PD measures the probability a borrower defaults within a given horizon; LGD estimates the fraction of exposure lost if that default happens. Together with EAD, they form the standard credit loss formula: EL = PD × LGD × EAD. Regulators expect long-run PD calibration for capital purposes and downturn-calibrated LGDs that hold up under stress, and this guide walks through both the methodology and the governance work that makes those numbers defensible.
TL;DR:
- Regulators require long-run PD calibration and downturn LGD estimates that reflect stressed recovery environments, not just historical averages.
- PD models should be built using either point-in-time or through-the-cycle assumptions, depending on their intended purpose, with the latter stabilized for capital calculations.
- Accurate LGD estimation depends on separating cure rates from true losses and applying downturn calibration based on worst-case recovery data.
- High-quality data, including collateral valuations and vintage segmentation, is critical for reliable PD and LGD modeling, especially for portfolios with limited default history.
- Model validation involves assessing discrimination, calibration, and stability through multiple tests, and stress testing is vital to prevent risk underestimation during downturns.
Automate Regulatory Model Risk Governance
Examine models against 32 qualitative criteria and resolve risk Tiers with pre-deployment checklists per OCC 2011-12 guidelines.
Table of Contents
- What Do PD, LGD, and EAD Actually Measure?
- What Do Regulators Expect From PD and LGD Models?
- How Do You Estimate PD With Limited Default Data?
- How Should You Estimate LGD and Apply Downturn Calibration?
- What Data Do You Need for Reliable PD/LGD Models?
- How Do You Validate and Stress Test PD/LGD Models?
- How Do AI Platforms Support PD/LGD Estimation in Practice?
- What Should Modelers Prioritize First?
- How RiskInMind Supports PD/LGD Estimation Workflows
- Sources
- FAQ
What Do PD, LGD, and EAD Actually Measure?
PD is a probability, LGD is a percentage of exposure, and EAD is a currency amount. Multiply the three together and you get expected loss, a formula so simple it hides how much modeling judgment sits behind each input, as MathWorks lays out in its explanation of PD, LGD, and EAD. Each parameter answers a distinct question, and mixing them up is a surprisingly common analyst error.
- PD: the likelihood a given obligor or facility defaults over a defined horizon, typically one year for regulatory capital purposes but extended to multi-year curves for lifetime provisioning.
- LGD: the share of exposure the lender does not recover after default, net of costs and time value of money.
- EAD: the outstanding exposure expected at the moment of default, accounting for undrawn commitments and amortization.
Expected loss feeds provisioning under CECL-style frameworks, while unexpected loss, the volatility around that expectation, drives regulatory capital. That split matters because a model tuned for one purpose often performs poorly for the other, particularly around whether you use point-in-time (PIT) or through-the-cycle (TTC) PD. PIT estimates track current economic conditions and suit staging and stress testing; TTC estimates smooth across a cycle and suit capital calculation, where volatility itself becomes the enemy of stability. Readers building out provisioning models alongside PD/LGD work should also see how CECL estimation connects the two.
What Do Regulators Expect From PD and LGD Models?
Basel-style IRB frameworks set a fairly specific bar, and supervisors have gotten less forgiving about vague documentation over the past several cycles. The Basel Committee's guidance on credit risk modeling practices requires PD calibration that reflects long-run averages for capital purposes, distinct from the point-in-time PDs used for IFRS 9 staging or internal stress testing. LGD carries its own conservatism requirement: models must incorporate downturn calibration, meaning LGD estimates cannot simply reflect an average economic environment.
Two additional concepts show up in nearly every supervisory review: LGD-in-default, which measures expected loss on exposures already in default, and ELBE (expected loss best estimate), the analog used for provisioning on defaulted assets. Supervisors typically probe three things during model review:
- Whether long-run PD and downturn LGD are documented with explicit conservatism adjustments, not just backward-looking averages.
- Whether LGD-in-default and ELBE calculations are separated clearly from performing-book LGD.
