IFRS 9: A modelling and judgement problem

It is the close of the reporting quarter, and a portfolio of performing mortgages has just added several million pounds to the loss allowance. No borrower has missed a payment. What changed is a forecast: the economics team revised its view of unemployment, the impairment models read that revision as higher risk, and a tranche of loans crossed from one provisioning stage into the next. A decade ago none of this would have touched the financial statements until the losses actually arrived. Under the International Financial Reporting Standard 9 (IFRS 9), a loss is recognised when it is expected, not when it is seen.

IFRS 9 is the accounting standard that governs how financial instruments are classified, measured, and impaired. It covers three areas:

• How assets are categorised on the balance sheet.
• How impairment, the allowance held against future losses, is calculated.
• How hedge accounting works.

The impairment model, built around expected credit loss (ECL), decides how much must be set aside against loans that may never default, and it is where the standard’s judgement and supervisory attention concentrate.

The standard has been mandatory for IFRS reporters since 1 January 2018, when it replaced International Accounting Standard 39 (IAS 39). It matters because the loan-loss allowance it produces is one of the largest judgement-based figures a lender reports, and the reasoning behind that figure is scrutinised as closely as its size.

IAS 39 vs IFRS 9: From incurred to expected

The change IFRS 9 introduced is best understood against the model it replaced. Under IAS 39, a credit loss could be recognised only once there was objective evidence that it had been incurred: a missed payment, a covenant breach, a restructuring. Losses everyone could see coming sat unprovisioned until that evidence crystallised. In the financial crisis this produced allowances that rose sharply only after the losses had already done their damage. The International Accounting Standards Board (IASB) accelerated its work in response, with the G20 among the parties pressing it on "the timeliness of recognition of expected credit losses"¹.

IFRS 9 reversed the logic. An allowance must now be held at all times, from the moment a loan is written, set at the losses the lender expects, not only those it can already prove. The shift looks subtle but it is not. It moves provisioning from a backward-looking record of what has happened to a forward-looking estimate of what might happen, turning the loss allowance into an attempt to predict losses before they arrive.

What are the three phases of IFRS 9?

IFRS 9 has three phases, but they differ in how much they leave to the lender. Two run largely on rules. One runs on judgement, and it is where almost all the modelling effort, and all the argument, goes.

The first phase, classification and measurement, is largely rule-driven. Every financial asset is grouped at the outset using two tests: the business model under which the asset is held, and whether its contractual cash flows are solely payments of principal and interest, known as the SPPI test. The outcome determines how the asset is measured:

IFRS 9 has three measurement categories, not two. Amortised cost holds an asset at what the borrower still owes, less the allowance. Fair value through other comprehensive income (FVOCI) and fair value through profit or loss (FVTPL) both mark it to what it would fetch if sold today; they differ in whether the movement runs through the balance sheet or through the income statement. Here are some practical examples:

• Retail mortgages are held at amortised cost because the bank’s sole objective is to collect principal and interest over their lifetime.
• Treasury debt (such as sovereign bonds) falls under FVOCI when held to earn interest while remaining available for frequent sale to manage daily liquidity.
• Trading book loans (like distressed corporate debt) are classified under FVTPL since they are actively managed for short-term trading gains rather than long-term cash flow collection.

The second phase is impairment, the ECL model, which applies to assets carried at amortised cost and FVOCI. Impairment is the phase that runs on judgement, and it is what the rest of this blog is about.

The third phase, hedge accounting, moved the other way: IFRS 9 dropped IAS 39’s quantitative bright lines in favour of an approach aligned to how a firm actually manages risk. For most lenders the hedges themselves stay operationally simple, whatever the accounting allows.

Staging: where 12 months becomes a lifetime

The heart of the impairment model is a three-stage classification that decides how far ahead a lender must look when sizing each loan’s allowance.

A loan whose credit risk has not risen meaningfully since it was written sits in Stage 1, and carries an allowance equal to 12-month expected credit loss. IFRS 9 is precise about what that means. A 12-month ECL is "a portion of the lifetime expected credit losses", representing "the lifetime cash shortfalls that will result if a default occurs in the 12 months after the reporting date", weighted by the probability of that default occurring² .

