Understanding Loan Loss & Credit Modeling

Introduction

Quick takeaway: Loan-loss and credit modeling estimate expected losses to size reserves, set price, and steer portfolio decisions. You need a repeatable model that translates borrower behavior and macro scenarios into a single allowance number you can defend-so the model must map PD (probability of default), LGD (loss given default), and exposure to default across scenarios. Provisioning matters because it directly hits earnings, regulatory capital, and investor trust; get it wrong and you misprice risk, tie up capital, or invite scrutiny. Here's the quick math: for a $10 billion loan book at an 0.7% expected loss you need about $70 million allowance (if EL rises to 1.5%, allowance moves to $150 million); what this estimate hides is timing, concentration, and LGD volatility. Next step: Model owner-run a two-year scenario backtest and deliver a defendable allowance and sensitivity table by Friday, so Finance can sign off; you can defintely use this to start.

Key Takeaways

• Build a repeatable model that maps PD, LGD, and EAD across macro scenarios to produce a defensible Expected Credit Loss (ECL) allowance.
• Provisioning materially affects earnings, regulatory capital, and investor trust-misstating allowance misprices risk and ties up capital.
• Size sensitivity matters: e.g., a $10B book at 0.7% EL → $70M allowance (at 1.5% → $150M); timing, concentration, and LGD volatility drive big swings.
• Immediate actions: assemble 24-36 months vintage data, build baseline PD/LGD, run three macro scenarios, and perform a two‑year backtest to deliver a defendable allowance and sensitivity table quickly.
• Governance required: versioning, documentation, independent validation, backtesting and stress‑testing; avoid overfitting, weak macro mapping, and data leakage.

Understanding Loan Loss & Credit Modeling: key concepts and metrics

You need a repeatable lift-from-data model that turns borrower behavior and macro scenarios into a single allowance number - fast and defensible. Here's the short takeaway: Expected Credit Loss (ECL) is the reserve output; it equals PD × LGD × EAD, applied across cohorts and scenarios. Build from clean loan-level data, validate by vintage, and stress with clear macro mappings.

Expected Credit Loss and Probability of Default

Expected Credit Loss (ECL) is the accounting and risk-reserve amount you book to cover forecasted losses. It's forward-looking and scenario-weighted. Probability of Default (PD) is the chance a borrower defaults within a defined window (monthly, annual, lifetime). Keep definitions consistent with accounting policy (e.g., 12-month vs lifetime horizon).

Steps to implement:

• Define default event clearly
• Choose horizon: 12-month or lifetime
• Segment borrowers by risk and product
• Estimate PD per segment and scenario
• Aggregate to portfolio ECL

Practical numbers and a quick math example: for a FY2025 retail book with an outstanding balance of $500,000,000, if segment PD = 2% (annual) and weighted LGD × EAD in next section equals 40%, ECL ≈ $4,000,000. Here's the quick math: $500,000,000 × 2% × 40% = $4,000,000. What this estimate hides: timing of default, prepayments, and scenario weights.

Best practices and checks:

• Backtest PDs by vintage quarterly
• Use survival analysis for time-to-default
• Cap PDs where data is sparse
• Document overrides and reason codes

Loss Given Default and Exposure at Default

Loss Given Default (LGD) is the percent of exposure you expect to lose after recoveries and workout costs. Exposure at Default (EAD) is the outstanding balance at the time of default, including drawn balances plus credit line utilization assumptions (conversion factors).

Practical estimation steps:

• Segment by product and collateral
• Calculate static and cure-adjusted recoveries
• Apply haircuts to collateral values under scenarios
• Estimate EAD using utilization curves
• Stress LGD/EAD under adverse macros

Common numeric guidances: unsecured retail LGD often sits between 50%-80%; secured loans after repossession often 20%-40%. For FY2025 modeling, assume credit-card EAD conversion factor ~75%, HELOCs ~50%, term loans ~100% unless prepayment evidence suggests otherwise. Example: if drawn balance $200,000, expected utilization adds $50,000, EAD = $250,000; with LGD 60%, loss = $150,000.

Best practices and caveats:

• Reconcile recoveries with accounting write-offs
• Separate unsecured vs secured LGDs
• Model cure (partial recovery) explicitly
• Avoid optimistic collateral valuations - stress by >=10%

Vintage, roll rates, and cure rates

Vintage analysis groups accounts by origination period to track lifecycle performance. Roll rates measure migration between delinquency buckets (current → 30 → 60 days). Cure rates are the share of delinquents that return to performing status. These lifecycle metrics are critical for PD timeliness and early-warning triggers.

