Stress Testing

Every tool so far — fat-tail models, EVT, copulas — is fitted to the past. That’s their power and their flaw: history is a finite, biased sample of what could happen, and the worst event in your data is rarely the worst event possible. Stress testing is the forward-looking complement. Instead of asking “what does the data say is likely?” it asks “what if this happened — could we survive it?” — and lets you probe scenarios history hasn’t dealt yet. This capstone covers the main flavors (scenario, historical, sensitivity, reverse stress tests), how to fuse stress testing with the EVT and copula machinery you’ve built, and — crucially — the honest limits of all of it.

Why models need a forward-looking complement

Analogy. Training only on past exams is great until the professor writes a brand-new question. Statistical risk models are the past-exam student: fluent in what already happened, mute on the genuinely novel. Stress testing is sitting down and asking yourself the hard new question before the exam.

The core problem. Three structural gaps in any data-fitted model:

1. The sample is finite and biased. Twenty years of data may contain no depression, no hyperinflation, no sovereign default in your home market. “Never happened in our window” is not “can’t happen.”
2. Structural breaks invalidate the past. A new regulatory regime, a currency peg breaking, a market structure change (e.g. the rise of algorithmic trading) can make historical relationships irrelevant overnight.
3. Correlations and tail dependence shift in crises. As the last lesson showed, calm-period dependence understates crisis co-movement — and even a t copula is calibrated to past crises, not the next one.

Stress testing doesn’t fix these by being more accurate — it sidesteps them by being hypothetical. You don’t need the scenario to be probable; you need to know whether it would kill you.

The flavors of stress test

Definition. A stress test evaluates portfolio losses under specified severe-but-plausible scenarios. The main types:

• Historical scenario tests replay an actual past crisis on today’s book: “what would the 2008 crash, the 1987 crash, or the 2020 COVID gap do to our current positions?” Concrete and credible, but limited to crises that happened.
• Hypothetical scenario tests construct a coherent imagined shock: “oil +50%, equities −30%, credit spreads +400bp, USD +15% — simultaneously.” Frees you from history but requires judgment to keep scenarios internally consistent.
• Sensitivity tests shock one factor at a time (“rates +100bp, all else equal”) to map which exposures hurt most. Simple and diagnostic, but ignores the joint moves that define real crises.
• Reverse stress tests invert the question entirely (next section).

A good stress program uses all of them: historical for credibility, hypothetical for imagination, sensitivity for diagnosis, reverse for the existential question.

Reverse stress testing: start from the funeral

Analogy. Ordinary stress testing asks “if a meteor hits, how hurt are we?” Reverse stress testing asks “what size of meteor would kill us — and how close is one?” You begin at the outcome (failure) and work backwards to the cause.

Definition. A reverse stress test starts from a defined catastrophic outcome — insolvency, breaching a capital ratio, a fund gating — and works backwards to identify the scenarios that would produce it. Instead of choosing shocks and reading off losses, you fix the fatal loss and solve for the shocks that get you there.

Why it’s uniquely valuable:

• It removes the anchoring bias of forward tests, where you only test scenarios you already imagined (and thus, implicitly, ones you suspect you’ll survive).
• It surfaces hidden concentrations and nonlinearities — you may discover that a surprisingly mild combination of moves, hitting a hidden concentration, is enough to break you.
• It forces an honest conversation: not “are we safe?” but “what exactly would it take to ruin us, and is that closer than we’re comfortable with?”

Worked example. A fund has $200M capital and wants to know what kills it. Working backwards: its book loses about 0.9% of $2B notional per 1% equity drop, plus convexity from short options. Solving for a $200M loss reveals it takes only an 18% equity drop combined with a volatility spike — well within historical experience (2008, 2020 both exceeded this). The forward tests had cheerfully reported survival of a “severe 15% drop”; the reverse test exposed that 18% with a vol spike is fatal and has happened twice in living memory. Same book, but reverse framing found the cliff edge the forward tests walked right up to and stopped short of.

Fusing stress tests with EVT and copulas

Stress testing and statistical tail models aren’t rivals — they’re complementary, and the strongest risk frameworks weave them together.

• EVT calibrates scenario severity. Instead of picking a shock size by gut, use your fitted GPD to ask “what’s a 1-in-200-year move for this factor?” and stress that. EVT supplies a principled magnitude; the stress test supplies the structure and the portfolio impact.
• Copulas make multi-factor scenarios coherent. A naive hypothetical (“equities −30% and credit +400bp”) needs those moves to be jointly plausible. A tail-dependent copula (a t copula) tells you how strongly factors crash together, so your scenario reflects realistic crisis co-movement rather than an arbitrary or impossibly mild combination.
• Stress tests cover EVT’s blind spot. EVT assumes the tail is “regular” and the past regime persists. Stress testing handles the structural breaks and novel scenarios EVT can’t — the genuinely unprecedented.

The integrated workflow: model each factor’s marginal tail with EVT → tie factors together with a tail-dependent copula → generate severe-but-coherent scenarios → run them through the portfolio → and add reverse stress tests to catch what you didn’t think to imagine. No single tool suffices; the discipline is in the combination.

The honest limitations of stress testing

Expertise means knowing where your tools fail — and stress testing has real weaknesses.

1. Scenario selection is subjective. You can only test scenarios you choose, and humans are bad at imagining the genuinely novel. The next crisis often rhymes with no past one. (Reverse stress testing mitigates but doesn’t eliminate this.)
2. No probabilities. A stress test tells you severity if a scenario occurs, not its likelihood. A terrifying loss under an essentially impossible scenario may matter less than a moderate loss under a likely one — yet stress tests don’t rank by probability. Pairing with EVT-derived likelihoods helps.
3. Static-book assumption. Most stress tests freeze the portfolio, but in a real crisis you’d trade, hedge, or be forced to liquidate (often at fire-sale prices). Dynamic effects, feedback loops, and liquidity spirals are hard to capture and usually make reality worse than the test.
4. Second-round effects are under-modeled. Counterparty defaults, margin calls, funding withdrawal, and contagion cascade in ways a first-order shock-and-revalue rarely captures. The 2008 collapse was largely a second-round funding and counterparty crisis.
5. Gaming and complacency. Tests can be calibrated to scenarios the firm is confident it survives, turning the exercise into theater. A passed stress test is reassurance, not proof.

The mature stance: stress testing is essential and insufficient. It’s the best tool for probing the unprecedented, and it still can’t tell you about the scenario you failed to imagine. Use it alongside EVT, copulas, and humility — never as a substitute for any of them.

That completes the working toolkit of tail risk: you can diagnose fat tails, read a tail index, model extremes with EVT, size the tail with expected shortfall, capture crash clustering with copulas, and probe the unprecedented with stress tests — all while knowing where each tool fails. The final exam pulls the whole structure together; no hints, no take-backs.