Reading time ~8 minutes
Ask most finance professionals to explain Expected Credit Loss under IFRS 9 and you will get a formula: PD × LGD × EAD, discounted. Technically correct — and almost entirely beside the point.
ECL is not a calculation. It is a framework. Its reliability rests on the quality of inputs, the logic used to group exposures, and the governance wrapped around management's judgment. In practice, the overwhelming majority of ECL problems trace back to weak data rather than weak mathematics.
Below are the five building blocks that determine whether an ECL framework holds up under scrutiny.
1 — Data quality: the non-negotiable starting point
Every downstream estimate inherits the flaws of the data beneath it. The essentials are unglamorous but decisive: accurate customer and exposure master data, clean aging reports reconciled to the general ledger, and consistent historical series captured on a comparable basis.
The most common failure is subtle — definitions that drift over time. If "default" meant 90 days past due in 2019 and something looser in 2023, any PD calibrated on that history is unreliable. Auditors probe exactly this point.
2 — Segmentation: grouping that reflects real risk behaviour
IFRS 9 requires exposures sharing similar credit risk characteristics to be assessed together. Useful dimensions include retail versus corporate, secured versus unsecured, internal risk grade, and product or geography where these genuinely drive loss behaviour.
The discipline lies in balance. Too few segments and distinct risk profiles are averaged into a blended rate describing no actual borrower. Too many and each segment lacks the observations to support a meaningful estimate.
3 — PD, LGD and EAD: grounding the core inputs
The three core risk parameters and what each must reflect
| Parameter | Requirement |
|---|---|
| PD Probability of default |
Must be point-in-time and forward-looking. Regulatory through-the-cycle PDs are a starting point, not a substitute. |
| LGD Loss given default |
Must reflect actual recovery timelines, enforcement costs, collateral haircuts, and the discounting of recoveries received years later. |
| EAD Exposure at default |
Must incorporate expected drawdown of undrawn limits — distressed borrowers characteristically draw down available credit before defaulting. |
4 — Forward-looking information and governance
This is where management adds genuine value, and where the greatest scope for error resides. Macroeconomic overlays must link economic indicators to loss outcomes, with multiple weighted scenarios capturing non-linearity.
Two disciplines make this credible: documented judgment behind every overlay and post-model adjustment, and back-testing against actual outcomes.
The fastest route to an audit finding
Undocumented overlays. If the rationale behind a post-model adjustment exists only in someone's head, it cannot be defended when challenged — and it will be challenged.
5 — The three-stage model
Staging under IFRS 9 and the interest calculation basis at each stage
| Stage | Measurement | Interest calculated on |
|---|---|---|
| Stage 1 Performing |
12-month ECL | Gross carrying amount |
| Stage 2 Significant increase in credit risk |
Lifetime ECL | Gross carrying amount |
| Stage 3 Credit-impaired |
Lifetime ECL | Net carrying amount |
The most misunderstood concept in IFRS 9
12-month ECL is not the losses expected over the next twelve months. It is the portion of lifetime losses attributable to defaults occurring within twelve months. The loss amount is the full lifetime loss; only the default window is restricted.
The consequential judgment is the SICR trigger — the point at which an exposure moves from Stage 1 to Stage 2. It is the single decision that moves the number most, and it deserves considerably more attention than it typically receives.
The core insight
A simple model built on strong data, sensible segmentation, and documented governance will outperform a sophisticated model resting on weak foundations. Every time.
Complexity is seductive because it looks like rigour. But an ECL framework is only as credible as the data it consumes and the judgment it can defend.
Start with the foundation. The mathematics is the easier part.