Loss Given Default (LGD) Modeling and Recovery Rates: A Practical Estimation Guide

Introduction

Loss Given Default, or LGD, decides what a default actually costs a bank. Probability of Default (PD) tells a lender how likely a default is. LGD tells them what it will cost when it happens. LGD is often the harder number to pin down. PD stays fairly stable across a cycle. Recovery outcomes don’t — they swing with collateral quality, legal jurisdiction, workout efficiency, and the state of the economy at the time of default.

This matters directly for capital and provisioning. Both the Internal Ratings-Based (IRB) approach to regulatory capital and the IFRS 9 Expected Credit Loss (ECL) framework use LGD as a direct multiplier: Expected Loss = PD × LGD × EAD. Underestimate LGD by even a few percentage points on a large secured portfolio, and a bank can materially understate its capital requirement or provisioning charge. Risk practitioners often find that LGD, not PD, generates the biggest disagreements between banks, auditors, and regulators. Recovery data is sparser. Workout periods run longer. And the underlying distribution of outcomes looks nothing like the neat bell curve that PD models assume.

Risk analysts, credit modelers, and finance professionals preparing for FRM, PRM, or model validation roles need to understand LGD well. That means knowing how recovery rates behave across secured and unsecured exposures, how collateral gets valued under stress, which statistical techniques actually fit LGD’s unusual distribution, and what regulators — from the Basel Committee to the Reserve Bank of India — expect from a defensible LGD framework. This guide walks through each building block with the technical depth a working risk professional needs.

What Is Loss Given Default (LGD)?

Definition. Loss Given Default is the share of an exposure a lender expects to lose if a borrower defaults. The calculation nets out recoveries from collateral liquidation, guarantees, insurance, and the workout process, then subtracts the costs of collecting. Analysts express LGD as a percentage of Exposure at Default (EAD).

The foundational relationship is simple:

LGD = 1 − Recovery Rate

The Recovery Rate is the present value of everything a lender ultimately recovers, divided by the exposure at the point of default. If a bank recovers 65% of a defaulted loan’s value after liquidating collateral and pursuing legal remedies, LGD works out to 35%.

A fuller formula used in practice

LGD = [EAD − PV(Recoveries) + PV(Workout Costs)] / EAD

Breaking down the components

Exposure at Default (EAD): The outstanding balance — principal, accrued interest, and undrawn commitments where applicable — at the moment of default.
Recoveries: Cash or asset value a bank collects through collateral sale, guarantee invocation, litigation settlement, or restructuring, often over a multi-year workout period.
Discounting: Recoveries arrive over time, not instantly, so analysts discount them back to the default date. The discount rate typically reflects the risk of the recovery cash flows — often the original effective interest rate or a risk-adjusted rate set by internal policy.
Workout costs: Legal fees, collateral maintenance costs, recovery agent fees, and administrative overhead all reduce net recovery and push LGD higher.

Why this matters for capital and provisioning: LGD isn’t a static number sitting in a spreadsheet. It flows directly into risk-weighted asset calculations under Basel and into Stage 1, 2, and 3 provisioning under IFRS 9. A retail mortgage with strong collateral coverage might carry an LGD of 10–20%. An unsecured personal loan or trade finance exposure might sit at 45–75%. That gap — often three to five times higher for unsecured versus well-secured exposures — shows why collateral structure, not just borrower creditworthiness, drives so much of a bank’s capital planning.

A simple example: A bank holds a ₹10 crore secured corporate loan. The borrower defaults. The bank spends ₹40 lakh liquidating the pledged collateral and litigating a personal guarantee, and recovers ₹7 crore in present value terms after 18 months. Net recovery works out to ₹7 crore minus ₹0.40 crore, or ₹6.6 crore. That’s a 66% recovery rate — and a 34% LGD.

Recovery Rate Analysis: Secured vs Unsecured Exposures

Recovery rates are the mirror image of LGD. The single biggest driver of variation in recovery outcomes is whether — and how well — an exposure is secured.

Secured exposures

When collateral backs a loan, the lender holds a claim on a specific asset it can liquidate or repossess. Recovery rates on secured exposures run generally higher and less volatile, but they still vary enormously by collateral type:

Real estate (commercial and residential): Usually the most stable collateral class. Recovery rates often land in the 60–85% range, depending on property type, location, and market liquidity at the time of sale. Residential mortgages in developed markets tend to sit at the higher end; commercial real estate swings more with the cycle.
Financial collateral (cash, government securities, listed equities): Produces the highest and most predictable recoveries, often near 90–100% after haircuts, because valuation and liquidation happen fast and transparently.
Receivables and inventory: More volatile. Recovery depends heavily on the quality of the underlying receivables book and how quickly a bank can sell inventory without a distressed-sale discount.
Plant, machinery, and other physical collateral: Recovery rates run lower and more dispersed, since specialized equipment often has a thin secondary market.