How to build and use vintages:

• Assemble 24-36 months of monthly vintages
• Compute cohort balances and charge-offs
• Derive monthly roll matrices
• Estimate forward PD from roll chains
• Use cure rates to reduce lifetime default exposure

Quick reproducible example: take a Jan-2023 vintage with $10,000,000 originations. If 3% roll to 30-day in month 1 and 20% of those cure, projected 12-month default path comes from chaining monthly roll rates into a cumulative PD. Here's the quick math: chain roll probabilities to get cumulative default; then apply LGD and EAD. What this hides: seasonality, origination quality shifts, and censoring from payoffs - defintely watch for left-censoring in short-lived products.

Validation and operational tips:

• Backtest vintages to realized charge-offs
• Flag cohorts with sudden roll-rate spikes
• Embed early-warning triggers from roll matrices
• Include reconciliations owner and cadence

Data and inputs for loan-loss and credit modeling

You need a repeatable feed of loan-level history, borrower attributes, and macro scenarios to turn borrower behavior into a defensible allowance (ECL). Takeaway: assemble a canonical data set covering at least 24-36 months, refresh monthly, and map three forward scenarios (baseline, adverse, severe) to PD/LGD shifts before modeling.

Core loan-level and borrower data

You're building models that predict future losses from past behavior, so capture the full loan lifecycle and borrower profile at the transaction level. Start by enforcing a single, unchanging loan identifier (master loan ID) and a monthly snapshot cadence that preserves historical balances and status.

Collect these canonical fields:

• origination date
• original and current outstanding balance
• term and remaining term
• coupon / effective interest rate
• amortization schedule
• payments: dates, amounts, and days past due buckets (30/60/90+)
• prepayments, charge-offs, recoveries
• collateral value and appraisal date
• loan purpose and product type

For borrower attributes capture credit score, income (stated or verified), debt-to-income (DTI), employment tenure, residence zip, and documented assets. Keep origination-time snapshots plus updated credit score and balance each month to measure drift. One-liner: canonical ID and monthly snapshots beat messy joins every time.

Practical steps: (1) Ingest raw feeds and build a monthly pivot table keyed by master loan ID; (2) store both flow events (payments, charge-offs) and snapshots; (3) retain pre-charge-off histories for at least 36 months to support vintage backtests.

Macroeconomic scenarios and mapping

Models must be forward-looking, so pick scenario variables investors and regulators expect: GDP growth, unemployment rate, short- and long-term interest rates, housing price index where relevant. Use three scenarios: baseline (consensus), adverse (stress), and severe (tail).

Steps to build and map scenarios:

• source scenarios monthly from public (Fed, Treasury) or consensus vendors
• select key drivers per portfolio (unemployment for consumer, GDP and rates for corporate)
• estimate PD/LGD elasticity to each driver using historical regressions by vintage
• apply scenario-paths to base PD/LGD, then weight scenarios for expected ECL

Here's the quick math: if base PD = 1.0% and unemployment elasticity = +20% per +1pp, a +2pp unemployment shock raises PD ≈ 1.0% × (1 + 0.20×2) = 1.4%. What this estimate hides: nonlinearity at tails and correlated LGD increases; calibrate with stress regressions and expert overlays.

Practical guardrails: document sources and dates for each scenario, run sensitivity bands (±1pp), and keep a reproducible mapping table so governance can review how each macro input moves PD and LGD. One-liner: scenarios without documented mappings are just stories.

Quality checks, governance, and data hygiene

Data quality failures make any model meaningless. Establish automated checks at ingest and before model training: field completeness, valid ranges, reconciliations to general ledger (GL), and time-order integrity to defintely avoid data leakage.

Essential QC rules:

• require > 95% completeness for key fields (balance, status, PD drivers)
• flag duplicates and missing master loan IDs
• ensure no future-dated payment entries in training windows
• reconcile monthly outstanding balances to GL within 0.5%
• apply censoring rules for live accounts and seasonality adjustments (holiday payment patterns)

Validation steps: (1) run automated data profiling and produce a data quality dashboard; (2) backfill missing historic values using amortization schedules or origination snapshots where documented; (3) freeze training data and log versions before model runs; (4) maintain a lineage table showing source system, extraction timestamp, and any transformations.

