Credit Risk 102: The Loss Nobody Models Until It's Too Late

The Loss Nobody Models Until It's Too Late

The loan defaulted exactly when the model said it would. The borrower missed payment number seven, right in the middle of the forecast window. The credit committee had priced for it, reserved for it, and when it happened, everyone nodded. The model worked.

Then the recovery process started, and the loss came in at nearly double what had been reserved. Not because the default was a surprise. Because nobody had modeled what happened after.

The lender had built a careful probability of default model, stress-tested it, validated it against historical cohorts. They knew with reasonable confidence which borrowers would fail to pay. What they had never separately estimated was how much they would lose once that failure occurred. They had assumed a standard recovery rate, applied it uniformly across the book, and called it conservative. It wasn't. It was a placeholder that had never been interrogated, and when the loan went into collections, the placeholder broke against reality.

This is the gap that Loss Given Default exists to fill, and it is the component of credit risk that gets the least rigorous attention until a portfolio starts bleeding in ways the top-line default rate never warned about.

A shorter version of this piece appears on LinkedIn.

The Structure Nobody Sees Until the Loss Arrives

Credit risk is not one question. It is three, asked in sequence, and most lenders stop modeling after the first.

The first question is: will this borrower default? That is what Probability of Default measures. It is the question that gets the models, the stress tests, the committee scrutiny.

The second question is: if they default, how much do we lose? That is Loss Given Default. It is the question that gets an assumption.

The third question is: how much exposure exists at the moment of default? That is Exposure at Default. It is the question that gets a balance sheet snapshot and a hope that it holds.

Expected Loss, the number a lender actually needs to reserve against, is the product of all three:

Most lenders model the first term carefully and treat the other two as constants. That works until it doesn't, and when it doesn't, the loss is always larger than the reserve.

What LGD Actually Measures

Loss Given Default is the proportion of exposure that will not be recovered after a borrower defaults. It is expressed as a percentage of the exposure at the time of default, and it sits between zero and one hundred percent.

If a borrower defaults on a loan with $100,000 outstanding, and the lender recovers $60,000 through collateral liquidation and collections, the LGD is 40%. The lender lost forty cents on every dollar that was owed at the moment of default.

That sounds simple until you try to forecast it. The recovery amount is not known at the time of underwriting. It is not known at the time of default. It is only known at the end of a recovery process that can take months or years, and the amount recovered depends on factors that have nothing to do with the borrower's creditworthiness at origination.

Collateral value at liquidation. The legal cost of enforcement. The time it takes to work through the queue. The state of the secondary market for the asset class. Whether the borrower declares bankruptcy, and if so, which chapter. Whether other creditors have priority. Whether the collateral was properly perfected in the first place.

Every one of those factors is variable, and most of them are correlated with the macroeconomic conditions that drove the default in the first place. Defaults cluster in downturns. Asset values fall in downturns. Legal queues lengthen in downturns. The LGD you observe in a benign credit environment is not the LGD you will observe when half your book defaults at once and the collateral market has frozen.

This is why assuming a constant recovery rate across a portfolio is not conservative. It is naive. Recovery rates are not stable. They are procyclical, and they are worst exactly when you need them most.

The Loan That Proved the Gap

The loan I opened with was a secured equipment loan to a small logistics company. The borrower had a solid payment history, reasonable leverage, and the loan was structured with a 70% loan-to-value ratio against the equipment. The lender's model forecast a PD of 12% over the loan's term, which was high but acceptable given the rate they were charging. They assumed a recovery rate of 50%, in line with their historical average for secured equipment loans. That implied an LGD of 50%, and an expected loss of 6% of the exposure.

The borrower defaulted in month seven. The equipment was repossessed and sent to auction. The auction took four months to organize because the logistics sector was shedding capacity and there were few buyers. When the equipment finally sold, it brought in 38% of the original appraised value. Legal and storage costs ate another 6% of the exposure. The net recovery was 32%. The realized LGD was 68%, not 50%.

The expected loss the lender had priced for was 6% of exposure. The realized loss was 8.16%. On a $100,000 loan, that is the difference between a $6,000 reserve and an $8,160 loss. Across a portfolio of two hundred similar loans, that gap compounds into a capital shortfall that was never modeled.

The lender had not failed to predict the default. They had failed to model what would happen after the default, and that failure cost them more than the default itself.

Why LGD Varies and What Drives It

LGD is not a borrower characteristic. It is a transaction and recovery characteristic, and it varies with factors that are only loosely related to the borrower's credit profile.

Collateral type and liquidity. A loan secured by a liquid financial asset will recover more than a loan secured by specialized industrial equipment. A mortgage on a residential property in a stable market will recover more than a mortgage on a commercial property in a declining sector. The collateral's liquidity is not a function of the borrower's creditworthiness. It is a function of the asset and the market for that asset at the time of liquidation.