Unsecured exposures

With no specific asset pledged, the lender’s claim ranks alongside other unsecured creditors in insolvency proceedings. Recovery depends on the borrower’s overall asset pool, the seniority of the claim, and the jurisdiction’s insolvency regime. Unsecured corporate exposures commonly show recovery rates of 20–40%. That’s precisely why Basel’s standard supervisory LGD for senior unsecured corporate exposures sits at 40–45% under the Foundation IRB approach, and why subordinated unsecured claims carry an even higher LGD of 75% — implying a recovery rate assumption of just 25%.

Guarantees

A third category deserves a mention: guaranteed exposures. A credible third-party guarantee — sovereign, bank, or corporate — can push recovery closer to the guarantor’s own credit strength rather than the original borrower’s. But this substitution effect only holds if the guarantee is legally robust and the guarantor itself stays out of distress. Many banks relearned this lesson during systemic downturns, when correlated defaults hit both borrower and guarantor at the same time.

The practical takeaway: recovery rate is never a single number for a portfolio. It’s a distribution shaped by collateral type, seniority, jurisdiction, and workout timing. Any LGD model that ignores this segmentation will systematically misestimate risk on the tails.

Collateral Valuation Methods for LGD Estimation

Collateral quality drives most of recovery, which makes collateral valuation one of the most consequential — and most frequently under-scrutinized — inputs into LGD estimation. Banks use three broad approaches, often in combination.

1. Market Approach

Analysts derive value from observed transaction prices for comparable assets — recent sales of similar properties, equipment, or securities. This method works best when an active, liquid market exists, such as listed securities or standard residential property in a liquid city market, because it reflects what the collateral would actually fetch if sold today. The limitation: during systemic stress, comparable transactions become scarce or reflect distressed-sale pricing themselves. That can understate true economic value, or overstate it if valuations lag a falling market.

2. Income Approach

For income-generating collateral such as commercial real estate or business assets, this method values collateral based on the present value of expected future cash flows — rental income or operating cash flows — discounted at an appropriate capitalization rate. It looks further forward than the market approach, but it’s sensitive to assumptions about occupancy, rental growth, and discount rates. All of those need independent validation to avoid overly optimistic collateral values.

3. Adjusted / Haircut Values

Regulatory and internal risk frameworks rarely accept raw market or income valuations at face value. Instead, they apply haircuts — percentage reductions that account for valuation uncertainty, price volatility, currency mismatch between the loan and collateral, and the time and cost required to liquidate. Basel’s standardised and IRB frameworks specify minimum haircuts by collateral class: small haircuts of a few percentage points for financial collateral, and larger haircuts — often 15–40% — for real estate and other physical collateral, reflecting realization risk. Banks then build internal haircut schedules on top of these regulatory floors, calibrated to their own historical liquidation experience.

Practical valuation governance issues risk teams should watch for

Stale valuations: collateral appraised at origination but never revalued through the credit cycle
Correlated collateral risk: collateral value and borrower default probability moving together — for example, a real estate developer whose loan is secured by real estate, where collateral value falls exactly when default risk rises
Valuation model independence: appraisers or valuation models controlled by the same business unit that originated the loan, which creates conflict-of-interest risk

Robust LGD models need a documented, periodically revalidated collateral valuation policy, not a one-time appraisal frozen at origination. Collateral value at the point of default — not at origination — determines actual recovery.

LGD Calculation Approaches: Market, Workout, and Statistical Methods

Banks use several distinct methodologies to derive LGD estimates, often blending them depending on data availability and asset class.

1. Market LGD

This approach uses observed market prices of defaulted debt instruments shortly after default — common for traded corporate bonds and syndicated loans. If a bond trades at 35 cents on the dollar right after default, the implied market LGD comes out to roughly 65%. The method is fast and market-based, but it only works where a liquid secondary market for distressed debt exists. That rules it out for most retail and SME lending.

2. Workout LGD

Banks use this approach most often for loan portfolios, especially retail and commercial lending without a tradable market. It tracks the actual cash flows recovered over the full workout period following default — collateral liquidation proceeds, guarantee payments, restructuring recoveries, litigation settlements — discounts them back to the default date, and nets off workout costs. Workout LGD demands a lot of data: it needs long, clean histories of defaulted accounts tracked to final resolution, which can take years. But it produces the most economically grounded estimate, because it reflects a bank’s own actual recovery experience.

3. Implied Market / Historical LGD

Analysts derive this indirectly from observed credit spreads or historical loss rates on a portfolio, back-solving for an implied LGD given known PD and observed loss experience. It works as a cross-check, or for asset classes with limited direct workout data, but it’s less precise than a direct workout LGD study.

