Forward-Looking IFRS 9 Is a Credit-Risk Transmission System, Not Three Forecasts Averaged Together

Forward-looking information is a transmission architecture, not an add-on

A historical credit model may explain yesterday's borrower risk well and still mismeasure tomorrow's loss. The simplistic response—create baseline, upside and downside paths, then average them—does not answer where unemployment, rates, growth, inflation, property prices or sector conditions enter the impairment system.

Forward-looking information can affect PD, LGD, EAD, SICR and overlays. The design problem is therefore omission and double counting at the same time: where does each economic risk transmission belong?

One average economic scenario can produce the wrong expected loss

Suppose favourable conditions reduce PD modestly but severe stress increases it sharply. The credit response is convex in the adverse region: evaluating a model at the average macro input can understate the loss obtained by evaluating each plausible path and weighting the outcomes.

The probability-weighted PD is 4.8%. A hypothetical model evaluated once at the average macro input might produce 4.5%. Both numbers are illustrative; the 0.3 percentage-point gap demonstrates why probability weighting should usually operate on loss outcomes when material nonlinearities exist.

A scenario is a coherent joint economic narrative

A baseline is a credible central path; an upside is a credible favourable state; a downside is a credible adverse state. Three scenarios are neither universally mandatory nor automatically sufficient. The architecture must be rich enough to capture material non-linearity, but no richer than the evidence can support.

A downside is not baseline made slightly worse variable by variable. GDP, unemployment, rates and inflation must move together plausibly under a shared narrative. Independently shocking every series can create combinations that have no economic interpretation.

Variable selection is the product of these six tests—not a contest for the lowest p-value. GDP, unemployment, income growth and confidence may carry overlapping information. Parsimony, sign, lag, stability and economic interpretation matter because correlated predictors can double-count a cycle or make the response unstable.

Likelihood and severity are different axes

An IFRS 9 scenario supports a probability-weighted expected outcome. A stress test assesses resilience under adversity and is often not weighted as an expected outcome. Severity alone does not convert one into the other.

Forecast quality and credit-risk transmission are separate questions

Unemployment can weaken income capacity; rates can increase debt-service burden; GDP and sector contraction can impair household or business cash flow. The same shock can depress collateral prices, lengthen recoveries and reduce cure, raising LGD. It can also increase utilisation, drawdown and balance persistence, raising EAD.

Under stress, PD↑, LGD↑ and EAD↑ can occur together. Independent parameter averages suppress their covariance. A reputable external GDP forecast is useful input, but it is not a PD model; forecast consensus and dispersion still require a governed transmission design.

Sparse macro history creates structural uncertainty

Millions of monthly account rows do not create millions of economic cycles. Short histories, common trends, policy regimes, structural breaks and product redesign can create spurious relationships. An institution may have one recession and a short digital-lending history. Complex regression does not remove that uncertainty.

Lag choices should reflect mechanism: PD_t may depend on Macro_t, Macro_t−1 and Macro_t−2, but adding lags mechanically creates over-parameterisation. Sign, magnitude and lag must survive out-of-time and regime challenge.

Calculate loss inside each scenario before weighting

Consider a fictional one-period €10 million revolving portfolio. Values are original, simplified and illustrative; EAD differs because utilisation responds to conditions.

If PD alone were weighted (3.13%) and then multiplied by weighted-average LGD (40%) and EAD (€9.54m), the result would be about €119,441—€4,609 below the correct scenario-level calculation. The shortcut loses the adverse co-movement of PD, LGD and EAD.

For lifetime PD, each scenario generates hazard h_t,s, survival S_t,s, marginal PD and cumulative PD. The identities described in Lifetime PD Term Structures must remain valid after scenario conditioning.

Scenario probability is not scenario contribution

Holding paths constant but moving weights from 20 / 60 / 20 to 10 / 50 / 40 increases weighted ECL from €124,050 to €176,853: a €52,803 weight effect. Holding the original weights but increasing downside ECL from €325,000 to €400,000 increases weighted ECL by €15,000: a path or severity effect.

This separation tells Finance whether allowance rose because the economic paths worsened, management assigned more likelihood to downside, or both. Monitor Δw_s and its ΔECL rather than describing all macro movement as “the outlook.”

