Quick answer: Loss Given Default (LGD) estimates the proportion of exposure a bank ultimately loses conditional on default. It is not the probability that default occurs. A useful LGD model follows the post-default cash-flow path: how much exposure is outstanding when default occurs, which recoveries arrive, when they arrive, what direct and indirect workout costs are incurred, whether collateral can actually be realised, whether the borrower cures or restructures, and how stressed economic conditions change recovery values. The mathematical task is therefore a discounted recovery problem rather than a simple “collateral value divided by loan balance” ratio.

PD asks whether default happens. LGD asks what remains lost after the recovery process finishes.

Page role: severity after default

This article owns a narrow job inside the finance-and-banking algorithms lane: how banks estimate loss severity after default. It does not replace probability-of-default modelling, exposure-at-default modelling, or expected-credit-loss accounting. Those components connect, but they answer different questions.

The European Banking Authority’s in-force guidelines on PD and LGD estimation were developed to reduce unjustified variability in internal-model outcomes while preserving risk sensitivity. They explicitly address LGD estimation for non-defaulted exposures and parameters for defaulted exposures. See EBA Guidelines on PD estimation, LGD estimation and treatment of defaulted assets.

1. The basic identity

For a defaulted exposure with exposure at default EAD and present value of net recoveries R, a teaching definition is:

LGD = (EAD − PV(net recoveries)) / EAD.

If EAD is S$1,000,000 and discounted net recoveries are S$650,000, LGD is 35%.

The simplicity of the ratio can be misleading. “Net recoveries” may arrive over years and can contain collateral sale proceeds, borrower payments, guarantor payments, insurance, legal recoveries, restructuring cash flows and recoveries from asset sales, minus direct and indirect workout costs. Every component needs a definition, date and evidence trail.

2. Recovery timing changes economic loss

Receiving S$600,000 one month after default is economically different from receiving S$600,000 four years later. The second recovery leaves the bank without the funds for longer and exposes the workout to more uncertainty and cost.

If recovery cash flow C_t arrives at time t and r is the relevant discount rate, a simplified present value is:

PV(recoveries) = Σ C_t / (1 + r)^t.

Suppose two defaults both recover S$700,000 gross. One recovers quickly; the other takes several years and incurs legal and servicing expense. Their undiscounted recovery rates look identical, but their economic LGDs do not.

3. Workout LGD reconstructs the post-default cash-flow path

A workout-LGD approach follows actual defaulted cases from the default date through recovery resolution. For each case, the bank should be able to reconstruct:

exposure at the default reference date;
cash collections after default;
collateral proceeds and sale dates;
guarantee or insurance payments;
asset-sale proceeds;
direct legal, repossession and realisation costs;
indirect workout costs where the methodology includes them;
write-offs and later recoveries;
final closure or an explicit treatment of incomplete workouts.

The algorithm is therefore partly a data-lineage problem. If historical recoveries are spread across collections, legal, collateral and finance systems, the model must first reconstruct one economic case before estimating any parameter.

4. Collateral value is not recovery value

Suppose a property is appraised at S$1 million when the loan is performing. That does not imply a S$1 million recovery after default. The bank may face a lower stressed sale price, senior claims, selling costs, legal delay, maintenance expense or a weak market in which the asset takes time to realise.

A more useful collateral-recovery chain is:

appraised value → stressed market value → enforceable collateral value → sale proceeds → net proceeds after costs → discounted recovery.

Each arrow can fail. Documentation may be defective. The collateral may rank behind another claim. The sale may take longer than expected. The asset can become less liquid exactly when defaults rise.

5. Wrong-way recovery risk

A common weak assumption is that collateral value is independent of borrower distress. Commercial-property lending provides an obvious counterexample: borrowers may default because rents, occupancy and refinancing conditions weaken at the same time property prices fall.

Then two things happen together:

default frequency rises; and
LGD rises because collateral recovery deteriorates.

This joint deterioration is why downturn LGD matters. A recovery estimate calibrated only on benign years can be most optimistic when the bank most needs conservatism.

6. Downturn LGD is not “average LGD plus a random buffer”

Regulatory downturn LGD asks for loss severity appropriate to adverse economic conditions. The EBA’s downturn-LGD guidelines distinguish approaches according to data availability and require institutions to identify relevant downturn periods and quantify the calibration target with evidence rather than merely adding a generic margin.

