Industry: Banking and Financial Services Product: ModelRisk Application: Credit Risk Assessment
A multinational bank's credit-risk model reported an expected loss of $15M on a $500M corporate loan portfolio, and the lending desk priced and reserved against that number. It was a point estimate built the orthodox way — multiply Exposure at Default by Probability of Default by Loss Given Default, plug in historical averages, read off the answer. The number was not wrong on its own terms. It was just answering a question nobody at the capital committee was actually asking.
The question they were asking was how much capital does this book need to survive a bad year — and that answer lives in the tail, not the mean. Re-run the same book as a Monte Carlo simulation, with each input carried as a distribution and defaults allowed to cluster, and the picture changes character entirely: a 99% VaR of $98M, an Expected Shortfall of $131M in the tail beyond it, and a 5.9% chance of losing more than $50M in a single year. The deterministic $15M sat at the bottom of a loss distribution whose dangerous end was more than six times larger — and invisible.
Expected loss is the product of three quantities, and the textbook treats all three as constants:
None of them is a constant. EAD drifts upward as troubled borrowers draw down committed lines on the way to default — a right-skewed, LogNormal shape. PD is a small, uncertain probability that moves with the borrower and the cycle, bounded on [0,1] and best carried as a Beta. LGD is the fraction not recovered after workout, governed by collateral value and market liquidity at exactly the wrong moment — a Triangular(15%, 45%, 85%) that the institution calibrated from its recovery archives. Multiply three point estimates and you get one number. Multiply three distributions — across 200 loans with correlated defaults, over 100,000 trials — and you get the shape that actually drives the capital decision.
The $15M is the mean of a loss distribution, and for a credit book the mean is the least useful summary statistic it has. Losses are bounded below at zero, unbounded above, and lumpy: defaults arrive together because borrowers share a macro factor, so the distribution is heavily right-skewed and the action is all in the upper tail. A mean tells you nothing about how fat that tail is. Capital is not held against the average year — it is held against the 1-in-100 year, and the gap between the two is the entire point of the exercise.
The simulated distribution makes the gap concrete. The mean sits at $15M, but the 95% VaR is $54M, the 99% VaR is $98M, and the Expected Shortfall — the average loss given you are already in the worst 1% of outcomes — is $131M, roughly 8.7 times the expected loss. There is a 5.9% probability of exceeding $50M, more than triple the mean. A model that reports only $15M is not conservative or aggressive; it is silent on the one region of the distribution where capital is destroyed.
Credit losses move with the cycle, so the same engine was run under three macro overlays — Optimistic, Baseline, and Adverse — by scaling PDs, shifting LGDs, and tightening default correlation as common-factor exposure takes over in a downturn.
The cumulative curves show why a single stressed mean is not enough. Moving from Baseline to Adverse, expected loss roughly triples — $15M to $44M — but the 99% VaR more than triples on top of that, from $98M to $299M. Tail risk grows far faster than the average, because rising correlation in stress makes defaults arrive in clusters rather than independently. The Optimistic and Baseline curves climb almost vertically, while the Adverse curve drags far out to the right. That non-linearity is precisely what a deterministic stressed-mean approach cannot capture, and precisely what a capital buffer has to be sized against.
With six candidate assumptions all uncertain, the next question is which ones actually move the loss number — so the validation budget goes where it pays.
The tornado answers it cleanly: LGD is the dominant driver, swinging the loss by ±$4.6M versus ±$3.0M for PD — about half again as much. Sector correlation (±$2.4M) ranks third, ahead of EAD, collateral haircut, and recovery time. The implication is operational: a percentage point of accuracy spent on collateral valuation and workout-recovery modelling buys more reduction in residual uncertainty than the same effort spent refining default probabilities. The bank reallocated analyst time accordingly.
The second source of tail risk is correlation between sectors — the thing a loan-by-loan deterministic model ignores entirely, because it sums losses as if each borrower fails on its own.
The matrix exposes concentrations the EL formula never saw. Retail and Hospitality default together at 0.62; Energy and Manufacturing at 0.42; Real Estate and Hospitality at 0.40. Those pairings are what turn a manageable scatter of independent defaults into the clustered, capital-destroying tail in the loss chart. Rebalancing away from the high-correlation pairs cut the 99% VaR materially without touching expected return — the textbook diversification benefit, but only visible because the simulation respects the correlation structure that a sum-of-means quietly discards.
The institution acted on the distribution, not the mean:
Monte Carlo turns credit risk from "what is the expected loss?" into "what is the distribution of losses, and what is the cheapest way to reshape its tail?" — and on this book the difference between those two questions was the difference between $15M and the $98M the capital had to actually cover.