Worst-Ever Fat-Tail Stress Testing
Stress test to the worst level ever recorded, not the level you find plausible.
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 90%
The mechanism replaces judgement-based scenario setting with a historical extreme. For each risk factor, you take the worst reading ever observed and make that the stress case: markets down 50%, rates at 8%, credit spreads back to their all-time wides. You then run those factors simultaneously rather than singly, because the correlation is precisely what breaks firms. The pass criterion is deliberately weak and deliberately binary: results will be worse, but are you still there and still able to serve clients and build the business? Because the test is cheap once instrumented, cadence matters more than precision: JP Morgan runs on the order of a hundred stress tests a week to cover a wide variety of combinations rather than one blessed scenario.
Origin
Dimon describes finding JP Morgan's high-yield stress test set at a 40% spread widening on arrival and replacing it with the worst-ever level, which the 2008 crisis then exceeded.
Core principles
- 01The worst thing that has ever happened is a reasonable planning assumption, not a tail.
- 02"The market is more sophisticated now" is the sentence that precedes the record being broken.
- 03The goal is not to forecast the shock but to remain operational through it.
- 04A regulator's scenario is a floor for your own testing, never the ceiling.
- 05Run many scenarios often, not one scenario annually.
How to run it
- 1
Inventory the risk factors you actually carry
List every exposure that can move against you (credit spreads, rates, equity levels, funding availability, concentration) before choosing any scenario values.
Watch out Exposures hidden by accounting treatment will not appear in this list unless you look for them deliberately.
- 2
Set each level to the historical worst, not a percentile
For each factor, find the most extreme value ever recorded and use it. Do not soften it because the market structure has changed.
Pro tip Go back further than the working memory of your team; the useful records are usually older than anyone's career.
Watch out Dimon's team argued a 17% high-yield spread could never recur because markets were more sophisticated. It hit 20% in 2008, with no bid.
- 3
Run the factors simultaneously
Combine the extremes into a single scenario rather than testing them one at a time. Fat tails arrive together.
- 4
Score for survival, not for loss size
The output is a yes/no: can you absorb this and continue building the business? Treat a worse P&L as acceptable and an inability to operate as a failure.
Pro tip Add a liquidity dimension; the 2008 lesson was not the mark, it was that there was no market to sell into.
- 5
Test beyond the regulatory scenario
Treat any supervisory stress test as one input among many. Maintain your own scenario library that is strictly harsher.
Watch out Passing the official test tells you what the regulator modelled, not what the world can do.
- 6
Raise the cadence until it is routine
Automate the runs so testing is continuous rather than a periodic exercise, and vary the combinations each cycle to cover the space.
Pro tip Weekly cadence turns stress testing into a management dashboard instead of a compliance artefact.
In the wild
On arriving at JP Morgan and reviewing the risk books, Dimon found the high-yield stress case assumed credit spreads widening 40% from a base of around 400bp (roughly 560bp). He replaced it with the worst level ever recorded, 17%. Internally the objection was that such a level could never recur because the market had become more sophisticated. In 2008 high-yield spreads reached 20% and, as he puts it, there was no market in which a bond could be sold at all.
→ The firm was sized for a shock that the consensus had declared impossible, and it arrived within two years.
Rather than maintaining one canonical downside case, JP Morgan runs on the order of a hundred stress tests weekly across combinations of market level, rates, spreads and other factors, specifically to confirm it can handle a wide variety of outcomes rather than one modelled one.
→ Tail coverage becomes a continuous operating metric instead of an annual set-piece.
Common mistakes
Letting sophistication justify a softer test
The claim that market structure has improved is the standard argument for lowering a stress level, and it has repeatedly preceded the record being broken.
Stopping at the regulator's scenario
Supervisory tests are designed for comparability across firms, not for your specific concentrations. Treating them as sufficient outsources your survival boundary.
Testing marks but not liquidity
A scenario that assumes you can sell at the stressed price misses the actual 2008 failure mode, where no bid existed at any level.
Is it for you?
Best for
Anyone carrying leveraged or correlated exposures who needs to know their survival boundary rather than their expected loss.
Not ideal for
Businesses with no balance-sheet exposure, where the binding constraint is demand rather than solvency.
From the transcript
“No, our stress test is going to be worst ever.”
“And not the stress test the Fed gives us, but all the fat tails. Markets down 50%, interest rates up to 8%, credit spreads back…”
“And the point isn't that you're trying to guess The point is you you can handle them, so you continue build your business.”
From the episode
The Jamie Dimon Interview