When credit risk increases significantly, the loan moves to Stage 2, and the allowance jumps to cover losses from defaults possible across the entire remaining life of the exposure. Stage 3 holds loans that are already credit-impaired; these too carry a lifetime allowance. A separate category, purchased or originated credit-impaired (POCI) assets, sits outside this flow with its own treatment.

The move from Stage 1 to Stage 2 can drastically increase a loan’s provision, yet the standard never says precisely when it should happen.

What triggers a significant increase in credit risk?

The transfer from Stage 1 to Stage 2 is the single most consequential judgement in the whole model. It is triggered by a significant increase in credit risk (SICR) since the loan was written. The standard deliberately leaves the trigger definition open but it does set a backstop: a loan more than 30 days past due (DPD) is presumed to have deteriorated significantly.

Most banks define their trigger by comparing the probability of default (PD) estimated today against the PD estimated when the loan was written, transferring the loan when the gap crosses an absolute or relative threshold.

Because the standard leaves staging and scenario design to each institution, Stage 2 ratios are not comparable across banks even within the same asset class.

Regulators expects firms to catch deterioration well before the 30-days-past-due backstop. The EBA in particular has guidelines that permit the presumption as a backstop "alongside other, earlier indicators for assessing significant increase in credit risk", while directing that institutions "should avoid using it as a primary indicator of transfer to lifetime ECL".³

Whether the PD-comparison method is even a good early-warning tool is contested. One study tested competing SICR definitions against real loan data and found that the choice materially changes which loans transfer, and when.⁴ The dominant practice is therefore a choice, not a rule.

How ECL drives the impairment number

Whichever stage a loan sits in, the ECL is built from three components estimated separately and then combined: the PD, the loss given default (LGD) and the exposure at default (EAD). LGD is the share of exposure not recovered after default; EAD is the balance expected to be outstanding when default occurs. Each of the three is a modelling problem in its own right, and our companion piece on ECL under IFRS 9 works through why a capital PD, LGD or EAD cannot simply be carried across into a provision.⁵

Two requirements make the ECL number distinctively demanding. The estimate must be probability-weighted, and it must be discounted to present value at the loan’s effective interest rate.

Probability-weighting is what forces forecasting into the accounts: a lender cannot simply run its central economic forecast and book the answer. The reason is more than procedural. A severe downturn drives losses up by more than an equally likely, equally sized upturn reduces them. So the loss from a single central forecast is not the same as the loss averaged across the range of outcomes.

Running and weighting three macroeconomic scenarios, typically a baseline, a downside, and an upside, is how a lender captures that asymmetry instead of missing it. The variables that move the number most are the familiar ones: GDP growth, unemployment, and, for mortgage books, house prices.

The choice of weights conceals a problem that we have noticed: two very similar banks in the same economy can report very different provisions because they weight their scenarios differently.

When is a post-model adjustment appropriate?

A model produces a figure, but the figure that reaches the balance sheet is rarely the raw model output. An adjustment sits on top, and it draws particular scrutiny.

The post-model adjustment (PMA), or overlay, is a manual addition that captures a risk the core model does not yet reflect, such as an emerging sector stress or a known data gap. Regulators treat overlays as legitimate but temporary. Coming back to the EBA, they mention that an overlay must target a specific identified risk, be quantified and documented, and carry a commitment to fix the model rather than become a permanent substitute for it.³ The EBA’s 2023 monitoring work points to "the supposed temporary nature of these adjustments and the limited governance of their application", and concludes that overlays should not be treated as equivalent to a proper collective assessment of significant increase in credit risk.⁶

Judgement is the work no equation shows

For a typical credit risk team, the lesson of IFRS 9 is that the model is one part. The other part is making the judgements the standard won’t make for you: when a loan has deteriorated enough to change stages, how severe the downside scenario should be and how likely, when an overlay is warranted and when it must be released. Each of these is a decision the lender makes, documents, and must be ready to explain.

An evidence trail is what most lenders will need. For every stage transfer, every scenario weight, and every overlay on the books this quarter, a lender should be able to show an auditor or supervisor why it moved and what information drove it. That is the work IFRS 9 also requires, and it is the work that appears in no equation.

Frequently asked questions