Practical considerations: treat new vintages as left-censored until they age; require independent sign-off for any manual overrides; and schedule quarterly data revalidations. One-liner: if you can't trace a field to its source, don't use it in the model.

Modeling approaches

You need a repeatable, auditable way to turn borrower behavior and macro scenarios into a forward-looking allowance number; the direct takeaway: match model complexity to use-cohorts for transparency, statistical models for calibrated PDs, and machine learning where nonlinearity matters, then force-forward scenarios into weighted ECLs.

Your situation: you must produce an allowance that regulators and auditors trust while giving traders and origination teams clear signals. Below I show practical steps, quick math, and governance actions to get that done using FY2025 vintage data and sample numbers.

Cohort-based roll-rate models

What it is: cohort models follow origination cohorts over time and measure transitions (current → 30 → 60 → 90+ days). They convert observed roll rates into short-horizon PDs and are the most transparent retail approach.

Steps to build

• Assemble cohort table covering 24-36 months of originations and balances, by product, vintage, and score band.
• Compute monthly transition matrix: probability of moving from bucket i to j in one month.
• Aggregate to horizon PDs (e.g., 12-month PD) by chaining monthly transitions.
• Apply cure rates (reaging) and right-censoring for recent vintages.
• Map to EAD and LGD to get ECL: ECL = Σ cohort balance × PD × LGD.

Best practices

• Segment by product and underwriting vintage.
• Control for seasonality (holiday payments, tax months).
• Backfill missing payment history carefully; defintely avoid data leakage from forward-looking fields.
• Benchmark LGD from recovery curves; for unsecured retail in FY2025 use an LGD range like 40-60% depending on collections performance.

Quick math example (FY2025 sample): a cohort with opening balance $100,000,000, observed 12‑month default rate 3.2%, and LGD 45% gives expected loss ≈ $1,440,000 (100,000,000 × 3.2% × 45%).

When to use it: retail portfolios with stable behavior and limited covariates. One-liner: cohort models trade sophistication for clarity.

Statistical PD/LGD models and machine learning

What they are: statistical models (logistic regression, survival analysis) estimate PD/LGD with clear parameters; machine learning (gradient boosting, GBM) captures nonlinearities but needs controls for interpretability and overfit.

Statistical modeling steps

• Define target: default within T months (binary) or time-to-default (survival).
• Choose features: score, LTV, DTI, term, payment history, vintage, and macro lags.
• Estimate with logistic regression or Cox proportional hazards for time-to-event; include regularization (L1/L2) and interaction terms sparingly.
• Calibrate model outputs to observed default rates (Platt or isotonic calibration) so model PDs map to absolute historical PDs-for FY2025 calibration anchor use most recent 12‑month vintage default rate.

Machine learning practical guidance

• Use GBM (e.g., XGBoost) for nonlinear effects but force monotonic constraints where economically required (score ↑ → PD ↓).
• Validate with k‑fold CV, time-split backtests, and out-of-time holdouts (last 6-12 months of FY2025 vintages).
• Produce explainability outputs (SHAP values) and monotonic feature checks for governance.
• Metrics to report: AUC (aim for 0.75-0.85), calibration slope (~1), and Brier score.

LGD modeling specifics

• Model recoveries on realized charge-offs using hurdle models or beta regressions; include cure and resale timing.
• Separate collateralized vs unsecured: FY2025 mortgage LGD example ranges 20-35%, unsecured 40-65% depending on collections.

Operational controls

• Document feature engineering and avoid leakage (no future payment info in inputs).
• Version models, freeze code, and require independent model validation.
• Translate relative model scores into portfolio-level PD curves used for provisioning and pricing.

One-liner: use statistical models for calibrated PDs and add ML only with strict guardrails.

Scenario weighting and forward-looking overlays

What it is: map macroeconomic scenarios (baseline, adverse, upside) into PD and LGD shifts, then weight scenario outcomes to produce a single ECL number consistent with accounting rules (IFRS 9 / CECL-style forward-looking requirements).

Steps to implement

• Define at least three macro scenarios with explicit paths for GDP, unemployment, and rates for FY2025-2027.
• Estimate macro-to-PD mapping via regression (panel or cross-sectional), or build stress multipliers: delta PD = f(Δunemployment, ΔGDP).
• Apply scenario PDs and LGDs to portfolio buckets, compute ECL per scenario, then weight scenarios to get final ECL.
• Document scenario weights and rationale; common starting weights are 60% baseline / 30% adverse / 10% upside but adjust to governance and regulator guidance.

