FRM Exam Part I · Learning From Financial Disasters

Model Risk and Valuation Failures in Financial Disasters

Updated 11 October 2026 · Fact-checked

Model risk is the chance of loss because a model is wrong, misused or poorly validated. Disasters such as the 2007-2009 crisis (Gaussian copula, CDOs) and the London Whale (VaR changes) show the pattern. To answer exam questions, name the flawed assumption, the control that failed, and the lesson.

Understand Model Risk and Valuation Failures

A model turns assumptions into numbers: prices, risk measures, capital. Model risk is the loss that arises when the model is wrong, applied outside its valid range, implemented with errors, or trusted without challenge. The model is a simplification. The danger is forgetting that.

The Gaussian copula is the classic case. A CDO's loss depends on how defaults cluster. The copula links individual default probabilities through one correlation number, and the dependence is modelled as normal-like. This gives little weight to many names defaulting together in a crisis (weak tail dependence). Correlation was also treated as stable and was calibrated on benign periods. When US house prices fell, defaults moved together and senior tranches that were rated very safe lost value.

Valuation failures also came from complexity and thin data. CDOs and CDO-squared products were rarely traded, so prices came from models (mark-to-model) rather than markets. Ratings were used as a substitute for analysis. Firms treated the AAA tranche as nearly risk-free and held it in large size, often funded short term. Model output, ratings and incentives all pointed the same way.

VaR failures follow a similar logic. VaR says nothing about losses beyond the cutoff. Models calibrated on calm history understate risk. Positions that are illiquid cannot be closed in the assumed holding period. The London Whale loss at JPMorgan's Chief Investment Office (2012) involved a synthetic credit portfolio. A new VaR model was introduced, reportedly implemented through spreadsheets with manual and formula errors, and it lowered the reported VaR when limits were being breached. Oversight was weak, and marks on illiquid positions were questioned. The lesson is governance: independent validation, change control and challenge.

So for each case, ask four things. What was the assumption? Why did it fail? Which control should have caught it? What is the fix? Common fixes are independent model validation, stress tests beyond history, reserves for model uncertainty, using several models, and senior management who understand the limits.

Key formulas to remember

Value at Risk (definition): VaR(α) = loss level L such that P(Loss > L) = 1 − α; Gives a threshold, not the size of losses beyond it. Expected shortfall measures the average loss beyond VaR.
Model risk sources: Model risk = wrong assumptions + implementation error + misuse outside valid range + weak validation; A checklist, not a numeric formula. Classify each case into one or more sources.
Default correlation effect on a senior tranche: Higher default correlation → higher probability of many joint defaults → senior tranche loses more, equity tranche loses less (relative); Direction rule. Senior tranches are short correlation; equity tranches are long correlation in value terms.
Normal VaR scaling: VaR(h days) = VaR(1 day) × √h; Holds only for independent, identically distributed returns with zero mean. It understates risk if returns are serially correlated or positions are illiquid.

How to solve Model Risk and Valuation Failures questions

Use this method for any case-study or concept question on model risk and valuation failures.

1Identify the case or product: Gaussian copula and CDOs, a VaR model, or the London Whale.
2State the model's key assumption, for example constant correlation, normal-like dependence, calm-period calibration, or tradable liquid positions.
3Say how reality differed: correlations rose, tails were fat, markets froze, or prices were not observable.
4Name the control that failed: validation, independent price verification, limit monitoring, change control, or governance.
5Link the failure to the loss: for example, senior tranches rated AAA lost value when defaults clustered.
6Check each answer option for overstated words such as 'always', 'eliminates' or 'proves'.
7Pick the option that names the cause and the remedy together, such as stress testing, validation or model reserves.

Quickest way: Assumption-Failure-Control scan

When to use it: For multiple-choice questions where you have about two minutes and four plausible options.

Underline the model or product in the stem.
Recall its one weak assumption (copula: stable correlation and thin tails; VaR: ignores losses beyond cutoff; London Whale: model change and weak oversight).
Remove options that blame only bad luck or only one person.
Remove options with absolute claims.
Choose the remaining option that fits the weak assumption.