4. Statistical / Regression-Based LGD Models

Once a bank assembles a workout LGD dataset — actual LGD outcomes for a sample of resolved defaults — it typically builds a statistical model to predict LGD for the broader portfolio. Explanatory variables include collateral type and loan-to-value ratio, industry, exposure size, time to resolution, macroeconomic conditions at default, and borrower or facility characteristics. This is where the modeling choice becomes critical, because LGD’s statistical properties are unusual and standard linear regression handles them poorly — the next section explains why.

Segmentation matters as much as method. Whichever base approach a bank uses, best practice segments LGD estimation by product type, collateral class, and — where sample size allows — geography or industry, rather than fitting one LGD model across a heterogeneous portfolio. A single blended LGD for “all secured corporate loans” hides enormous variation. A loan secured by liquid financial collateral behaves very differently from one secured by specialized machinery.

Beta Regression for LGD Modeling

Why Ordinary Least Squares (OLS) regression fails for LGD. LGD data has three statistical properties that violate the core assumptions of OLS:

Bounded support: LGD sits between 0 and 1 (it can occasionally exceed 1 in cases of negative recovery, where workout costs surpass recoveries, though banks typically cap this). OLS assumes an unbounded continuous response, so it will happily predict LGD values below 0% or above 100% for extreme covariate combinations — numbers that make no economic sense.
Non-normal, often bimodal distribution: LGD outcomes frequently cluster near the two extremes: many defaulted loans get fully recovered (LGD near 0), while others are almost entirely lost (LGD near 1), with fewer observations in between. This U-shaped or bimodal pattern breaks OLS’s assumption of normally distributed residuals.
Heteroscedasticity: The variance of LGD outcomes isn’t constant across the range of predicted values — it typically differs by collateral type and loan segment. That further violates OLS assumptions and produces unreliable standard errors and confidence intervals.

Why Beta regression fits better. The Beta distribution lives naturally on the (0, 1) interval and can take on a wide range of shapes — including the U-shaped, bimodal pattern common in LGD data — by adjusting its two shape parameters. A Beta regression model links the mean of this Beta-distributed response to explanatory variables through a logit-type link function. That guarantees predicted LGD values stay within the valid 0–1 range, something OLS can’t promise.

Practical advantages for LGD modeling

Predictions stay automatically bounded, which removes the need for ad hoc capping or flooring rules on OLS outputs
The model can separately parameterize the mean and the dispersion of LGD, so analysts can model not just expected LGD but how much it varies by segment — useful for stress testing and downturn LGD estimation
Tail behavior calibrates better, which matters because LGD’s tails — very low or very high loss — are exactly where capital and provisioning are most sensitive

A common workaround worth knowing: LGD data often includes exact zeros and ones, which creates a practical problem since pure Beta distributions require values strictly between 0 and 1. Analysts typically handle this with a small transformation (the Smithson-Verkuilen adjustment, for example) or a zero-and-one-inflated Beta regression, which models the probability mass at the boundaries separately from the continuous distribution in between. This usually fits better than forcing all observations into the open interval.

Analysts also use other techniques alongside or instead of Beta regression: Tobit models for censored LGD data, fractional response regression, and machine learning approaches such as gradient boosting or random forests. The machine learning methods can capture non-linear interactions between collateral, borrower, and macroeconomic variables, but they demand more careful validation and explainability work before regulators will accept them.

Basel III and RBI LGD Standards

Basel Foundation IRB (F-IRB) supervisory LGD values

Under Basel III, banks using the Foundation IRB approach don’t estimate their own LGD for corporate exposures — they apply supervisor-prescribed values instead. Senior unsecured claims on corporates, sovereigns, and banks get a standard LGD of 45%, recalibrated to 40% for senior unsecured exposures specifically to non-financial corporates under the Basel III finalisation reforms. All subordinated claims carry a 75% LGD. For secured exposures, a prescribed formula adjusts LGD downward using collateral-specific haircuts, subject to input floors that stop banks from modeling implausibly low LGDs even with strong collateral coverage.

Advanced IRB (A-IRB)

Banks approved for the Advanced IRB approach can use their own internal LGD estimates, built from workout data as described earlier. These estimates still face regulatory floors, and they must reflect downturn conditions — meaning LGD estimates need to capture the possibility that recoveries fall during an economic downturn, when collateral values drop and workout costs rise, not just long-run average experience. This “downturn LGD” requirement ranks among the more technically demanding parts of IRB model validation, since a bank needs historical data spanning at least one full economic cycle, including a stress period, to calibrate it credibly.

RBI’s regulatory position

India has taken a notably different path from many global peers. Rather than adopt the IRB approach for credit risk capital, the Reserve Bank of India issued the Commercial Banks – Capital Charge for Credit Risk (Standardised Approach) Directions, 2026, mandating the Standardised Approach for all commercial banks under its jurisdiction — excluding small finance banks, payments banks, and local area banks — effective April 1, 2027. No Indian bank currently operates under IRB for credit risk capital purposes.

