ECL under IFRS 9: an unbiased view of future credit losses

A model risk committee is reviewing a Stage 2 provision that has doubled since last quarter. Stage 2 means the allowance now has to cover losses across each loan’s whole remaining life, not just the next twelve months. No borrowers have defaulted. Every account is performing. Where does the number come from and is it a fair reflection of future expected losses? Trace it back and the answer runs through four separate estimates, each resting on its own assumptions:

• Probability of default (PD): how likely borrowers are to default.
• Loss given default (LGD): how much is lost when they do.
• Exposure at default (EAD): how much will be owed at that point.
• Scenario weights: how probable a recession is this year, and how much weight the provision gives that possibility.

This four-part structure is what the industry has built in practice: PD, LGD, EAD and scenario weights. The International Financial Reporting Standard 9 (IFRS 9) is clear about what it wants. Its measurement requirement asks for an allowance that is unbiased, weighted across a range of possible outcomes rather than pinned to one. It should also be discounted for the time value of money, and built from whatever the lender can reasonably find out about past events, current conditions and the economic outlook.¹

IAS 39 looked back, IFRS 9 looks forward

Under International Accounting Standard 39 (IAS 39), impairment required an objective loss event before any allowance could be booked, as our article on IFRS 9 explains². Something had to have gone wrong first, and the lender recognised a loss only once it had.

IFRS 9 changed the question. The standard does not ask what has happened; it asks what is expected to happen, weighted across all outcomes, over the instrument’s full remaining life. Staging is the mechanic that carries the change, and we broke it down in that earlier piece: Stage 1 instruments attract a 12-month expected credit loss (ECL), while Stage 2 (where a significant increase in credit risk, or SICR, has been identified) and Stage 3 (where the asset is credit-impaired) both attract a lifetime ECL.

Why a borrower’s PD is not constant over the life of a loan

Stage 1 loans require a 12-month PD. The twelve months run from the reporting date, not from origination. The question is how likely this borrower is to default at some point in the coming year, asked again at every reporting end date. Stage 2 requires a curve instead: the probability that the borrower defaults in year 1, year 2, year 3, and so on through to maturity, with the 12-month figure sitting inside it as the first year.³

Take a borrower in the first year of a mortgage and the same borrower ten years in. They are not the same risk. Early on, little has been repaid, there is no track record of payments under pressure, and default is comparatively likely; a decade of instalments later, it is not. The curve has to reflect that shape, and there is no shortage of statistical tools for fitting it, from period-by-period regressions to models that follow loans through arrears, default and cure and back again.⁴

Why capital LGD and EAD models need recalibrating for IFRS 9

LGD

Most mature lenders came to IFRS 9 with internal ratings-based (IRB) LGD models already built for regulatory capital and hence started there. The European Banking Authority (EBA) has repeatedly found in its monitoring work that they do not transfer cleanly. Its 2023 exercise saw a material share of institutions applying flat or minimally adjusted LGD curves across the lifetime horizon, consistent with transplanting IRB parameters into the ECL framework without recalibrating them.⁵ A model built to set capital is not usable as it stands for setting a provision, because the two are calibrated to answer different questions.

The difference is in the calibration.

• IRB LGD (capital): set through-the-cycle, one figure meant to hold on average across good years and bad, then floored so that it stays conservative when conditions are benign.
• IFRS 9 LGD (provision): a point-in-time estimate of what would actually be lost on a default happening now, under today’s conditions and today’s forecast,⁵ with the recovery cash flows discounted back at the rate the loan was originally priced at.¹

A through-the-cycle LGD is a long-run average that barely moves. A point-in-time LGD moves with the cycle, which is exactly what makes it right for a provision and wrong to lift from capital.

Why exposure at default is harder to estimate for a revolving facility

EAD is straightforward for an amortising term loan: the outstanding balance at each horizon, less the repayments scheduled before it. A £200,000 mortgage might stand at £193,000 after a year and £185,000 after two, so those are the exposures the model carries into year one and year two. Nothing the borrower chooses to do changes them.

Revolving facilities are different from amortising term loans, because there the borrower sets the balance. Take a credit card with a £10,000 limit and £2,000 drawn, meaning £2,000 actually borrowed against that limit and £8,000 still available. By the time it defaults it could be at £2,000 or at £9,500.

On Gini’s reading of revolving portfolios, the share of the limit already drawn is among the strongest predictors of how much more will be taken, and in a downturn borrowers draw harder still as other funding dries up.

Forcing the forecast in

IFRS 9 requires the allowance to take account of what is expected to happen to the economy, so far as that can be worked out from information the lender can reasonably get hold of. Past loss experience is the starting point but cannot be the finishing point, because it has to be adjusted for what is observably true now: a book calibrated on the last decade cannot be left to speak for a year in which unemployment is climbing. The EBA puts the same point as an expectation, that information on historical loss experience "may not fully reflect the credit risk in lending exposures".⁶ The standard does not prescribe how to do this. It requires only that the resulting allowance span a range of possible outcomes rather than a single forecast, and that it be unbiased.

In practice this has settled into three scenarios: baseline, downside and upside. Most institutions in the EBA’s 2023 sample were doing exactly that, but the weights they landed on varied substantially: some assigned almost the entire weight to the baseline, in effect running a single-scenario model, while others spread it more evenly across all three.⁵

Average scenario weights across the EBA’s 2023 monitoring sample. The average hides wide dispersion: individual institutions range from putting almost the whole weight on the baseline to spreading it evenly.

In practice the scenario weighting barely moves the resulting allowance. When the EBA compared the weighted answer against the baseline on its own, the two came out close enough that the weighting was doing little work. The reason it gives is not that the downside is too mild. It is that the baseline assumptions dominate the answer from the start. On top of that, the models translating economic forecasts into losses often barely react to the variables that make a downside a downside. Where that happens, the downside contributes little whatever weight it is given.⁵

This is our own reading rather than a finding of the report. One further mechanism is visible in the shape of the forecasts themselves. Macroeconomic paths get smoothed, and a downside that returns to the long-run average within a year or two ends up looking much like the central case, which leaves the weighting exercise little left to do.

What an unbiased ECL means

Unbiased is a property of the whole calculation, not of any one part of it. Each of the four estimates can pass its own test and the ECL can still come out high or low.

The model risk committee’s better question, then, is not which model moved. A Stage 2 provision doubles for one of two reasons: balances have crossed the SICR line and switched from a 12-month measurement to a lifetime one, or the lifetime numbers themselves have risen. Those are different problems, and only one of them is about the models.

Frequently asked questions