IFRS 9 Model Validation in Practice

By Jonas Osman Abdelghafour.

IFRS 9 expected credit loss models have now been through enough cycles that the interesting validation questions have shifted. The mechanics of a PD term structure are well understood. What still fails independent challenge is the judgement layer sitting on top: staging thresholds, the forward-looking scenario set, and the overlays that were introduced as temporary and became structural.

Staging and significant increase in credit risk

SICR determination drives the single largest discontinuity in the ECL number — the move from twelve-month to lifetime loss. Validation should test whether the threshold is calibrated to identify deterioration early enough to be meaningful, and whether it behaves sensibly across the portfolio rather than only in aggregate.

Useful tests include migration analysis over time, backtesting whether Stage 2 accounts subsequently default at rates materially above Stage 1, sensitivity of the ECL to threshold movement, and examination of the backstops. Heavy reliance on the thirty-days-past-due presumption is a signal that the quantitative trigger is not doing its job. So is a Stage 2 population that is stable to three decimal places through a changing economic environment — real portfolios do not behave that way.

Forward-looking information

The scenario set and its weights are where a small judgement moves a large number. Challenge should cover whether the scenarios are genuinely distinct in the variables that drive the models, whether the weights are supported by evidence rather than convention, whether the non-linearity between economic conditions and losses is actually captured — the entire purpose of multiple scenarios — and whether the macro variables used are the ones with demonstrated explanatory power for the portfolio rather than the ones the group forecast happens to publish.

A test worth running every year: recompute ECL under equal weights and under alternative severities. If the reported number is highly sensitive to a weighting choice that has thin documentation, that is the finding.

Component models

For PD, the question is whether the point-in-time conversion is genuinely responsive to the cycle and whether the through-the-cycle anchor is stable. For LGD, the recurring weaknesses are recovery data depth, discounting conventions, collateral valuation haircuts and cure-rate assumptions — especially where collateral markets are illiquid. For EAD, the credit conversion factors on revolving exposures are frequently calibrated on periods with no drawdown stress, and behave badly when one arrives.

Across all three, the validator should check the consistency of definitions with the regulatory models. Different default definitions between IFRS 9 and IRB estimates are defensible; undocumented differences are not.

Post-model adjustments

Overlays are the honest response to a model that cannot yet capture a known risk. They become a governance problem when they persist without an owner, a rationale, a quantification method, or a plan to model the underlying driver. Validation should inventory every adjustment, test its calculation, assess whether the driver is now modellable, and report the total as a share of ECL. A book where overlays represent a material fraction of the provision has a model problem that has been reclassified as a judgement.

Monitoring and disclosure

Backtesting should compare predicted against realised losses at a segment level, not just in total, where offsetting errors flatter the result. Disclosure consistency is the last check and an easy one to fail: the sensitivities, scenario weights and staging criteria published in the accounts must match what the models actually do.

A short validation agenda

• Test SICR calibration, backstop reliance and staging stability across segments.
• Recompute ECL under alternative scenario weights and severities; document the sensitivity.
• Challenge LGD recovery data, discounting and collateral haircuts before touching PD methodology.
• Inventory, quantify and time-limit every post-model adjustment.
• Backtest by segment and reconcile disclosures to model behaviour.

Independent challenge on IFRS 9 is most valuable when it stops re-deriving components and starts interrogating the three or four judgements that actually set the provision.

Primary sources: IFRS 9 Financial Instruments (IFRS Foundation); EBA supervisory and risk assessment publications on IFRS 9 implementation.