Quick answer: expected-credit-loss (ECL) accounting asks banks to recognise credit losses before they are fully realised. But two major frameworks do this differently. IFRS 9 uses a staging architecture: broadly, performing assets begin with 12-month expected credit losses, assets whose credit risk has increased significantly move to lifetime ECL, and credit-impaired assets remain on lifetime ECL with additional accounting consequences. US CECL does not use the same three-stage trigger; for covered amortised-cost assets it generally estimates lifetime expected credit losses from the start. Both frameworks force the bank to combine historical evidence, current conditions, forward-looking information and model judgment.

The difficult question is no longer only “Who has already failed?” It is “What loss distribution is already embedded in the loans we still expect mostly to repay?”

Why this belongs in mathematics

Expected-credit-loss estimation combines conditional default probabilities, survival, exposure at default, recovery, discounted cash shortfalls, scenario probabilities, segmentation, forecasting and model uncertainty. It also exposes a philosophical boundary in quantitative work: the number reported today depends partly on a model of futures that have not happened.

This is different from bank regulatory capital. Accounting provisions and regulatory capital interact, but they have different rules and purposes. A mathematically similar PD × LGD × EAD engine can feed more than one system without making those systems identical.

1. Start with a simple expected-loss identity

A simplified one-period credit-loss model is:

Expected Loss = PD × LGD × EAD

PD — probability of default over the relevant horizon.
LGD — proportion of exposure lost if default occurs after recoveries and collateral.
EAD — exposure expected to be outstanding when default occurs.

If PD = 2%, LGD = 40% and EAD = S$1 million, the simplified expected loss is S$8,000. That does not mean the loan will lose exactly S$8,000. The borrower may pay in full, or the bank may lose hundreds of thousands. ECL is a probability-weighted quantity across possible outcomes.

For a multi-period portfolio, the model becomes a time series: marginal default probabilities, future exposures, recoveries and discount factors must be considered across the life of the instrument.

2. IFRS 9: the three-stage architecture

IFRS 9 stage	Broad credit state	Loss allowance
Stage 1	Credit risk has not increased significantly since initial recognition	12-month expected credit losses
Stage 2	Significant increase in credit risk (SICR), but not credit-impaired	Lifetime expected credit losses
Stage 3	Credit-impaired	Lifetime expected credit losses, with credit-impaired interest-recognition treatment

The BIS summary of IFRS 9 describes this as a three-stage expected-loss framework introduced to replace the earlier incurred-loss approach. See IFRS 9 and expected loss provisioning.

3. “12-month ECL” does not mean losses expected only in the next 12 months

This is one of the easiest ideas to misunderstand. IFRS 9’s 12-month ECL is not simply the cash loss expected to occur within twelve months. It is the portion of lifetime cash shortfalls associated with defaults that are possible within the next twelve months.

Suppose a five-year loan could default in month 10 and then generate recovery shortfalls extending beyond month 12. Those later shortfalls can still be part of the 12-month ECL because the triggering default occurred inside the next twelve months.

This distinction matters because horizon definitions are mathematical operators, not ordinary-language labels.

4. Stage 2 is a change-detection problem

The move from Stage 1 to Stage 2 is driven by a significant increase in credit risk since initial recognition. That makes staging a relative-state problem, not simply a current-rating problem.

A borrower may still be paying on time but have become materially riskier because of deteriorating financial performance, adverse sector conditions, a downgrade, weakening collateral or forward-looking macroeconomic information.

The algorithm therefore needs two states:

credit risk at origination, and
credit risk at the reporting date.

This is why a fixed threshold such as “PD above 5% = Stage 2” can be too crude. Moving from 0.2% to 1.5% may represent a much larger relative deterioration than moving from 7% to 8%, depending on product and policy.

5. CECL: lifetime loss from the beginning

The US Current Expected Credit Losses methodology takes a different route. For financial assets measured at amortised cost within its scope, the allowance reflects expected credit losses over the contractual term, adjusted for expected prepayments where appropriate.

CECL therefore does not wait for an IFRS-9-style Stage 2 transition before moving a performing loan to a lifetime expected-loss horizon. The Federal Reserve describes CECL as replacing the incurred-loss methodology with a forward-looking lifetime-loss objective that uses past events, current conditions and reasonable and supportable forecasts.