- Whether the institution can trace every calibration choice back to a governance decision with an owner and a date.
Statistic to remember: Downturn LGD calibration and LGD-in-default handling are among the most frequent validation failure points supervisors flag when institutions cannot evidence how those adjustments were derived.
How Do You Estimate PD With Limited Default Data?
Grade-based calibration, where you map internal risk grades to historical default frequencies, remains the workhorse approach for retail and middle-market portfolios with enough volume. Logistic regression and survival models extend that further, letting you build multi-year PD curves that behave differently from a flat one-year assumption, which matters when you need lifetime PD for provisioning rather than a single annual figure for capital.
- Choose your calibration anchor. Decide upfront whether you're building a PIT or TTC PD, because retrofitting one into the other later is painful.
- Build the term structure. Convert one-year PD into a multi-year curve using survival analysis or Markov transition matrices across rating grades.
- Address sparse defaults. For low-default portfolios, pool internal data with external benchmarks or apply Bayesian shrinkage toward an expert prior, an approach regulators accept specifically to reduce sampling variance while preserving conservatism.
- Backtest continuously. Compare predicted versus realized default rates by vintage and grade, not just in aggregate.
Pro Tip: Run your backtest by origination vintage before you look at portfolio-wide averages. A model that looks fine in aggregate can be quietly overpredicting for 2023 originations and underpredicting for 2020 ones, and averaging hides exactly that.
How Should You Estimate LGD and Apply Downturn Calibration?
LGD estimation starts with recovery cash flows, discounted back to the default date, net of collection costs. That gives you a workable per-facility LGD, but the real modeling work is in separating cure rates (accounts that self-resolve without loss) from true loss-given-default, because blending the two understates severity on the exposures that actually go bad.
- Segment LGD by collateral type, lien seniority, and jurisdiction before pooling, since a first-lien mortgage in one state recovers on a completely different timeline than a second-lien position in another.
- Apply downturn calibration by identifying the worst historical recovery period in your data (or a comparable stressed benchmark) and using that period's severity, not the long-run average, as your floor.
- Document LGD-in-default and ELBE separately, showing how each was derived from the same underlying recovery data but adjusted for different accounting purposes.
The most common pitfall is mixing data sources with inconsistent recovery timing, blending workout-unit recoveries measured over 36 months with charged-off recoveries measured over 12, which quietly biases the LGD estimate low.
Pro Tip: If your recovery data spans a merger or a change in your workout process, split the historical window there. A single LGD curve stitched across two different collections regimes will always underperform in validation.
What Data Do You Need for Reliable PD/LGD Models?
Model quality tracks input quality almost every time, and that "every time" is not an exaggeration. Practitioners consistently find that output reliability depends more on data quality than on model architecture choice, particularly for LGD, where sparse recovery histories are the norm rather than the exception.
- Full origination-to-resolution histories, including collateral valuations at multiple points in time, not just at origination.
- Cure and recovery flags that let you separate self-resolved accounts from true defaults.
- Vintage-level segmentation so pooling doesn't blend structurally different cohorts.
- External benchmark data or industry pools for segments too thin to model internally, a use case where peer benchmarking tools add real value.
- Documented, reproducible data transformations, since an undocumented pipeline is a validation finding waiting to happen.
Collateral valuation timing deserves particular attention: if your LGD calculation uses a valuation from 18 months before default, you're measuring against stale collateral value, not the value actually realized at recovery.
How Do You Validate and Stress Test PD/LGD Models?
Validation is not a single test, it's a set of complementary checks that together answer whether the model discriminates well, calibrates accurately, and stays stable over time. The Bank of England's guidance on model risk and validation frames this as a lifecycle discipline, not a one-time exercise.
- Discrimination: AUC or Gini coefficient to confirm the model ranks risk correctly across the portfolio.
- Calibration: Brier scores and decile-level backtests comparing predicted versus actual default and loss rates.
- Stability: population stability index and characteristic stability index to catch drift in the underlying portfolio.