Seniority and security. A senior secured loan will recover more than a subordinated unsecured loan to the same borrower. The borrower's default probability is the same. The loss given default is not. If the borrower has multiple creditors, the recovery depends on where you sit in the capital structure, and that is a contractual question, not a credit question.

Time to resolution. The longer it takes to recover, the more the recovery costs, and the more the collateral value can deteriorate. A loan that goes into a contested bankruptcy can take years to resolve, and the legal fees accumulate while the asset depreciates. Time is a cost, and it is a cost that is not captured in the appraised value at origination.

Macroeconomic conditions at default. This is the factor that makes LGD procyclical. In a benign environment, defaults are idiosyncratic and collateral markets are functioning. Recovery rates are high. In a downturn, defaults are correlated, collateral markets are distressed, and recovery rates collapse. The LGD you observe in 2019 is not the LGD you will observe in 2021, even if the borrower population and the collateral types are identical.

None of these factors are modeled when a lender assumes a constant recovery rate. All of them matter, and all of them vary.

The Second-Order Effect Nobody Prices

The procyclicality of LGD has a second-order effect that is almost never priced into the initial credit decision. When a portfolio starts to default in a downturn, the realized LGD rises at exactly the same time that the default rate rises. The two risks multiply.

If a lender assumes a PD of 5% and an LGD of 40%, they are pricing for an expected loss of 2%. If a downturn doubles the PD to 10% and raises the LGD to 60%, the expected loss is now 6%. The loss has tripled, not doubled, because both terms in the equation moved in the same direction.

This is the failure mode that stress tests are supposed to catch, but most stress tests stress the PD and leave the LGD assumption untouched. They model what happens if more borrowers default. They do not model what happens if the losses per default also rise. That gap is where the unexpected loss lives.

The Portfolio That Looked Safe Until It Wasn't

A lender I worked with had a portfolio of asset-backed loans, all secured by vehicles. The historical recovery rate on the portfolio was 55%, and they had used that rate in their pricing and reserving for five years. The portfolio had performed well. Defaults were low, recoveries were in line with expectations, and the book was profitable.

Then the used vehicle market softened. Not a collapse, just a 15% decline in wholesale prices over six months. The default rate on the portfolio did not change. The borrowers who were going to default still defaulted, at the same rate the model had predicted. But the recovery rate fell from 55% to 42%, because the collateral was worth less when it was liquidated.

The lender had not modeled that scenario. They had stress-tested the PD. They had not stress-tested the collateral value, because the collateral value was not part of the credit model. It was an input to the LGD assumption, and the LGD assumption was not a model. It was a historical average, applied uniformly, and it had no sensitivity to market conditions.

The result was a portfolio that looked safe by every credit metric but was quietly losing more per default than had been reserved. The expected loss was constant. The realized loss was rising. The gap showed up in the quarterly results as a reserve shortfall, and the explanation was not that the credit model had failed. It was that the credit model had never included half the risk.

How to Model LGD When Most Lenders Don't

Modeling LGD properly requires treating it as a separate estimation problem, not a fixed assumption derived from historical averages. It requires data on recoveries, not just defaults, and that data is harder to gather because recovery processes are long and the outcomes are noisy.

The starting point is to segment the portfolio by the factors that actually drive recovery. Collateral type. Seniority. Geography. Loan size. Time to default. Then estimate the recovery rate for each segment separately, using historical recovery data where it exists and market data on collateral values where it does not.

The next step is to make the LGD estimate conditional on the scenario. In a baseline scenario, use the historical recovery rate for the segment. In a stress scenario, adjust the recovery rate down to reflect the expected decline in collateral values and the lengthening of recovery times. The adjustment does not need to be precise. It needs to be directional, and it needs to move in the same direction as the PD.

The final step is to validate the model against realized recoveries, not just at the portfolio level but at the segment level. A model that gets the average recovery rate right but misses the variation across segments is not useful. The variation is where the risk is.

This is more work than assuming a constant recovery rate. It is also the only way to reserve accurately for a portfolio that will default in a downturn, which is the only scenario that matters.

Why This Matters Now

The credit environment is tightening. Default rates are rising across most asset classes, and they are rising faster in sectors where collateral values are falling. The lenders who have modeled LGD separately will see the losses coming and can adjust their reserves and their pricing. The lenders who have assumed a constant recovery rate will see the losses arrive and will have no explanation for why the reserves were wrong.

This is not a theoretical risk. It is the risk that materialized in 2008, when mortgage LGDs doubled and the loss severity on defaulted loans far exceeded what had been modeled. It is the risk that materialized in 2020, when commercial real estate collateral lost liquidity overnight and recovery timelines stretched from months to years. It is the risk that will materialize in the next downturn, and the lenders who are not modeling it now will be reserving for it after the fact.

Next week: Exposure at Default, and why the loan balance at origination is not the number that matters when the borrower stops paying.

If you are building credit models that need to hold through a cycle and stand up to regulatory review, the work is detailed and the assumptions matter: calendly.com/muhammed-adediran/30min.