IFRS 9 PD, LGD and EAD: Calibration and Validation in Practice

IFRS 9 replaced the incurred-loss impairment model of IAS 39 with an expected-credit-loss (ECL) model that requires forward-looking, probability-weighted estimates for every financial asset held at amortised cost or at fair value through OCI. In its simplest form the ECL for a facility is:

ECL = Σ_t PD(t) · LGD(t) · EAD(t) · DF(t) for the 12-month horizon (Stage 1) or the full lifetime (Stages 2 and 3), where DF(t) is the discount factor at the effective interest rate.

That formula hides a decade of implementation debate. This article walks through what "PD", "LGD" and "EAD" actually mean under IFRS 9, how each component is calibrated, and what an independent validator looks for.

PD under IFRS 9: from regulatory anchors to PIT curves

Banks under the IRB approach already have regulatory PD models calibrated through the cycle (TTC) — a long-run average default rate by grade. IFRS 9 requires a point-in-time (PIT) view: the probability that this obligor defaults given today's macro state and the forward path.

Two families of approaches dominate:

1. Scalar/Vasicek transformation. A TTC PD is transformed to a PIT PD via the single-factor Merton–Vasicek framework: PD_PIT = Φ((Φ⁻¹(PD_TTC) − √ρ · Z) / √(1−ρ)), where Z is a standardised systemic factor derived from macro variables and ρ is the asset correlation. The advantage is auditability — every step is closed form. The disadvantage is that PIT-ness depends heavily on how Z is constructed.
2. Direct macro regression. A logistic model links observed default rates to macro variables (GDP growth, unemployment, house-price index) with lags, and produces PIT PDs directly. This is more flexible but harder to reconcile with the regulatory TTC anchor.

Under either approach, the PD term structure needs to extend to the maturity of the facility. Common techniques are Markov migration matrices calibrated PIT, or survival-function extrapolation. Validators should verify that the term structure integrates to reasonable cumulative default probabilities and does not oscillate or explode at long horizons.

Multiple scenarios and probability weighting

IFRS 9 requires probability-weighted outcomes. In practice most institutions use three to five macro scenarios — typically baseline, upside and one or two downsides — combined with subjective probability weights. The scenarios must be internally consistent (an unemployment path that contradicts the GDP path fails validation immediately) and the weights must be documented. Validators check that the resulting weighted PD is not simply the baseline in disguise, and that the downside scenarios genuinely reflect adverse tails — a lesson reinforced by macro stress testing for ICAAP.

LGD: downturn, cure and the collateral question

LGD is the loss given default, expressed as a fraction of EAD. Empirical LGD is bimodal for many portfolios: full recovery (near zero LGD) for cured cases, and a heavier distribution around 40–60% for genuine liquidation cases. Averaging over this bimodality produces LGD estimates that describe neither pole well.

A robust LGD framework separates:

• Cure LGD: for exposures that default and return to performing. Usually low but positive after direct costs and time value.
• Recovery LGD: for exposures resolved through workout or collateral realisation, with recovery times of 12–48 months for corporate secured exposures.
• Discounting: cash flows discounted at the effective interest rate, not at a workout-specific rate — an IFRS-9-specific choice that differs from IRB practice.

Downturn LGD under IFRS 9 is forward-looking: it applies whenever the macro scenario used to compute the ECL implies downturn conditions. A common failure mode is to apply a static downturn add-on across all scenarios, which double-counts the downturn in the adverse scenarios and understates the baseline separation.

Cure and redefault dynamics matter for stage-transfer accounting. A facility that cures out of Stage 3 but re-defaults within 12 months should be flagged in the cure LGD calibration, or the model will systematically overstate cure recoveries.

EAD: commitments, prepayments and behavioural balances

EAD is straightforward for a fully drawn term loan and challenging for almost everything else. For revolving facilities and undrawn commitments, EAD depends on the credit conversion factor (CCF) — the fraction of undrawn commitment expected to be drawn at default. Regulatory CCFs are TTC and often conservative; IFRS 9 CCFs need to be PIT and best-estimate.