LGD becomes central for Indian banks instead through the RBI’s Expected Credit Loss (ECL) framework, finalized in April 2026 and also effective April 1, 2027. It requires scheduled commercial banks to move from the current incurred-loss provisioning model to a forward-looking ECL approach built on internally estimated PD, LGD, and EAD — closely mirroring IFRS 9. Because IFRS 9 and RBI’s ECL directions stay principle-based rather than prescriptive on LGD methodology, RBI has added product-wise prudential floors for Stage 1 and Stage 2 provisioning as a regulatory backstop, along with specific expectations around collateral valuation techniques for LGD and the distressed value of collateral for retail exposures. Individual banks carry the burden of building defensible, well-governed internal LGD models — the exact skill set this guide covers — even though India hasn’t adopted IRB for capital purposes.

Case Study: Recovery Rates in Practice

A widely studied example from the 2008–2009 global financial crisis shows why segmentation and downturn calibration matter so much in LGD modeling. Post-crisis studies of European and US bank loan portfolios found that recovery rates on defaulted bank loans and bonds fell sharply during the crisis compared to long-run historical averages. Unsecured and subordinated corporate exposures took the biggest hit. Recoveries partially rebounded in later years as distressed-asset markets stabilized and collateral prices recovered.

Secured exposures with liquid financial or real estate collateral held up better than unsecured corporate exposures over the same period. Even real-estate-backed recoveries took a hit, though, dragged down by the simultaneous collapse in property prices — a textbook case of the collateral correlation risk covered earlier, where the value of the security and the borrower’s default risk deteriorate together.

Risk teams draw a consistent lesson from this and similar downturn episodes: an LGD model calibrated purely on a benign credit cycle will understate loss severity exactly when it matters most. That’s why Basel’s Advanced IRB framework mandates downturn LGD estimation rather than allowing long-run average LGD alone. It’s also why banks building internal LGD models — for IRB or for IFRS 9/ECL purposes — need workout data spanning a genuine stress period, not just a benign multi-year average, before regulators or internal capital committees will accept the model.

Frequently Asked Questions

Q1. What is the difference between LGD and recovery rate?

Recovery rate measures the share of an exposure a lender gets back after default. LGD is simply 1 minus the recovery rate. They’re two sides of the same calculation — recovery rate measures what a bank regains, LGD measures what it loses.

Q2. Why can’t banks just use OLS regression to model LGD?

LGD sits between 0 and 1, often clusters near full recovery or full loss, and shows different variance across segments — all of which break OLS’s core assumptions. Beta regression and related bounded-response models handle these properties directly and keep predictions within the valid range.

Q3. What is downturn LGD, and why does it matter?

Downturn LGD reflects economic stress conditions — depressed collateral values and elevated workout costs — rather than long-run average experience. Basel’s Advanced IRB framework requires it because average-cycle LGD alone can understate capital needs exactly when losses are most likely to spike.

Q4. Does India use the IRB approach for LGD under Basel?

No. The RBI has mandated the Standardised Approach for credit risk capital for all commercial banks under its jurisdiction, effective April 1, 2027, rather than adopting Internal Ratings-Based approaches. LGD modeling still matters for Indian banks under RBI’s parallel Expected Credit Loss (ECL) provisioning framework, which also takes effect April 1, 2027.

Q5. How does collateral type affect LGD?

Collateral type drives recovery outcomes more than almost any other factor. Financial collateral — cash, government securities — typically produces the highest, most stable recoveries. Real estate stays generally strong but cyclical. Receivables, inventory, and specialized physical assets tend to show lower and more volatile recovery rates, and unsecured exposures show the lowest recoveries of all.

Conclusion

LGD estimation sits at the intersection of statistics, legal recovery process, and collateral economics, which is exactly why it resists simple modeling shortcuts. A defensible LGD framework needs clean segmentation by collateral and seniority, a disciplined collateral valuation methodology, a modeling technique that respects LGD’s bounded and skewed distribution, and calibration that accounts for downturn conditions rather than relying solely on benign-cycle averages. Whether you’re building models under Basel’s Advanced IRB approach or an IFRS 9/RBI-style ECL framework, the fundamentals covered here — from workout LGD construction to Beta regression — form the technical backbone of credible LGD estimation.

Ready to Build These Skills Hands-On?

Understanding the theory behind PD, LGD, and EAD is the first step. Building bankable, interview-ready models — in Python or SAS, on real credit datasets, aligned to Basel and IFRS 9 — is what actually moves a career forward.

Explore Dexlab Analytics’ Credit Risk Modeling certification program to build PD, LGD, and EAD models from scratch, work through IFRS 9 ECL frameworks, and learn model validation techniques used by practicing risk teams.