Forecast confidence should decay as horizon extends

Short-term forecasts can support detailed paths; medium-term uncertainty widens; beyond a reasonable and supportable horizon, methodology may transition toward longer-run relationships. A model should not pretend equal confidence at years one and ten.

Speed, shape and validation are portfolio-specific. Abrupt jumps from forecast PD to a long-run PD create artificial term-structure discontinuities; controlled mean reversion should preserve probability identities and an economically coherent transition.

Exposure horizon is decisive. A six-month product gains little from a ten-year macro path. Forecast horizon should align with expected exposure life, while longer-duration portfolios require explicit tail and convergence treatment.

Macroeconomic deterioration can change both stage and measurement

Forward-looking deterioration can increase current lifetime PD and affect SICR and Stage 2 even when borrower data are unchanged. That can be economically coherent, but governance must explain whether migration arises from borrower-specific deterioration, broad macro deterioration or both—and track a conceptual Stage2_{macro effect}.

Every material driver needs one coherent primary transmission path, or a documented reason for appearing across components. Stress that moves accounts to Stage 2, raises the PD curve, raises LGD and then receives a duplicate overlay is not conservatism; it is an uncontrolled measurement stack.

An overlay needs an entry condition and an exit condition

A management overlay may be justified by an unprecedented event, delayed data, a structural break or a known model blind spot. It requires rationale, quantification, approval, sensitivity, double-counting assessment, review frequency and release logic.

An overlay should not remain because it existed last quarter. The boundary question is: is this risk temporary and outside model design, or persistent enough to require redevelopment? Repeated overlays often reveal structural model deficiency.

Versioned attribution makes the allowance explainable and reproducible

Compare M_t,s^Q1 with M_t,s^Q2, trace forecast revisions into parameters, and separate them from portfolio composition and strategy. If defaults rise while the lender enters a riskier acquisition channel or loosens its cut-off strategy, observed deterioration is Macro Effect + Mix Effect—not proof of a larger macro coefficient.

Each reporting cycle should retain scenario ID, forecast date, complete paths, weights, source, approvals and effective period, alongside model, parameter and overlay versions. For reporting date T, the institution should be able to reconstruct ECL_T exactly.

Vintage analysis protects another distinction: portfolios under the same current economy can differ because origination standards, pricing, channels and customer mix differed. Macro attribution should not erase those cohort effects.

Validate the economic claim, not only the regression fit

Backtesting should compare historical forecast-time predictions with outcomes using only information available at that reporting date. Realised macro data inserted retrospectively create look-ahead bias. Out-of-time performance, coefficient stability, sign, lag, horizon and sensitivity matter more than a polished in-sample fit.

Challenge should ask whether downside is sufficiently adverse, upside plausible, correlations coherent, weights defensible, non-linearity captured, extrapolation visible and double counting absent. Scenario weights remain governed judgement—not objective truth merely because they sum to 100%.

Short-tenor non-bank portfolios need their own economic clock

In high-risk consumer lending, macro transmission can emerge quickly, data matures quickly, customer mix changes rapidly, acquisition channels can dominate and seasonality can be strong. An unsecured six-month product should not imitate a long-duration mortgage architecture.

Segment sensitivity should follow portfolio economics: mortgages may respond to rates and property prices; unsecured consumers to unemployment and income stress; SMEs to GDP and sector conditions. Product, channel, vintage and strategy effects must be separated before a movement is labelled macroeconomic.

Twenty failure modes that weaken forward-looking ECL

A Macro Scenario & ECL Sensitivity Agent can maintain evidence—not approve judgement

A bounded recurring agent could ingest approved forecasts, maintain scenario versions, validate path coherence, apply governed PD/LGD/EAD models, calculate scenario and weighted ECL, test alternative weights, measure contributions, flag nonlinearities and possible double counting, compare quarters, and prepare attribution evidence.

Its role is scenario analytics + ECL sensitivity + attribution + governance support. It must not autonomously approve economic scenarios, accounting judgements, weights or overlays. The service architecture connects naturally to Credit Risk and CFO & Finance, because the same assumptions affect risk measurement and reported impairment.

Continue through the IFRS 9 research chain: Expected Credit Loss, Lifetime PD, LGD, EAD & CCF, and SICR. Related diagnostic foundations include Model Calibration Drift, Credit Risk Model Validation and Credit Vintage Analysis.