- Scenario simulation: Monte Carlo techniques to test model performance against adverse, low-probability scenarios, an approach Deloitte identifies as essential for surfacing "known unknowns" that historical backtests miss.
Statistic to remember: Regulators and industry commentary consistently stress that models are tools, not crystal balls; rigorous stress testing is what prevents systematic underestimation of risk during a downturn. A presentation-ready validation pack should include versioning history, independent validator sign-off, and a documented remediation log, points covered in more depth in this CECL model validation walkthrough.
How Do AI Platforms Support PD/LGD Estimation in Practice?
Machine learning can sharpen PD and LGD predictions, but it raises governance stakes: explainability and monitoring become non-negotiable once a model moves beyond logistic regression. Applied well, automated platforms translate the estimation and validation work above into operational capability rather than a one-off analytics project.
- Automated data pipelines that preserve lineage from origination through resolution, cutting the manual reconciliation that usually delays validation.
- Scenario simulation and drift monitoring that flag when a PD or LGD model needs recalibration before a backtest failure forces the issue, a discipline detailed in this drift monitoring plan.
- Audit-ready dashboards that assemble documentation supervisors expect without a separate reporting build each cycle.
What Should Modelers Prioritize First?
Chasing model sophistication before data quality is backward. A logistic regression built on clean, well-lineaged data will outperform a gradient-boosted model built on inconsistent recovery histories almost every time, and validators know the difference immediately. If there's one underappreciated fix, it's this: document your downturn LGD assumptions with the same rigor you'd document a PD model, including exactly which historical period you treated as the downturn benchmark and why. That single artifact resolves more supervisory findings than any model upgrade.
— Raj
How RiskInMind Supports PD/LGD Estimation Workflows
Building and maintaining defensible PD and LGD models means juggling data lineage, recovery segmentation, scenario simulation, and audit documentation, often across separate spreadsheets and one-off scripts that nobody fully trusts by exam time. Riskinmind consolidates that workflow into one platform: automated data pipelines handle lineage tracking, its simulation tools run the scenario testing supervisors expect, and drift monitoring flags recalibration needs before a backtest failure does it for you.
Some AI risk management platforms offer security certifications and real-time processing, helping credit teams access audit-ready dashboards instead of managing disconnected validation memos. For institutions evaluating how a modern underwriting stack handles PD estimation end to end, the loan application platform is the practical starting point. Request a demo to see how your own portfolio data maps onto these workflows before your next model review cycle.
This article is general information, not a substitute for advice from a qualified financial advisor. Consult a qualified financial professional about your own circumstances before acting on anything here.
Sources
- MathWorks — PD/LGD/EAD explanation
- Basel Committee — Credit risk modelling: current practices and applications
- Bank of England — model risk and validation guidance
FAQ
How Are PD and LGD Calculated?
PD is typically calculated through grade-based calibration or statistical models like logistic regression against historical default outcomes, while LGD is calculated from discounted recovery cash flows net of collection costs, adjusted for downturn conditions per regulatory guidance.
What Is the Difference Between PD and LGD?
PD is the probability a borrower defaults within a given horizon, while LGD is the percentage of exposure the lender loses if that default actually occurs; multiplying PD by LGD and EAD produces expected loss.
What Is the Difference Between EAD, PD, and LGD?
PD measures default likelihood, LGD measures loss severity as a percentage, and EAD measures the exposure amount outstanding at the moment of default, with all three multiplying together to form the expected loss formula.
How Is LGD (Loss Given Default) Calculated?
LGD is calculated by discounting actual or projected recovery cash flows back to the default date, netting out collection costs, and separating cured accounts from true losses, then applying downturn calibration to reflect stressed recovery conditions rather than long-run averages.
Why Does Downturn Calibration Matter for LGD?
Downturn calibration prevents LGD estimates from understating losses during economic stress, since regulators require LGD to reflect the worst historical recovery environment rather than an average across the full cycle.