Example calculation (FY2025 sample)

• Baseline PD for cohort = 2.0%.
• Adverse scenario unemployment shock maps to PD multiplier = 1.8x → adverse PD = 3.6%.
• Upside multiplier = 0.6x → upside PD = 1.2%.
• Weighted PD = 0.6×2.0% + 0.3×3.6% + 0.1×1.2% = 2.28%; plug into ECL formula with EAD and LGD to get weighted expected loss.

Best practices and guards

• Ensure macro mapping is economically sensible and stable across vintages; run sensitivity on multipliers.
• Avoid double-counting: don't apply blind overlays on top of models that already incorporate macro variables unless explicit gap analysis supports it.
• Keep scenario design and weights under Risk committee control and refresh quarterly during FY2025 monitoring.

Governance and next steps

• Risk: define three scenario paths and PD multipliers and deliver mapping table by Friday, 2025-12-12.
• Finance: run weighted ECLs and produce a 13-week cash and P&L impact report by Friday, 2025-12-19.

One-liner: tie macro paths to PD multipliers, weight them, and lock scenario governance promptly to avoid ad-hoc overlays.

Validation, governance, and regulatory expectations

You're running a credit model that must survive auditors, regulators, and board scrutiny - so validate it, govern it, and prove it ties to capital. The direct takeaway: build repeatable backtests, tie stress scenarios to capital planning, and formalize model risk controls that catch drift early.

Backtesting - compare predicted vs. realized losses by vintage

Start by aligning model outputs to realized outcomes on the same definition of default and timing; mismatched definitions cause false alarms. Use vintage cohorts (origin month/year) and track performance for at least 24-36 months where possible.

Practical steps:

• Assemble cohort table by origination month and month-on-book.
• Compute cohort metrics: observed default rate, observed LGD, cumulative loss.
• Compare predicted ECL to realized losses at matching horizons (12m, 24m).
• Report discrimination (AUC/Gini) and calibration (Brier, calibration plots).
• Use statistical tests (e.g., Hosmer-Lemeshow) and confidence intervals for small samples.

Best practices:

• Require minimum sample: prefer portfolios with > 1,000 accounts or > 200 defaults for stable bucket estimates; flag small-sample cohorts.
• Hold out an out-of-time validation (OOTV) window and re-run monthly/quarterly.
• Adjust for censoring and prepayments; align exposure at default (EAD) methodology.
• Log all backtest discrepancies and require root-cause within 30 days for material misses.

One clean line: Backtest by vintage, not by calendar, so you compare apples to apples.

Stress testing - link to capital planning (ICAAP / CCAR where applicable)

Stress testing translates macro shocks into PD/LGD/EAD changes and then into capital impacts. For provisioning use a 12‑month forward view; for regulatory capital link to supervisory horizons (for example, CCAR-style runs over multiple quarters).

Operational steps:

• Define scenarios: baseline, adverse, severely adverse (document assumptions and paths for GDP, unemployment, rates).
• Estimate elasticities: map 1 percentage point change in macro to PD/LGD changes by segment (create segment-level sensitivity tables).
• Run portfolio ECL under each scenario; produce CET1 / capital ratio impact for the planning horizon.
• Do reverse stress testing: find macro breakpoints that breach capital or liquidity triggers.
• Weight scenarios for accounting where required; keep scenario weights auditable.

Practical calibration tips:

• Derive macro mappings from historical episodes where possible; supplement with expert overlays where data is thin.
• Apply severity caps (floor/ceil) to avoid implausible extrapolations.
• Document judgmental overlays and require Finance plus Risk sign-off; track who changed what and why.

One clean line: Map macro movements to model inputs, then map model outputs to capital - every link must be auditable.

Model risk management, documentation, and common failures

Model risk management (MRM) is the control framework that keeps models reliable. Key pillars are inventory, versioning, documentation, independent validation, monitoring, and escalation rules.

Concrete controls to implement now:

• Maintain a model inventory with risk rating, owner, last validation date.
• Version-control code and data pipelines (tag releases with date and change log).
• Produce a Model Governance File: purpose, assumptions, data lineage, performance metrics, limitations, and remediation plan.
• Schedule independent validation annually for high-risk models and on material change; validators must test code, inputs, backtests, and stress mappings.
• Set monitoring cadence: monthly for performance metrics, quarterly for full reconciliation to GL/accounts.
• Define escalation triggers: PSI > 0.25, realized losses > predicted by > 20% on a rolling 12m, or a material change in macro elasticity.

