example)* | 20% | 23% | 0.0042 |
| **PSI** *(Example)* | | | **0.0141** |

A PSI of 0.014 is well below the bank's amber threshold of 0.10, so the
2025 population is similar to the development population.

### 7.3 Calibration test

Compare predicted and observed defaults by grade. A simple test for each
grade uses the normal approximation to the binomial distribution:

$$
z = \frac{O - nPD}{\sqrt{nPD(1 - PD)}}
$$

where $n$ is the number of loans in the grade, $PD$ the grade PD and $O$
the observed number of defaults. Values of $|z|$ above 1.96 indicate a
deviation at the 95% level. The table uses the grade PDs before the
margin of conservatism, as is common for the accuracy test, and the
2025 cohort of 172,000 loans. The cohort's average predicted PD is
1,967.5 / 172,000 = 1.14%, below the 1.35% central tendency, because
the 2025 book holds fewer loans in the high-risk grades than the
reference portfolio.

| Grade | Loans | Grade PD | Expected defaults | Observed defaults | Observed rate | z |
|---|---|---|---|---|---|---|
| 1 *(Example)* | 40,000 | 0.10% | 40.0 | 31 | 0.08% | -1.42 |
| 2 *(Example)* | 45,000 | 0.25% | 112.5 | 98 | 0.22% | -1.37 |
| 3 *(Example)* | 35,000 | 0.50% | 175.0 | 160 | 0.46% | -1.14 |
| 4 *(Example)* | 22,000 | 1.00% | 220.0 | 231 | 1.05% | 0.75 |
| 5 *(Example)* | 15,000 | 2.00% | 300.0 | 312 | 2.08% | 0.70 |
| 6 *(Example)* | 8,000 | 4.00% | 320.0 | 301 | 3.76% | -1.08 |
| 7 *(Example)* | 5,000 | 8.00% | 400.0 | 428 | 8.56% | 1.46 |
| 8 *(Example)* | 2,000 | 20.00% | 400.0 | 377 | 18.85% | -1.29 |
| **Total** *(Example)* | **172,000** | | **1,967.5** | **1,938** | **1.13%** | |

No grade shows a deviation at the 95% level. The portfolio's observed
default rate in 2025 (1.13%) is below the long-run average of 1.35%,
which is expected in a year with falling unemployment and supports the
choice to calibrate to the long-run average rather than to recent
years. Note that the per-grade test ignores default correlation, so it tends to
flag deviations too often in a year with a common shock; the validator
may apply tests that allow for correlation, such as a Vasicek-based
confidence interval.

### 7.4 Sensitivity and benchmarking

Test how outputs change when key inputs and assumptions change, and
compare results with an alternative model or external benchmark.

*Example:* A 10% fall in house prices raises the average predicted PD
of the reference portfolio, including the margin, from 1.50% to 1.71%
through the indexed loan-to-value driver. Removing
the margin of conservatism lowers own funds requirements for the
portfolio by about 6% (CRR retail mortgage formula with a 15%
asset correlation and constant LGD; the capital engine gives the
exact figure). The challenger gradient boosted model ranks 91%
of loans within one grade of the scorecard.

## 8. Assumptions, expert judgement and limitations

<!-- Guidance: List every assumption and judgement that a reasonable expert could have made differently, with its support and its effect. Users and validators rely on this section more than any other. Pitfall: a limitations list that is generic ("past performance may not predict the future") instead of specific to this model. -->

| ID | Type | Statement | Support | Effect if wrong | Mitigant |
|---|---|---|---|---|---|
| ASM-01 *(Example)* | Assumption | The 2010 to 2024 period represents the likely range of default rate variability | Includes two stress periods | PD under-estimated in a more severe downturn | Margin of conservatism; downturn tested in the stress testing programme |
| ASM-02 *(Example)* | Expert judgement | Loans under 3 months on book receive grade 4 | Too little behaviour to score; grade 4 matches early default experience | Mis-grading of new loans | Monitoring of early defaults by origination month |
| LIM-01 *(Example)* | Limitation | Few observations in sustained high-rate periods | Data only since 2022 | Affordability effects may be understated | Debt service variable; monitoring of DSTI segment; recalibration trigger |
| LIM-02 *(Example)* | Limitation | Buy-to-let has 4% of observations | Small segment | Wider estimation error for buy-to-let | Separate monitoring; category C conservatism |
| LIM-03 *(Example)* | Limitation | Model not designed for lifetime PD | Purpose statement | Misuse in IFRS 9 if used directly | Use restriction in inventory and model card |

Carry the limitations to the Model Card (`ai-data/model-card`) and the
inventory entry, so that users see them before relying on the output.

