IFRS 9 Expected Credit Loss: Models, Staging and Management Judgement

The measurement mechanics

In its standard decomposition, expected credit loss for a period is the product of the probability of default in that period, the loss given default and the exposure at default, discounted to the reporting date. Lifetime expected credit loss sums this across all remaining periods, weighted by survival to each period.

ECL = Σ_t [ PD_t × LGD_t × EAD_t × DF_t ]

The parameters differ from their regulatory-capital equivalents in important respects. IFRS 9 requires point-in-time, unbiased and forward-looking estimates; regulatory PD is often through-the-cycle, and regulatory LGD typically incorporates downturn conservatism. Reusing capital models without recalibration produces estimates that are neither unbiased nor responsive to current conditions.

Staging: where most of the volatility comes from

Exposures move from Stage 1 to Stage 2 when credit risk has increased significantly since initial recognition. That transfer changes the measurement horizon from twelve months to lifetime, which for a long-dated loan can multiply the provision several times over. Small movements in the transfer criterion therefore produce large movements in reported provisions.

• Quantitative criteria are usually based on relative change in lifetime PD since origination, sometimes combined with an absolute threshold.
• Qualitative criteria capture watch-list status, forbearance and other indicators not reflected in the PD.
• The thirty-days-past-due presumption is a backstop, not a primary trigger; relying on it alone defers recognition.
• Calibration of relative thresholds should be tested for the proportion of exposures transferring under different economic conditions, since a threshold that is stable in benign conditions can produce cliff effects in a downturn.

Forward-looking information and scenario weighting

Because the relationship between macroeconomic conditions and credit loss is convex, the expected loss under an average scenario is lower than the probability-weighted average of losses across scenarios. The standard therefore requires an unbiased probability-weighted estimate rather than a single central projection.

1. Define a small number of scenarios — commonly three to five — spanning a range wide enough to capture the convexity.
2. Assign probability weights with a documented rationale, reviewed each period rather than fixed indefinitely.
3. Link macroeconomic variables to credit parameters through estimated relationships, and test those relationships out of sample.
4. Disclose sensitivity: how much the provision changes if the weights shift, since this is often more informative than the headline number.

A recurring weakness is scenario ranges that are too narrow. If the adverse scenario is only mildly adverse, the convexity that motivates the multi-scenario requirement is not captured and the estimate is close to a single-path calculation with extra process.

Post-model adjustments and governance

Overlays exist because models are estimated on history and conditions arise that history does not contain. They are a legitimate response to known model limitations, emerging risks not yet in the data, and data quality issues. They become a governance problem when they persist without a remediation path, when their calculation is undocumented, or when their aggregate size exceeds the modelled component without comment.

• Each overlay should have a stated rationale, a quantification method, an owner and an expected resolution — either incorporation into the model or release.
• Aggregate overlay levels should be reported to the audit and risk committees with movement analysis.
• Overlays should be subject to independent review; they carry the same decision consequence as model output.
• Directional consistency should be checked: an overlay increasing provisions while the underlying models fall requires an explanation beyond prudence.