Common failures and how to avoid them:

• Overfitting - prevent with out-of-time validation and parsimony in features.
• Weak macro mapping - prevent by anchoring elasticities to historical episodes and documenting expert judgment.
• Data leakage - defintely avoid including post-default information in PD training; enforce strict feature cut dates.
• Undocumented overrides - require written rationale, quantitative impact, and sign-off whenever a manual override changes reserve levels by more than 10%.
• Stale parameters - automate a drift check and require re-estimation when triggers fire.

Next step (owner action): Risk: schedule independent validation kickoff within 6 weeks; Finance: deliver cleaned 24-36 months vintage dataset by next Friday.

One clean line: Treat MRM as repeatable plumbing - if it's manual, undocumented, or hit‑or‑miss, it will fail an auditor or regulator.

From model output to business action

You need a repeatable path from model ECL to decisions - reserves booked, prices set, portfolio limits changed, and operations run. Here's the direct takeaway: translate ECL into accounting entries and capital impact, convert risk lifts into pricing and origination moves, and turn signals into concrete monitoring and reconciliations so you can act before losses compound.

Reserves and P&L

You're translating a modeled Expected Credit Loss (ECL) into a reserve number and an earnings hit. Start with a clear mapping to accounting: under US GAAP CECL or IFRS 9, the ECL (12‑month or lifetime as required) becomes the allowance for credit losses (a contra-asset) and the offset posts to provision expense, lowering pre-tax income and retained earnings.

Step-by-step:

• Calculate portfolio ECL by segment: ECL = Balance × PD × LGD × EAD adjustments.
• Post journal: Debit Provision Expense; Credit Allowance for Credit Losses.
• Apply tax effect and update retained earnings; quantify CET1/regulatory effect.

Quick example math so you can see the hit: a loan book of $5,000,000,000 with weighted PD 1.8% and LGD 40% yields ECL = $36,000,000 (5bn × 1.8% × 40%). If pre-tax income was $200,000,000, provision reduces it to $164,000,000. What this estimate hides: timing of cashflows, collateral recoveries, and discounting assumptions.

Regulatory capital example: with RWAs of $40,000,000,000 and CET1 = 12% (capital = $4,800,000,000), a $36,000,000 provision lowers CET1 to $4,764,000,000, moving the ratio to roughly 11.91% (~9 bps drop). Make those bps visible on your capital plan.

Best practices:

• Version ECL inputs and publish model changes; defintely avoid data leakage.
• Publish sensitivity tables (PD ±50bps, LGD ±5pp) and scenario-weight breakdowns.
• Backtest by vintage monthly and disclose material overlays with approval.

Pricing and origination policy and portfolio management

Use the model to price new originations and to tilt portfolio exposures. Translate ECL into a required spread, then layer in cost of funds and target return to set offered rates and eligibility.

Pricing steps and example:

• Compute expected loss for the borrower: PD × LGD (example PD 3.0%, LGD 45% → ECL = 1.35%).
• Set required spread = cost of funds + operating cost + target return + ECL. Example: cost 2.00% + ops 0.50% + target 4.00% + ECL 1.35% = required rate 7.85%.
• Compare to market elasticity; if price exceeds demand, tighten credit (score, LTV) instead.

Origination policy levers:

• Tighten score cutoffs or add additional covenants.
• Reduce max LTV or shorten amortization to cut LGD and EAD.
• Apply credit overlays for high-correlation sectors (energy, CRE).

Portfolio management actions tied to model outputs:

• Set concentration limits by borrower, sector, or geography (e.g., cap single-borrower exposure ≤ 10% of tier-1 capital or ≤ 2% of portfolio balance).
• Trigger workouts when vintage NPL exceed thresholds (e.g., 90+ day delinquency > 3% for two months).
• Consider securitization or whole-loan sales for distressed cohorts above a size threshold (example: sell pools > $50,000,000 to remove risk-weighted assets and recover liquidity).

What to watch: over-pricing kills originations; under-pricing creates compounding losses. Run simple take-rate elasticity tests and a break-even price table for each credit segment.

Operations

Operationalize model outputs with a clear cadence, concrete triggers, and owner responsibilities so the model actually changes behavior.