See the Federal Reserve’s CECL Frequently Asked Questions and the FDIC CECL resource centre.

6. IFRS 9 and CECL should not be collapsed into one algorithm

Feature	IFRS 9 ECL	US CECL
Performing asset at initial recognition	Generally 12-month ECL	Generally lifetime expected credit loss
Staging	Three-stage impairment architecture	No equivalent IFRS 9 three-stage model for amortised-cost assets
Forward-looking information	Required; probability-weighted outcomes and non-linearity matter	Past events, current conditions and reasonable/supportable forecasts required
Beyond forecastable horizon	Framework-specific lifetime measurement using reasonable/supportable information	Reversion to historical loss information/proxy beyond reasonable-supportable forecast period
Method	Principle-based; multiple modelling approaches possible	No single mandatory estimation technique; reasonable supported methods permitted

The shared concept is forward-looking expected loss. The recognition mechanics are different. An educational article that merges them into “banks use lifetime ECL” would erase one of the most important distinctions.

7. Probability-weighted scenarios: the mean is not the most likely case

Suppose a bank considers three macroeconomic scenarios for a portfolio:

Scenario	Probability	Modelled lifetime credit loss
Upside	20%	S$20m
Baseline	55%	S$35m
Downside	25%	S$90m

A simple probability-weighted ECL is:

0.20×20 + 0.55×35 + 0.25×90 = S$45.75 million.

The baseline is the single most likely scenario, yet the probability-weighted result is much higher than the S$35 million baseline loss because the downside scenario is nonlinear and severe.

IFRS material on forward-looking information emphasises that “expected” is a probability-weighted concept rather than simply management’s most likely forecast. See IFRS 9: Forward-looking information and multiple scenarios.

8. Macroeconomic variables enter through credit behaviour

Unemployment, GDP growth, interest rates, property prices, commodity prices and sector-specific indicators can affect PD, LGD, EAD or cash-flow expectations. Different portfolios require different drivers.

A mortgage book may be sensitive to unemployment, house prices and borrowing costs. Commercial real-estate loans may depend on vacancy, rent, refinancing rates and collateral values. Credit cards may respond differently to labour-market and household-income shocks.

This creates a model-selection problem: adding every macro variable can overfit history, while using too few variables can miss the transmission channel that matters in the next recession.

9. Lifetime ECL needs a time path, not one annual PD multiplied by years

If annual PD is 2%, five-year cumulative default probability is not automatically 10%. Survival changes the population at risk each year, and PD itself may vary with age and scenario.

With a constant 2% conditional annual PD and no competing exits, five-year survival is approximately 0.98⁵ ≈ 90.4%, so cumulative default probability is about 9.6%. If PD rises through a recession, the path changes again.

A lifetime ECL engine therefore often works with marginal PDs or survival curves, future EAD and period-specific LGD rather than simply multiplying a one-year loss rate by remaining maturity.

10. Prepayment and credit-line drawdown change EAD

Expected exposure is not static. A mortgage can prepay before default, reducing future EAD. A revolving credit line can be drawn further as a borrower becomes stressed, increasing EAD. The ECL model therefore needs behavioural assumptions as well as credit assumptions.

That connects ECL to mortgage prepayment models and to liquidity models for contingent credit lines.

11. Model overlays: when the model is known to be incomplete

Sometimes management knows the statistical model is not capturing a material emerging risk. Historical data may contain no pandemic, no sudden war, no new regulatory shock, no rapid change in underwriting, or no comparable digital run on a borrower sector. Institutions may use documented qualitative adjustments or post-model overlays where permitted by the applicable framework and governance process.

An overlay should not mean “management dislikes the model answer.” It needs a clearly identified missing risk, evidence, a defined calculation or judgment process, governance, expiry criteria and a plan for eventual model incorporation where appropriate.

US supervisory policy explicitly recognises qualitative adjustments when historical loss information does not fully capture current conditions and reasonable/supportable forecasts. See the Interagency Policy Statement on Allowances for Credit Losses.