## 9. Implementation

<!-- Guidance: Show that the production model matches the developed model. Most implementation errors come from data mapping, not from coefficients. Pitfall: testing the scoring code with the development dataset only, missing differences in production data feeds. -->

Describe the production platform, the data feeds, the scoring frequency
and the controls on changes. Record the implementation tests and their
results.

| Test ID | Test | Result *(Example)* |
|---|---|---|
| IMP-01 | Parallel run: production scores against development code on the same 50,000 loans | 100% of grades identical; maximum PD difference below 0.0001 |
| IMP-02 | Production data feed mapping for each risk driver | All six drivers reconciled; one code mapping corrected for employment type |
| IMP-03 | Edge cases: missing values, new products, extreme values | Defaults applied as documented in section 4.4 |
| IMP-04 | User acceptance by underwriting and pricing | Signed off 2026-02-20 |
| IMP-05 | Access and change controls on the rating engine | Changes require dual approval; logs retained |

## 10. Ongoing monitoring plan

<!-- Guidance: Define what will be measured, how often, the thresholds and who acts. The plan must exist before the model goes live. Pitfall: thresholds so loose that nothing ever breaches. -->

| Metric | Frequency | Green | Amber | Red | Owner |
|---|---|---|---|---|---|
| Gini on rolling 12-month defaults *(Example)* | Quarterly | At least 0.55 | 0.50 to 0.55 | Below 0.50 | Head of retail credit risk |
| PSI of score distribution *(Example)* | Monthly | Below 0.10 | 0.10 to 0.25 | Above 0.25 | Modelling team |
| Grades with z beyond 1.96 *(Example)* | Annual | 0 or 1 | 2 | 3 or more | Modelling team |
| Override rate *(Example)* | Monthly | Below 5% | 5% to 10% | Above 10% | Underwriting |
| Early defaults (within 12 months of origination) *(Example)* | Quarterly | Below 0.30% | 0.30% to 0.50% | Above 0.50% | Head of retail credit |
| Missing input rate *(Example)* | Each run | Below 1% | 1% to 3% | Above 3% | Data owner |

A red result triggers investigation within the period set in the Model
Risk Management Policy (`finance/model-risk-management-policy`) and a
documented response: no action with reasons, overlay, recalibration or
redevelopment. Use the Model Monitoring Plan
(`ai-data/model-monitoring-plan`) for the detailed design, data
sources and report layout.

## 11. Regulatory requirements traceability

<!-- Guidance: Map each applicable requirement to the section and evidence that meets it. Cite numbers and names only and obtain the source text. Pitfall: claiming compliance with a requirement that the document does not actually evidence. -->

| Requirement | Source | Section | Evidence |
|---|---|---|---|
| Clear statement of purpose and use | SR 26-2, model development | 2 | Purpose table, user workshops |
| Developmental testing including out-of-sample and out-of-time | SR 26-2, model development | 7 | Discrimination, stability and calibration tables |
| Documented limitations | SR 26-2, model use; PRA SS1/23 Principle 3 | 8 | Assumptions and limitations register |
| Implementation and use controls | PRA SS1/23 Principle 3 | 9 | Implementation tests IMP-01 to IMP-05 |
| Default definition | CRR Article 178; EBA Guidelines on the definition of default as amended | 4.2 | Default rule and re-derivation |
| Long-run average and margin of conservatism | EBA/GL/2017/16 | 6 | Calibration and MoC table |
| Validation of internal estimates | CRR Article 185 | 10, 7 | Monitoring plan and testing results provided for validation |
| Model use and internal validation expectations | ECB guide to internal models (revised 28 July 2025), overarching principles and credit risk chapters (definition of default, PD estimation) | 2, 7, 10 | As above |

## 12. Adaptations for market risk models

<!-- Guidance: Use this section instead of sections 4 to 7 detail when the model is a value-at-risk, expected shortfall or pricing model. Pitfall: documenting a VaR model as if it were a credit model, without risk factor coverage or P&L analysis. -->

For a market risk model, replace the credit-specific content as
follows, keeping the same section structure.

