Tackle Data Quality Issues with These Effective Solutions

Why data quality matters for IFRS 9 and ECL

IFRS 9 requires forward‑looking Expected Credit Loss calculations that are model-driven and evidence-based. Poor data quality undermines the ECL Methodology, leads to material misstatements, and triggers regulatory scrutiny. Even a 2–5% error in key inputs (PD, exposures, collateral values) can change provision levels materially — potentially moving a bank between provisioning categories and affecting capital planning.

Common operational consequences include misallocation between stages (the Three‑Stage Classification), incorrect lifetime vs. 12‑month ECL splits, and unreliable forward scenario overlays. These problems are often rooted in upstream processes; see our guidance on IFRS 9 data shortage for typical root causes and diagnostic steps.

High-quality data supports robust Sensitivity Testing, stronger Model Validation outcomes, and clearer Risk Committee Reports — all necessary to demonstrate governance and defensibility during audits and inspections.

Core concept: What we mean by data quality (definition, components, examples)

Definition and dimensions

Data quality is the degree to which data are fit for purpose. For ECL that means data should be accurate, complete, timely, consistent, traceable and documented. These dimensions enable repeatable PD, LGD and EAD Models and defensible staging decisions under IFRS 9.

Practical examples tied to ECL inputs

• PD inputs: inaccurate default flags or late updates to payment history can understate short‑term PD by 20–50% in some portfolios; missing 30‑90 day past‑due indicators leads to staging errors.
• LGD inputs: incorrect collateral valuations or incomplete recovery costs produce biased LGD estimates — e.g., overstated collateral can reduce LGD by 10–30% artificially.
• EAD inputs: facility limit misreporting or wrong utilisation assumptions inflate EAD; a 5% overstatement in EAD across a retail book can propagate directly to ECL increases.

Data lineage and traceability

Traceability means you can follow a figure from the Risk Committee Reports back to source systems and transformation logic. Without lineage, Model Validation teams struggle to resolve exceptions and auditors will question model outputs during reviews — a common symptom we see when teams neglect ECL data lineage documentation.

Practical use cases and scenarios

Below are recurring situations where data quality interventions create immediate value.

Monthly provisioning run

Scenario: The monthly ECL run produces a provision that swings by 15% month‑on‑month. Diagnostics: compare staging movements, PD changes and EAD inflows; profile missing values and look for changes in upstream feeds. Operational fix: set automated alerts on source‑to‑model reconciliations; apply quick fixes to null collateral values with conservative fallbacks.

New model roll‑out (PD, LGD and EAD Models)

Scenario: New PD model shows improved discrimination but produces materially different lifetime PDs. Diagnostics: perform back‑testing on historical cohorts and run robust Sensitivity Testing. Address data gaps in training sets and document assumptions to support the implementation in Model Validation.

Sensitivity Testing and stress scenarios

Regular Sensitivity Testing highlights model dependence on particular inputs. For portfolios with limited historical defaults, consider scenario‑based overlays and alternative data sources; this is where guidance on Data collection challenges and creative data enrichment becomes important.

Advanced analytics: Using alternative and large datasets

Where appropriate, integrate external signals and machine learning features. Our cluster includes advice on Using big data in ECL for enrichment, but only after careful validation and governance to avoid introducing bias.

Impact on decisions, performance and outcomes

Reliable data quality influences multiple dimensions of bank performance:

• Regulatory compliance: Higher quality data reduces query volume from supervisors and improves Model Validation outcomes.
• Financial reporting accuracy: It reduces unexpected provisioning volatility and the need for late adjustments to financial statements.
• Capital planning: Accurate ECL feeds better economic capital estimates and credit loss forecasts.
• Governance and oversight: The Risk Committee gains more confidence in reports and decisions when staging, sensitivity and scenario assumptions are traceable and defensible.

Achieving these outcomes requires not just tools but people: develop Technical skills for ECL inside risk teams and ensure Model Validation and Data Governance collaborate early in model lifecycles.

Common mistakes and how to avoid them

Mistake 1 — Treating data as an IT problem

Fix: Establish cross‑functional ownership (risk, finance, credit ops) and map critical data elements for ECL by business impact.

Mistake 2 — Insufficient profiling and no baseline metrics

Fix: Run automated data profiling each month and capture baselines (completeness %, null rates, cardinality) before changes.

Mistake 3 — Overfitting models to noisy or incomplete data

Fix: Use conservative feature selection, regularisation and keep a holdout sample for back‑testing; incorporate model governance steps into your ECL Methodology.

Mistake 4 — Ignoring staging sensitivity

Fix: Add staging stability checks and document triggers that move accounts between the Three‑Stage Classification buckets. Include Sensitivity Testing on staging thresholds during Model Validation.

Mistake 5 — Weak documentation and non‑reproducible pipelines

Fix: Implement reproducible ETL pipelines with versioning, store transformation logic in a central repository and align with ECL best practices for audit readiness.

Practical, actionable tips and a checklist

Use this practical checklist when you diagnose or improve data quality for ECL:

1. Data discovery: Inventory all ECL inputs, systems and owners. Map to PD, LGD and EAD Models and to required outputs for Risk Committee Reports.
2. Profiling baseline: Record completeness, accuracy, timeliness and uniqueness metrics for each critical field.
3. Quick triage rules: For null collateral values, apply conservative LGD assumptions by asset class (e.g., +10–20% LGD uplift until validated).
4. Reconciliation layer: Implement source-to-model reconciliations with automated daily or weekly checks and exception dashboards.
5. Validation and testing: Integrate Sensitivity Testing into monthly runs and require model owners to explain drivers for >5% ECL movement.
6. Enrichment plan: Identify priority external sources (credit bureaus, property indexes) and a timeline to onboard — see common ECL data sources.
7. Governance: Update your ECL Methodology to include data quality KPIs and approval thresholds for manual overrides.
8. Continuous monitoring: Implement drift detection for PD distributions, EAD utilisation rates and LGD recoveries; document remediation SLAs.
9. Operationalize best practices: Convert fixes into permanent rules and code; align with broader ECL data practices for consistency across portfolios.

Tip: start with the highest-impact portfolio segment (e.g., corporate exposures >€5m) and scale remediation efforts based on quantified ECL sensitivity.

KPIs & success metrics

• Data completeness rate for critical fields (target: ≥99% for core fields, ≥95% for secondary).
• Null/invalid value reduction rate (month-on-month % reduction target).
• Staging accuracy: percentage difference between automated staging and model-validated staging (target: <2%).
• PD calibration metrics: Population Stability Index (PSI) and Brier score improvements after cleanup.
• LGD predictiveness: RMSE or mean absolute error against realized recoveries (improvement target of 10–20% post-remediation).
• Time to produce Risk Committee Reports (target reduction in report preparation time by 25–40%).
• Number of Model Validation findings related to data quality (target: zero high-severity findings).
• Volume of manual interventions/overrides per run (target: reduce by 50% year-on-year).

FAQ

Reference pillar article

This article is part of our data and ECL content cluster. For a broader treatment of why data is central to ECL models, forecasting risk and complying with IFRS 9, see our pillar guide: The Ultimate Guide: The importance of data in calculating expected credit losses – why data is central to ECL models and its role in forecasting risk and complying with IFRS 9.

Additional resources referenced

Related short reads: ECL data sources, ECL data practices, ECL data, and discussions on Data collection challenges. For governance and validation skillsets, consult Technical skills for ECL, and explore our notes on ECL best practices. If you are assessing external enrichment, see Using big data in ECL.