12. Creative-work lens: The Big Short and loss recognition before default

The Big Short is not an accounting manual, but it provides a useful attention lens: deterioration can become visible in underwriting quality, borrower structure and market conditions before contractual default is finally recorded. ECL accounting formalises a related discipline—though with very different rules and controls—by requiring expected deterioration to enter measurement before every borrower has actually stopped paying.

The creative work asks, “What evidence are people ignoring?” The accounting model must answer with documented data, scenarios and recognised standards rather than narrative conviction.

13. The algorithmic pipeline

Define the accounting framework and portfolio scope. Do not mix IFRS 9 and CECL recognition logic.
Segment assets by shared risk characteristics.
Build historical default and loss data.
Estimate PD, LGD, EAD or alternative loss-rate/cash-flow models.
For IFRS 9, assess significant increase in credit risk and stage classification.
Determine the required loss horizon. 12-month or lifetime under IFRS 9; lifetime objective for covered CECL assets.
Develop forward-looking macroeconomic scenarios or forecasts.
Translate scenarios into credit-risk parameters or cash shortfalls.
Probability-weight outcomes where the framework and nonlinearity require it.
Discount expected cash shortfalls according to applicable requirements.
Apply justified qualitative adjustments/overlays.
Aggregate allowance results and reconcile to the general ledger.
Backtest defaults, recoveries and scenario sensitivity.
Update staging, segmentation and forecasts each reporting period.

14. Failure modes

Framework collapse. IFRS 9 staging and CECL lifetime recognition are treated as the same rule.
Most-likely-scenario bias. The baseline forecast is substituted for a probability-weighted expected value.
Static PD. One annual default rate is multiplied mechanically across the remaining life.
SICR cliff design. Stage 2 relies on one crude threshold that misses meaningful relative deterioration.
Historical regime lock. Models assume the future loss relationship must resemble the calibration period.
Overlay permanence. Temporary judgment adjustments become unchallenged recurring reserves.
Double counting. A risk already captured in the model is added again through an overlay.
Data leakage. Information unavailable at the historical reporting date is accidentally used in backtesting, making the model look better than it was.

15. Diagnostics and falsifiers

How much of the allowance comes from Stage 2 migration rather than parameter changes?
What portfolio would change most if the downside scenario probability doubled?
Are realised defaults persistently outside predicted PD bands?
Does LGD worsen when collateral prices and recovery timelines are stressed together?
Which overlay has remained in place longest, and why has it not been modelled?
Would the same loan receive a different horizon treatment under IFRS 9 and CECL?
What happens if the reasonable-and-supportable forecast period is shortened?
Does an independent challenger model produce materially different ECL?

Suppose someone claims, “The baseline scenario is most likely, so we can use its loss estimate as ECL.” A falsifier is a nonlinear downside scenario whose probability-weighted contribution materially raises the expected value. The most likely outcome and the mathematical expectation are not the same object.

16. Verification and update triggers

backtest PD, LGD and EAD by segment and vintage;
compare predicted stage migration with realised deterioration;
reconcile allowance movements to drivers rather than only total change;
challenge macroeconomic scenario weights independently;
document why qualitative adjustments exist and when they should expire;
update segmentation when assets no longer share risk characteristics;
re-estimate models after underwriting or economic regimes change;
keep accounting-model validation separate from regulatory-capital validation even when shared data are used.

Connections across the finance-and-banking algorithms lane

Credit-scoring algorithms — borrower-level probability estimation.
Credit-rating migration models — state transitions that can drive lifetime credit deterioration.
Bank capital models — related but distinct prudential use of credit-risk parameters.
Probability Is Not a Promise — the core statistical discipline behind expected loss.

Research anchors

The deeper lesson

Expected-credit-loss accounting turns uncertainty into a reported number before the uncertainty resolves. That makes model discipline unusually important. The bank must distinguish horizon from event timing, expected value from most likely outcome, historical evidence from forecast, and model output from management judgment. The aim is not to predict which exact borrower will fail. It is to make the portfolio’s already-present credit uncertainty visible without pretending the future is known.

Educational note: This article explains public accounting and banking mathematics. It is not accounting advice, audit advice, investment advice or guidance for preparing any institution’s financial statements.

Small Group Tutorials

How Banks Estimate Expected Credit Loss: IFRS 9 Staging, CECL Lifetime Losses, Probability-Weighted Scenarios and Model Overlays