- **Purpose and scope:** desks, books and products covered; risk
  measure (for example, one-day 99% VaR or 97.5% expected shortfall),
  holding period and use (limits, own funds, or both).
- **Data:** risk factor time series, sources, cleaning, proxies for
  missing factors, and the observation window and weighting.
- **Methodology:** historical simulation, Monte Carlo or parametric
  approach; full revaluation or sensitivities; scaling from one day to
  longer horizons, with the justification.
- **Testing:** back-testing of actual and hypothetical P&L against the
  risk measure, P&L attribution where required, risk factor coverage
  analysis and stress period selection. Back-testing for the Basel
  framework is set out in chapter MAR32 (cited by name) and, for EU
  institutions, in the CRR; record live results in the VaR Backtesting
  Report (`finance/var-backtesting-report`).
- **Limitations:** risks not in VaR, illiquid positions, proxy risk
  factors, and their treatment under the Market Risk Limit Framework
  (`finance/market-risk-limit-framework`).

*Example:* A one-day 99% historical simulation VaR uses 500 days of
equally weighted risk factor returns. Over 250 trading days, the
expected number of exceptions at 99% is 2.5; observing four exceptions
would be within the range the developer should anticipate and explain,
not by itself proof of a defect.

## 13. Glossary and references

<!-- Guidance: Define statistical terms once for non-specialist reviewers such as committee members. -->

| Term | Meaning |
|---|---|
| PD | Probability of default over a stated horizon, usually one year |
| LGD | Loss given default, the share of exposure lost if default occurs |
| EAD | Exposure at default |
| AUC | Area under the receiver operating characteristic curve |
| Gini | Discrimination measure equal to twice the AUC minus one |
| PSI | Population stability index |
| Long-run average default rate | Average of observed annual default rates over a period that reflects the likely range of variability |
| Margin of conservatism | Add-on to an estimate to cover identified deficiencies and estimation error |
| Out-of-time test | Test on data from a period after the development sample |
| Rating philosophy | How far grades respond to the economic cycle (point-in-time) or remain stable (through-the-cycle) |
| VaR | Value at risk, a loss level not expected to be exceeded at a stated confidence over a stated horizon |

References:

- Board of Governors of the Federal Reserve System, FDIC and OCC,
  Revised Guidance on Model Risk Management (SR 26-2, OCC Bulletin
  2026-13, FDIC FIL-15-2026), 2026-04-17 (public domain).
- Prudential Regulation Authority, SS1/23 Model risk management
  principles for banks, effective 2024-05-17 (cited by name and
  principle number only).
- European Central Bank, ECB guide to internal models, revised
  version published 28 July 2025 (replaces the February 2024 version).
- Regulation (EU) No 575/2013 as amended by Regulation (EU) 2024/1623,
  Articles 178 and 185. EUR-Lex.
- European Banking Authority, Guidelines on PD estimation, LGD
  estimation and treatment of defaulted assets (EBA/GL/2017/16,
  published 2017-11-20, applicable from 2021-01-01).
- European Banking Authority, Guidelines on the application of the
  definition of default (EBA/GL/2016/07), as amended by EBA/GL/2026/05
  (applicable from
  2026-10-19).
- Basel Committee on Banking Supervision, Basel Framework chapter MAR32
  (back-testing), cited by name only.

## Standard and licence note

Based on public-domain US federal guidance (SR 26-2, OCC Bulletin
2026-13, FDIC FIL-15-2026) and EU public law and guidelines
(Regulation (EU) No 575/2013 as amended, EUR-Lex; EBA guidelines, used
with attribution to the European Banking Authority). Structure also
aligned with PRA SS1/23 and the ECB guide to internal models, cited by
name and number only, and the Basel Framework, cited by chapter name
only; no text from these sources is reproduced. All text, formulas
presentation and worked examples are original; every number in the
example has been recomputed. Fernhollow Bank and its model are
fictional. Check the current official version of each guidance and
regulation before use. Template licence: MIT.