Monitoring cadence and controls:

• Daily: collections and cure-rate dashboard (top 50 accounts).
• Weekly: PD and roll-rate deltas, exception list to collections and underwriting.
• Monthly: full ECL run, GL reconciliation, and provisioning entries.
• Quarterly: model revalidation, governance committee review, and macro scenario refresh.

Early-warning triggers (examples):

• Trigger A: 30-day delinquency roll-rate increases > 200 bps versus prior month → escalate to Collections Head.
• Trigger B: Model PD for a cohort rises > 50% from baseline → pause new originations in that cohort.
• Trigger C: Macro unemployment delta > 100 bps in adverse scenario weight → implement contingency pricing and loss mitigations.

Reconciliation and ownership:

• Owner: Risk runs the ECL model and publishes inputs weekly.
• Owner: Finance posts provisions monthly and owns GL reconciliation between model allowance and accounting balance.
• Owner: Operations implements collection actions and reports cure/roll metrics to Risk.

Concrete next step and owner: Risk: assemble 24-36 months vintage data and deliver baseline PD/LGD by Friday; Finance: run the initial ECL and post the provisioning timeline within 6 weeks.

Conclusion

You're building a repeatable ECL (expected credit loss) process; the direct takeaway: assemble the right vintages, build baseline PD/LGD, run three forward-looking scenarios, and deliver a first ECL and backtest in 4-6 weeks. You need clear owners and short feedback loops so reserves, pricing, and capital reflect current risk.

Core next steps: assemble data, build baseline PD/LGD, run scenarios

Start with the borrower-level and loan-level history that covers the last performance cycle so your model learns real behavior. Aim for 24-36 months of vintage data (for example, Jan 1, 2023-Dec 31, 2025 if you use calendar-year FY2025), with cohort IDs, origination date, current balance, contractual term, delinquency tags, collateral values, and payment history.

Practical steps:

• Pull: raw loan file, payment ledger, and charge-off records.
• Map: credit score, income bands, and collateral buckets to each loan.
• Clean: backfill missing months, flag censored accounts, and remove data leakage (defintely check timestamp joins).
• Aggregate: create vintage cohorts by origination month and compute roll rates, cure rates, and cumulative default counts.
• Estimate: build baseline PDs using vintage roll-rate or logistic models and LGD by workout data or collateral haircuts.
• Document: inputs, transformations, and assumptions in a single model spec.

Here's the quick math: ECL per loan = PD × LGD × EAD; aggregate to get your reserve number. What this estimate hides: time-to-default, unsecured vs. secured splits, and forward macro sensitivity-so tag cohorts accordingly.

Short timeline: deliver initial ECL and backtest within weeks; assign owners

Commit to a tight sprint: an initial, auditable ECL and vintage backtest in 4-6 weeks. That gives you a working model for accounting and risk decisions while leaving room for refinement.

Week-by-week sample plan:

• Week 1 - Data pull and reconciliation (Owner: Finance).
• Week 2 - Cohort build and exploratory analytics (Owner: Risk modeller).
• Week 3 - Baseline PD/LGD estimation and one-off checks (Owner: Risk modeller).
• Week 4 - Map macro scenarios and run ECL; produce P&L reserve impact (Owners: Risk + Finance).
• Week 5 - Backtest by vintage, reconcile to charge-offs, and draft model documentation (Owner: Independent validator or second modeller).
• Week 6 - Governance packet and sign-off for provisional reserve booking (Owners: Risk lead and Finance owner).

Assign named owners now: Risk: nominate a lead model owner; Finance: nominate an accounting/reconciliation owner; Compliance: provide independent reviewer. One-liner: pick owners this week and lock the sprint plan.

Watch list: model drift, macro surprises, and data gaps; plan revalidation

Monitor three failure modes: model drift (performance decay), unmodeled macro shocks, and data quality gaps. Set clear thresholds and owners so issues surface before they hit reserves or capital.

• Track: monthly PD vs realized default by vintage; trigger review if PD moves > 20% vs baseline.
• Stress: rerun worst-case macro scenario when unemployment or GDP surprises exceed historical vol-define shock bands (e.g., +200 bps unemployment).
• Data: run automated completeness checks; flag cohorts with > 5% missing critical fields.
• Governance: revalidate models on a quarterly cadence and after any material override or policy change.

One-liner: set automated alerts and a quarterly revalidation calendar so drift never surprises earnings or regulators.

Next step: Risk lead - name the model owner and deliver the model spec by Friday; Finance - reconcile loan-level balances to the general ledger by Friday.