For amortising products with prepayment optionality, EAD needs a behavioural balance model. Prepayment tends to accelerate in falling-rate environments and stall in rising-rate ones — the same dynamics that drive IRRBB behavioural modelling. Ignoring the interaction between prepayment and default (they compete for the borrower's cash flow) leads to double-counted expected balances.

For credit cards, EAD is essentially a behavioural forecast: what utilisation does the account reach in the months before default? Regression models on account age, utilisation trajectory and payment behaviour are standard.

Staging: the discipline that ECL depends on

The stage-transfer criteria — a "significant increase in credit risk" (SICR) — are as material as any of the components above. The paragraph 5.5.11 backstop of a 30-days-past-due trigger is a floor, not a ceiling; institutions supplement it with:

• Absolute thresholds (e.g. lifetime PD > x%).
• Relative thresholds (e.g. lifetime PD at reporting date more than y× the lifetime PD at origination for the residual maturity).
• Qualitative triggers (forbearance flags, watchlist status).

Validators should check that staging is symmetric — an asset that improves back below the SICR threshold, with a defined probation, moves back to Stage 1 — and that the SICR calibration produces a Stage 2 population that is plausibly ahead of default but distinguishable from Stage 1.

Backtesting and outcomes analysis

IFRS 9 does not prescribe a specific backtesting framework, but ITS and EBA guidance on ICAAP and IRB carry over in spirit. Practical tests:

• PD calibration: realised default rates versus predicted, by grade and stage, with binomial confidence intervals.
• PD discrimination: Gini/AUC on defaults observed in the reporting window.
• LGD calibration: realised loss on resolved cases versus predicted, split by cure and recovery, with the workout period accounted for.
• EAD calibration: realised balance at default versus predicted, by product and time-to-default bucket.
• Stage migration: transition matrices should be stable in benign periods and respond in stressed ones without whipsawing.

For all three components, backtest results should feed a documented feedback loop: findings inform calibration cycles, override registers and, where needed, independent validation lifecycle actions.

Management overlays: the honest end of the ECL number

IFRS 9 permits — and in practice requires — post-model adjustments for risks the models cannot capture yet: novel sector stress, geopolitical events, model-recalibration lags. Overlays are legitimate; opaque overlays are not. Best practice includes an overlay register with the driver, the quantitative basis, the exposure covered, the sign-off body and the sunset condition.

Governance and disclosure

Boards should see the ECL sensitivity to macro scenarios and to key model assumptions (SICR thresholds, downturn LGD add-ons, cure rates). IFRS 7 disclosure requirements — quantitative and qualitative — provide the disclosure architecture, and ESMA's public statements on IFRS 9 have consistently pushed for granularity on scenario weights and overlays.

Limitations

Every ECL number is a point estimate of a distribution; a probability-weighted mean of three or five scenarios understates the true tail. Institutions that use ECL as their only credit-loss estimate — without a parallel stress or economic-capital view — are working with a compressed picture of credit risk.

Conclusion

IFRS 9 rewards rigour in every component. The banks that produce credible ECL numbers are the ones that treat PD, LGD, EAD, staging and overlays as five separate calibration problems with five separate validation plans — not one blended exercise.

Related reading

• The Model Validation Lifecycle
• Credit Risk Stress Testing for ICAAP
• IRRBB and CSRBB Behavioural Models

References and further reading

• IFRS Foundation, IFRS 9 Financial Instruments.
• European Banking Authority, Guidelines on credit institutions' credit risk management practices and accounting for expected credit losses (EBA/GL/2017/06).
• Basel Committee on Banking Supervision, Guidance on credit risk and accounting for expected credit losses, BCBS 350.
• European Securities and Markets Authority, public statements on IFRS 9 implementation.

About the author

This piece continues an ongoing series on credit and capital modelling — more about the author.