Common mistakes in Model Risk and Valuation Failures

Saying the Gaussian copula was 'wrong mathematics'.
Students remember the blame but not the reason.
Fix: Say the mathematics was fine; the assumptions (single stable correlation, thin joint tails, calm data) and over-reliance on it were the problem.
Claiming a higher correlation hurts all tranches.
Correlation is treated as simply 'more risk'.
Fix: Higher correlation raises senior tranche risk and lowers equity tranche risk, because losses become all-or-nothing.
Thinking VaR states the maximum possible loss.
The word 'risk' suggests a worst case.
Fix: VaR is a loss threshold at a confidence level. Losses beyond it can be much larger.
Blaming the London Whale loss on VaR being 'inaccurate' only.
Students ignore the governance side.
Fix: Include the model change, implementation errors, limit breaches, weak validation and questions about position marks.
Treating credit ratings as a model validation.
Ratings looked like independent confirmation.
Fix: Ratings relied on similar models and assumptions, and could not replace a firm's own analysis.
Applying √h scaling to every situation.
It is a quick formula that is easy to memorise.
Fix: State its conditions: iid returns, zero mean. Note that illiquid positions need longer liquidation horizons.

Worked examples

Example 1

A risk manager uses a one-factor Gaussian copula with a constant correlation of 0.30 calibrated on 2003-2005 data to value a CDO. In 2007 default correlations rise sharply. Which tranche is most likely to suffer a larger loss in value than the model predicted, and why?
A. Equity tranche, because it absorbs first losses
B. Senior tranche, because joint defaults become more likely
C. All tranches equally, because correlation is a single number
D. No tranche, because the portfolio's expected loss is unchanged

Show the solution

Identify the change: realised default correlation is higher than the 0.30 assumed.
Recall the effect: higher correlation increases the probability of many simultaneous defaults.
The expected portfolio loss is unchanged, but the distribution shifts: more probability at both very low and very high losses.
The senior tranche only loses when losses are very large, so it is the one that gains probability of loss.
The equity tranche already takes first losses; with higher correlation its expected loss actually falls slightly.
Option D is wrong because tranche values depend on the loss distribution, not just expected loss. Option C is wrong because the effect differs by tranche.

Answer: B. The senior tranche is hurt most relative to the model, because higher correlation makes large joint losses more likely.

Example 2

A bank reports a 1-day 99% VaR of USD 4 million on a trading portfolio. Assuming iid returns with zero mean, compute the 10-day 99% VaR. Then state one reason this figure may still understate risk in a stressed market.

Show the solution

Use the scaling rule: VaR(10 days) = VaR(1 day) × √10.
√10 = 3.1623 (approximately).
VaR(10 days) = 4 × 3.1623 = USD 12.649 million, about USD 12.65 million.
Reason for understatement: the rule assumes iid returns and that positions can be held or exited over the period. In stress, returns can be serially correlated, volatility rises, and illiquid positions cannot be closed in 10 days.
Also, VaR gives no information on losses beyond the 99% threshold.

Answer: 10-day 99% VaR is about USD 12.65 million. It may still understate risk because of fat tails, rising volatility, illiquidity and the iid assumption.

Exam tips

Expect questions that ask for the cause of a failure. Match the assumption to the failure rather than recalling the story.
Learn the direction of the correlation effect on equity versus senior tranches. It is a frequent test.
For VaR questions, remember the limits: no information beyond the cutoff, dependence on history, and a liquidity assumption.
For London Whale, think governance: model change, validation, limit breaches, valuation of illiquid positions.
Do the √h arithmetic carefully. Use a calculator's square root key and keep four decimals.

Practice questions from Learning From Financial Disasters

Model Risk and Valuation Failures: frequently asked questions

What is model risk in FRM Part I?

Model risk is the loss from errors in a model's design, data, implementation or use. It includes using a sound model outside its valid range. FRM questions often ask which assumption or control failed.

Why did the Gaussian copula fail in the 2008 crisis?

It summarised default dependence with a single correlation that was assumed stable and calibrated on calm periods. It gave little weight to joint extreme defaults. Firms also relied on it and on ratings without enough stress testing.

What were the main model risk lessons of the London Whale?

A new VaR model was implemented with errors and lowered reported risk while limits were under pressure. Validation, oversight and valuation controls were weak. The lesson is that models need independent review and change control.

Does VaR fail as a risk measure?

VaR is useful but limited. It ignores losses beyond the cutoff, depends on past data and assumes positions can be liquidated. It should be supplemented by expected shortfall and stress tests.