LGD & EAD Validation: Common Pitfalls

Ask most credit-risk teams which model matters most and they will say PD. Ask an experienced validator, and they will say it is a coin toss between LGD and EAD — because these are the models nobody looks at closely enough, and small biases in either one move ECL and RWA more than a well-calibrated PD ever will.

LGD is not one number

Loss Given Default is often reported as a single percentage per grade or product. That obscures three distinct components:

1. Cure rate. What share of defaults roll back to performing without a workout loss? Cure rates depend heavily on the cure definition (90 days? 12 months?), and small changes here move the LGD headline substantially.
2. Recovery rate on non-cured defaults. Discounted cash flows recovered, net of workout costs.
3. Downturn add-on. A margin over long-run LGD to reflect stressed recovery environments — required under IRB, and creeping into IFRS 9 stress overlays as well.

Validation typically finds problems in one of two places: the cure rate is applied inconsistently between historical calibration and current application, or the downturn add-on is a flat percentage bolted on rather than a data-supported adjustment. Either can move LGD by 5–15 percentage points, and 10 points of LGD on a secured book is a very large ECL number.

EAD is where hidden risk lives

For term products EAD is close to outstanding balance, and the model is largely mechanical. For revolving credit — cards, overdrafts, undrawn corporate facilities — EAD depends on the credit conversion factor (CCF): the share of the undrawn commitment expected to be drawn between today and default.

Where EAD models go wrong:

• CCFs estimated in benign times. Historical CCFs from 2015–2019 look nothing like actual behaviour in a genuine stress period. Facilities that were 30% drawn on average moved to 70% or higher during 2020 for some segments.
• Cohort mixing. Combining low-utilisation prime customers with high-utilisation subprime in the same CCF grade masks the real dispersion.
• Time-to-default assumptions. A 12-month CCF is not a 3-month CCF. Some frameworks silently use one when reporting the other.

Backtesting: what to actually test

Meaningful LGD and EAD backtesting compares model estimates against realised losses and drawn balances from a closed cohort — defaults that occurred long enough ago that workouts are effectively complete.

Three tests that catch the majority of real problems:

• Central tendency test. Weighted mean of realised LGD versus predicted LGD by grade, product, and vintage. A persistent one-sided bias is the biggest single red flag.
• Discriminatory power. Somers' D or a similar rank measure to check the model separates high-loss from low-loss defaults, not just gets the average right.
• Realised CCF distribution. For revolving products, plot the histogram of realised CCFs against the modelled point estimate. Long right tails are usually where the model breaks.

Benchmarking: don't skip it

External LGD and EAD benchmarks are patchier than PD benchmarks, but they exist — supervisory studies (EBA benchmarking exercises, Fed CCAR public disclosures), rating-agency historical recovery studies (Moody's Ultimate Recovery Database, S&P LossStats), and industry pools. A model that is 15 percentage points below industry average LGD on senior unsecured corporate debt needs a very good explanation.

The governance layer

Model risk framework standards (SR 11-7 in the US, TRIM findings in the EU, PRA SS1/23 in the UK) all treat LGD and EAD as full first-tier models that require the same conceptual review, implementation testing, and outcomes analysis as PDs. In practice, resource often gets skewed toward PD work and LGD/EAD are validated more superficially. That gap is where regulators tend to score their findings.

What good validation looks like

• Cure definition, recovery calculation, and downturn add-on all documented and reproducible from raw workout data.
• CCFs by product segment with visible cohort dispersion — not just point estimates.
• Backtesting on truly closed cohorts, published each cycle with variance explained.
• External benchmarks referenced explicitly, with any material gap investigated.
• Independent second-line validation opinion that separates conceptually sound but calibrated conservatively from conceptually flawed and materially biased.

LGD and EAD are not glamorous, and they will never make a headline in the way a PD model can. But they carry as much of the ECL and RWA weight — often more. Any model risk framework that lets them coast is understating the true model risk profile of the credit book.