What volatility does to you depends on which way your curve bends
2026-08-26 · Fragility, resilience and early warning
"Gets stronger under stress" sounds like a motivational slogan. It is in fact a geometric property you can compute — and once you compute it, most of the things marketed as antifragile turn out not to be.
A person can drown in a river of average depth 1.2 metres. That is not a witticism; it is the seed of this whole issue. Whenever the relationship between an outcome and a variable is curved, any conclusion computed from the average is wrong, and wrong in a fixed direction — either systematically too flattering or systematically too harsh.
The counterintuitive part comes next. If a curve can bend either way, then there must exist things whose situation improves as volatility rises. Not "survives" volatility — profits from it. Such structures exist, and there is nothing mystical about them: you can draw one, measure one, and build one.
Three issues in this phase do three different jobs, and blur together if read carelessly. Topic 37 asked what the distribution of the world looks like (how heavy the tail is). Topic 38 asked whether you return to normal after a hit (resilience). This issue asks one thing only: whether you can end up better than before — and how that claim can be checked. The answer sits in something concrete: the curvature of your payoff curve.
Start with a curve you travel along every day.
Say your commute takes 20 minutes. One morning there is 20% more traffic than usual, and the extra time you lose is rarely 20% — it can be ten minutes, or twenty. Now take 20% less traffic. You save two or three minutes at most, because the road was never going to let you exceed the speed limit.
Traffic engineering has a standard way of putting this: delay grows far faster than volume, and the empirical formula highways agencies use has that growth at roughly a fourth power. In plain terms: the congested side takes away much more time than the empty side gives back.
Draw it. The horizontal axis is how far traffic deviates from normal; the vertical axis is your outcome (higher is better — say, how unhurried your arrival is). The curve bends downward: what the good news pays you cannot keep up with what the bad news takes. Bending down is called concave; bending up is convex.
One warning so the rest reads cleanly: change the vertical axis and the curvature flips. "Travel time as a function of traffic" is convex (time rises faster than volume); "your outcome as a function of traffic" is therefore concave. Throughout this issue the vertical axis is always your outcome, higher is better.
Why does the bend matter so much? Because it settles a rule almost everybody breaks: when the relationship is curved, the outcome computed from the average is not the average of the outcomes. The rule is Jensen's inequality, proved by the Danish mathematician Johan Jensen in 1906 → ref · Convexity & Jensen's Inequality.
Under a concave curve, plugging in the average overstates what you will actually get; under a convex one it understates it. That river of average depth 1.2 metres drowns people for exactly this reason: survival is concave in depth — the shallow stretches save you nothing, the deep one kills you — so "average" carries no safety information at all.
Look at the third panel. When the curve bends up, the same two-sided shock leaves the average of the two outcomes above the outcome with no volatility at all. Volatility itself became a source of income. That is the entire content of the word antifragile: not a good attitude, not a thick skin — a curve that bends upward.
The middle panel is last issue's resilience: flat within the range it can take, unaffected by the shock, back to baseline afterwards. It neither loses nor profits. The three differ not by degree but by shape: concave, flat, convex.
Every conclusion you reached by plugging in an average — average schedule, average load, average footfall, average return — deserves one question first: is this relationship curved? If it is concave, that conclusion is not an expected value but an optimistic ceiling, and should be treated as one: don't promise on it, don't size capacity with it, don't compute your safety margin from it.
"That plan feels a bit fragile" — anyone can say it, and saying it conveys nothing. But fragility can be reduced to a number, by a method simple enough to arouse suspicion.
Two steps. First, pick the variable you care about most: footfall, interest rate, schedule, exchange rate, dose, concurrency — anything the outcome is sensitive to. Second, push it 20% up and 20% down, and compute the outcome both times.
Then compare the two magnitudes. A drop of 45 against a gain of 12: if the downside magnitude clearly exceeds the upside, your curve is concave and you are fragile. The reverse means convex. Roughly equal means you are the flat line.
What makes this test powerful is not its simplicity but that it needs no probabilities. In the 2013 paper that gave the concept its formal definition, Taleb and Douady make this the headline: you do not have to know whether footfall will collapse next year, you do not have to fit a distribution, you do not have to decide between a bell curve and a power law — you only have to compute both sides. And estimating probabilities is precisely the least reliable link in the whole chain of risk work. A test that does not require you to be right about the odds is a test you can actually use. → ref · Monte Carlo & Ensemble Simulation
Follow the same thread and something normally invisible falls out: "same total" often is not the same. If damage is convex in the size of a single dose, then splitting the same total into several doses produces markedly less cumulative damage than delivering it at once. Conversely, if payoff is convex in the size of a single commitment, spreading it thin is worse than concentrating it.
This is not a figure of speech, it is a checklist item. Whenever someone tells you "the total is the same anyway", ask whether the relationship is curved. If it is, how you split it is itself a decision variable that changes the outcome — usually the cheapest one available, since it demands no extra resources at all.
Run a two-sided stress test on your most important system: key variable ±20%, compute both, and record the asymmetry = downside change ÷ upside change. Above one means concave. Then spend your improvement budget on flattening the downside curvature — buffers, caps, removing irreversible steps — rather than on raising the average. On a concave curve, gains in the average get eaten straight back by the downside.
