Expected Credit Loss is not only a model output. It is a reporting process that joins source data, staging, model assumptions, forward-looking scenarios, management judgement, approval and disclosure. A weakness anywhere in that chain can make an otherwise reasonable allowance difficult to reproduce or defend.
An audit-ready process does not wait until year-end to assemble evidence. It creates evidence while the work is performed. The following ten controls provide a practical baseline for lenders, NBFCs, banks and corporate finance teams.
At-a-glance control map
| # | Control | Principal risk addressed | Core evidence |
|---|---|---|---|
| 1 | Population completeness | Missing or duplicated exposures | Source-to-model reconciliation |
| 2 | Data quality and mapping | Incorrect attributes or transformations | Validation report and exception log |
| 3 | Segmentation governance | Inappropriate pooling | Approved segmentation rationale |
| 4 | SICR and staging control | Incorrect 12-month or lifetime ECL | Stage movement and override report |
| 5 | Model and parameter versioning | Wrong model or assumptions used | Run manifest and change log |
| 6 | Scenario governance | Unsupported forward-looking effects | Scenario and weighting paper |
| 7 | Overlay governance | Uncontrolled management adjustment | Overlay memo and approval trail |
| 8 | Independent review | Unchallenged error or bias | Maker-checker sign-off |
| 9 | Output-to-ledger reconciliation | Incorrect amount posted | Approved journal bridge |
| 10 | Disclosure and archive control | Inconsistent reporting or lost evidence | Disclosure tie-out and period archive |
1. Population completeness and reconciliation
The first question is not whether the model is sophisticated. It is whether every in-scope exposure entered the process once—and only once.
Reconcile the model population to controlled source totals such as the lending subledger, receivables ledger, trial balance and off-balance-sheet records. Differences should be classified, investigated and approved. The reconciliation should cover both count and value, with special treatment for accrued interest, write-offs, guarantees, commitments and manual additions.
Evidence to retain: source extracts, control totals, reconciliation, reconciling-item explanation and reviewer sign-off.
2. Data quality and mapping validation
Completeness does not establish accuracy. ECL depends on fields such as origination date, maturity, days past due, internal rating, collateral, product, geography and restructuring status. Missing or mis-mapped attributes can change segmentation, stage and loss measurement.
Validation should combine schema checks, acceptable-value tests, date logic, duplicate detection, outlier rules and period-on-period reasonableness checks. Exceptions should not disappear through manual overwriting; they should be logged with ownership and resolution.
Evidence to retain: validation results, mapping specification, exception log and approved workaround.
3. Segmentation approval and monitoring
Collective ECL assumes that exposures grouped together share meaningful credit-risk characteristics. Segmentation therefore requires governance, not merely a convenient filter.
Document the risk rationale for each segment and test whether pooling remains appropriate as products, underwriting, borrower mix and economic conditions change. Sparse segments and rapidly changing books may need special treatment.
Evidence to retain: segmentation policy, mapping table, population analysis, change assessment and approval.
4. SICR, staging and override control
Stage allocation changes the measurement horizon. It deserves a dedicated control over quantitative thresholds, qualitative indicators, delinquency backstops, cure rules and overrides.
Review stage movements at exposure and portfolio level. Investigate unusual migrations, stable Stage 2 populations during visible stress, and large manual overrides. An override should record the system result, revised conclusion, rationale, support, owner, approver and review date.
Evidence to retain: stage movement report, rule results, watchlist/restructuring inputs, override register and approval.
5. Model, code and parameter version control
The institution should be able to identify exactly which data snapshot, model version, parameter set and configuration produced the approved result.
A controlled run manifest should prevent draft models, stale assumptions or unapproved code from entering production. Changes should be tested, impact-assessed and approved before use. Emergency changes need an explicit exception route and retrospective review.
Evidence to retain: model inventory, version identifier, deployment approval, run log, parameter file and change record.
6. Forward-looking scenario governance
ECL should reflect reasonable and supportable forward-looking information, not a single unchallenged forecast. Scenario design should address relevant variables, forecast horizons, probability weights, non-linearity and consistency with the institution's broader planning view.
The control should separate preparation from approval and show why the selected scenarios remain relevant. A mechanically repeated scenario set can be as weak as an arbitrary last-minute change.
Evidence to retain: macroeconomic paper, variable selection rationale, model linkage, weights, sensitivities, challenge and approval.
7. Management overlay control
Overlays may be necessary where model outputs do not capture a material emerging risk or a known limitation. Their judgemental nature makes them a high-risk control point.
Every overlay should define the risk, affected population, reason the base model is insufficient, method, amount, sensitivity, double-counting check, approval and exit criteria. The overlay should have an owner and expiry or reassessment date.
Evidence to retain: overlay inventory, calculation, supporting indicators, challenge, approval, release criteria and outcome review.
8. Maker-checker review and management challenge
Preparation and approval should not rest with the same person for critical steps. Review must be more than a tick box: it should consider unexpected movements, sensitivity, limitations and consistency with observed portfolio conditions.
Material judgements may require escalation to an impairment or governance committee. Minutes should capture challenges and decisions, not simply attendance.
Evidence to retain: review checklist, comments and resolution, committee papers, minutes, approval status and unresolved actions.
9. Output, journal and movement reconciliation
The final approved model output must reconcile to the amount posted in the ledger. The movement should also be explainable through drivers such as new business, repayments, stage migration, parameter changes, write-offs, recoveries, overlays and foreign exchange.
This control catches stale files, incorrect signs, duplicate journals and late changes that did not pass through approval.
Evidence to retain: model-to-ledger bridge, movement analysis, journal support, preparer/reviewer approval and posting confirmation.
10. Disclosure tie-out and period-end archive
Disclosures should come from the same controlled outputs as the booked allowance. Stage balances, loss-allowance reconciliation, credit-risk narrative, assumptions and sensitivity commentary should agree with approved records.
The final archive should preserve inputs, configurations, results, reconciliations, papers, approvals and journal evidence in a read-only or otherwise controlled form. A repeatable naming and retention standard makes future comparison significantly easier.
Evidence to retain: disclosure tie-out, final financial statement support, archive index, access record and retention confirmation.
How to assess whether a control is genuinely effective
For each control, ask five questions:
- Is the objective clearly defined?
- Is an accountable owner named?
- Is the control performed at the right frequency and level of detail?
- Is there evidence that a reviewer challenged the output?
- Are exceptions tracked through resolution rather than silently cleared?
A process can contain many checklists and still be weak if evidence is retrospective, approvals are perfunctory or exceptions recur without remediation.
Frequently asked questions
Does an audit trail make the ECL model accurate?
No. An audit trail improves traceability and accountability, but model design and performance still require validation and monitoring. Governance and modelling are complementary.
Are all ten controls necessary for a small portfolio?
The principles remain relevant, but execution should be proportionate. A smaller corporate receivables portfolio may use simpler methods and fewer systems while still reconciling data, governing assumptions, separating preparation from approval and preserving evidence.
Can spreadsheets form part of a controlled ECL process?
Yes, but material spreadsheets require controlled access, protected logic, version discipline, input/output checks and review evidence. Risk rises quickly when multiple linked files and manual handoffs become the production environment.
Build evidence before the question is asked
Assess your ECL governance maturity or explore the audit-ready ECL reporting workflow to identify where your current process loses traceability.
Technical references
- IFRS Foundation, IFRS 9 project summary
- IAASB, Auditing Accounting Estimates—ISA 540 (Revised)