See EBA Guidelines on the estimation of LGD under an economic downturn.

At a conceptual level, downturn LGD can increase because:

collateral prices fall;
time to recovery lengthens;
legal/workout capacity becomes congested;
secondary markets become less liquid;
more borrowers default simultaneously, weakening bargaining and sale conditions;
guarantors become weaker in the same macroeconomic shock.

The mechanism matters because different portfolios have different downturn drivers. A mortgage portfolio and an unsecured corporate portfolio need not share the same recovery-cycle structure.

7. A worked recovery example

Assume a corporate loan defaults with EAD = S$2,000,000. Over the next three years the bank receives:

Time after default	Cash flow	Source
6 months	S$150,000	Borrower collections
18 months	S$700,000	Collateral sale
30 months	S$300,000	Guarantor settlement
Throughout workout	−S$100,000	Legal and workout costs

Undiscounted net recovery is S$1,050,000, implying an undiscounted loss of S$950,000 or 47.5%. After discounting later recoveries back to the default date, net recovery is lower and LGD is therefore higher.

The example illustrates why a model needs both amount and timing. A recovery ledger that stores only final totals throws away information needed to estimate economic loss.

8. Cure, restructuring and return to performing status

Not every defaulted exposure ends in liquidation. Some borrowers cure, restructure successfully or return to a performing state. The model must define how such cases are treated.

A cure should not be assumed to mean zero LGD. The bank may already have incurred interest shortfalls, concessions, legal expense or delayed cash flows. Conversely, a default label does not imply a 100% loss if the borrower ultimately repays most of the exposure.

This connects to How Banks Forecast Loan Delinquency and Cure: delinquency-state migration predicts the route into and out of distress; LGD quantifies the economic severity once default has occurred under the chosen definition.

9. Incomplete workouts create censored data

Recent defaults are often unresolved. If the bank calibrates LGD only on completed workouts, it can create selection bias because fast-resolving cases may be systematically easier or harder than long-running cases.

Alternatives include:

restricting calibration to sufficiently mature cohorts and accepting fewer observations;
estimating remaining recoveries for unresolved cases under a validated method;
using survival/hazard methods for time-to-recovery;
segmenting by workout age;
applying explicit margins where uncertainty is material.

There is no free solution. Excluding unresolved cases loses information; estimating them adds model risk. The method should make that trade-off visible.

10. Segmentation: one LGD cannot represent every recovery mechanism

LGD commonly varies with:

seniority;
collateral type and lien position;
loan-to-value;
borrower type;
jurisdiction and enforcement regime;
product structure;
guarantee quality;
time in default;
macroeconomic state.

A model can segment explicitly or use regression/tree-based methods that estimate conditional LGD. But segmentation has a bias-variance trade-off: too broad and the model mixes unlike recoveries; too narrow and each cell contains too little default data.

11. Fractional outcomes need careful model choice

LGD is often represented between 0 and 100%, but realised workout calculations can produce unusual values if costs exceed recoveries, if additional exposure arises after the reference date, or if the accounting and modelling definitions differ. A simple ordinary least-squares regression can predict economically impossible values below 0 or above 1 unless constrained or transformed.

Possible modelling approaches include:

segment averages;
beta or fractional regressions where appropriate;
two-stage models separating cure/recovery occurrence from severity;
decision trees and ensembles;
survival models for recovery timing;
cash-flow simulation at individual workout level.

The right choice depends on the question, data and validation evidence—not on which algorithm sounds most advanced.

12. LGD floors and regulatory use are not the same as economic recovery forecasts

Regulatory capital frameworks can impose supervisory LGDs, floors or other constraints depending on the exposure class and approach. Those rules answer a prudential-capital question. An internal workout forecast may answer a different question about expected recovery for pricing, collections strategy or provisioning.

That distinction mirrors a recurring banking principle: same variable name does not guarantee same model purpose. A regulatory LGD, accounting LGD and economic pricing LGD can be related but should not be silently substituted for one another.

For the Basel IRB context, see Basel Framework CRE31 and CRE32.

13. Evidence polarity: what should make the model more pessimistic?

Evidence should be able to move LGD in both directions. Examples that can justify a higher loss estimate include:

longer recovery times;
lower collateral sale proceeds relative to valuation;
higher workout costs;
weaker guarantor performance;
higher loss severity in downturn cohorts;
more unresolved cases than historical calibration assumes.

Evidence that may support lower LGD includes faster recoveries, stronger collateral realisation, improved seniority, more effective guarantees or better workout performance—provided the change is durable and not merely the result of an unusually strong economy.

A model that can only explain why LGD should fall is not a risk model. It is a one-way justification machine.

14. Counterexamples that break naive LGD rules

“Secured means low LGD.” Counterexample: collateral is junior, illiquid, legally unenforceable or highly correlated with borrower distress.
“High recovery rate means low economic loss.” Counterexample: recovery arrives years later after large costs.
“Cure means no loss.” Counterexample: the bank grants concessions and absorbs a material economic shortfall before cure.
“Average LGD is enough.” Counterexample: loss severity doubles in recessions precisely when defaults become numerous.
“Recent defaults are the freshest evidence.” Counterexample: they are unresolved, so final recovery severity is unknown.

15. The LGD algorithmic pipeline

Define default and recovery closure consistently.
Freeze the correct EAD reference amount.
Reconstruct post-default cash flows by case.
Capture collateral, guarantees, costs and dates.
Discount net recoveries to the chosen reference date.
Calculate realised workout LGD.
Handle cures, restructurings and unresolved cases explicitly.
Segment by recovery mechanism and risk driver.
Estimate conditional LGD using an appropriate method.
Identify downturn periods and stress recovery drivers.
Apply required regulatory constraints for the model’s use.
Validate calibration, ranking and stability.
Backtest predicted versus realised recovery amount and timing.
Update after legal, collateral, servicing or macroeconomic regime changes.

16. Failure modes and weak links

Collateral shortcut. Appraisal value substitutes for realised net recovery.
Timing blindness. Late cash is valued like immediate cash.
Cost omission. Legal and workout costs vanish from the numerator.
Benign-cycle calibration. Strong years dominate the sample and understate downturn severity.
Incomplete-workout bias. Only resolved defaults enter the dataset.
Definition drift. Default, cure or closure rules change through history without adjustment.
Jurisdiction pooling. Recoveries from different legal regimes are treated as interchangeable.
Guarantee double count. Guarantee value is reflected both in LGD and elsewhere without consistent treatment.

17. Diagnostics and falsifiers

What percentage of predicted recovery is collateral, borrower cash, guarantees and other sources?
What is median and tail time-to-recovery?
How much does LGD change when recovery timing is stressed?
How far below appraisal values are realised collateral sale proceeds?
Do downturn cohorts show higher LGD after controlling for portfolio mix?
Which unresolved defaults could materially change calibration?
Does model error cluster by seniority, jurisdiction or collateral type?
What evidence would force the bank to increase LGD tomorrow?

Suppose someone claims, “The loan is 80% collateralised, so LGD cannot exceed 20%.” A falsifier is a workout in which collateral is sold at a stressed discount, recovery takes two years, senior claims absorb part of the proceeds and legal costs are material. Nominal collateral coverage is not an upper bound on economic LGD.

18. Verification and update triggers

reconcile recovery cash flows to servicing, legal and finance records;
audit collateral lien, ranking and enforceability data;
backtest both recovery amount and recovery timing;
compare simple segment averages with more complex model outputs;
test performance separately in downturn periods;
revalidate after changes in workout strategy, law or collateral markets;
track unresolved-case assumptions until actual recoveries replace estimates;
retain historical model versions so realised defaults can be compared with the prediction actually made at the time.

Connections across the finance-and-banking algorithms lane

Credit scoring and PD — how likely default is.
Revolving-credit utilisation and EAD — how large exposure may be when default occurs.
Expected credit loss — combines credit-state forecasts with severity and exposure.
Model validation — the governance layer that checks LGD assumptions against realised evidence.

Research anchors

The deeper lesson

LGD is not a label attached to a default. It is a reconstruction of what happens after the relationship breaks. Recovery amount matters. Recovery timing matters. Collateral enforceability matters. Costs matter. The economic cycle matters. A strong LGD model therefore remains close to the actual workout path and keeps the model correctable by realised cash flows. When the recovery mechanism changes, the parameter should change with it.

Educational note: This article explains public credit-risk mathematics. It is not lending advice, recovery advice, legal advice, valuation advice or institution-specific regulatory guidance.

Small Group Tutorials

How Banks Estimate Loss Given Default: Workout Cash Flows, Collateral Recoveries, Discounting, Cure, Downturn LGD and Validation