Convex curves do not fall from the sky. The most common way of manufacturing one is to keep an option — a right rather than an obligation.
You pay a fixed, modest amount in exchange for being allowed to act later if you want to, and to walk away if you don't. The loss is capped at what you paid; the gain is not capped at all. Your payoff curve now has a kink in it: floored on the left, open on the right. That is a convex curve, and it was built on purpose.
The straight line on the left is a commitment: you gain as things improve and you keep losing as they worsen, with no floor. The kinked line is the option: past a certain point it stops falling, because you can simply not proceed. On the right the two run almost parallel — the little the option gives up is the premium you paid.
But an option is not free convexity. It rests on four preconditions, and drops if any one fails:
One: the cost per try is small and known in advance. The premium must be a fixed number. If "how much can I lose at most" is only knowable afterwards, this is not an option.
Two: quitting is genuinely cheap. This is the one that gets faked. Being contractually free to exit is not the same as being able to exit: already announced publicly, already staffed with three people, already written into the quarterly goals — those are real costs of abandonment. Once quitting is expensive the option has degraded into a commitment, while you are still approving it under option rules.
Three: you can repeat many times, and the tries do not share a single fatal risk. Convexity pays out through repetition: nineteen small losses and one large win. If the twenty tries are tied to the same rope, that is not twenty options, it is one.
Four: the upside really is heavy-tailed. If the best possible result is only slightly better than average, the upside you bought by capping your downside is not worth the premium. → ref · Identifying and Misidentifying Power Laws
Put the four together and an unintuitive division of labour appears: the value of an option comes mostly not from picking well, but from being able to quit cheaply. Picking well raises your hit rate per try; being able to quit changes the shape of the whole curve. The first is a linear improvement, the second bends concave into convex.
For every "try", write down two things in advance: the abandonment condition (what signal stops it) and the abandonment cost (what you actually forfeit by stopping). If the second is not a small number, stop approving it as a try — it is a commitment and must be judged as one: can you survive the worst case, rather than how big the upside looks. This single rule filters out most "let's just run a small pilot" proposals, whose exit cost was never small from day one.
Push convexity up to the level of allocation and you get an odd-looking conclusion: rather than putting everything into "moderate risk", stand at both ends — most of it extremely safe, a small slice extremely risky, the middle deliberately empty. That structure is the barbell.
What makes it odd is that on average the barbell and the middle-heavy allocation can carry exactly the same risk. If the average is the same, why bother splitting it in two?
Three reasons, each harder than the last.
First, the left tail is truncated. The 90% that is extremely safe guarantees that however the other end blows up, you are still at the table. And still being at the table is the precondition for every convexity: nineteen small losses plus one big win only adds up if you survive to the twentieth. (This is the direct consequence of last issue's ergodicity problem: get wiped out once and every expected value after that is zero.)
Second, the middle is what leans hardest on your estimates. At the extremely safe end you barely need to estimate anything; at the extremely risky end you already assume it can go to zero. Only the middle depends on a parameter you had to estimate to be called "moderate" at all — and in a heavy-tailed world that estimate is the least trustworthy number you own. What the barbell really removes is not risk, it is dependence on your own precision.
Third, the shape of the middle is itself concave. A moderate-risk position usually carries substantial downside and capped upside. That is the definition of concave. You are not compromising; you are buying concave curves in bulk.
One caveat. The barbell is not a licence for "more extreme is better". Its entire legitimacy comes from the safe end — if what you believed was safe turns out not to be (cash-equivalents that fall along with everything else in a crisis), the structure collapses back into a high-risk middle. A barbell is only as strong as its safe end, never as strong as the imagination of its risky end.
Sort your resources — money, time, attention, headcount — into three piles: can lose it all without consequence / middle / absolutely untouchable. Then force yourself to write a justification for the middle pile: why must this stay in the middle instead of being split to the ends? Whatever has no justification gets split. The value of the exercise is less in how much you move than in turning "what I assumed was moderate risk" from a feeling into a position that has to be defended.
Now the other side. "Antifragile" is among the most abused words of the past decade: half of its content is hard, the other half is rhetoric. It is worth drawing the line precisely.
First, convexity is local and bounded by dose. Nothing is antifragile to every stressor at every dose. Toxicology has spent decades on a phenomenon called hormesis: many substances stimulate an organism at low doses and poison it at high ones, so the dose-response curve is an inverted U. Strength training is the same — too little load and muscle atrophies, the right load and it strengthens, too much and you get tears and overtraining.
Note the dashed line in the figure: at the same dose, with recovery cut short, the entire curve drops and the window narrows sharply or vanishes. So calling something antifragile says nothing at all unless you can fill three blanks: antifragile to which stressor, over what range of doses, and at which level.
Second, levels get swapped. The restaurant industry gets better through the relentless failure of individual restaurants — the sector's convexity is bought with the fragility of its members. Likewise, the mechanism that strengthens a muscle involves damage to individual fibres. So "be antifragile" means opposite things at different levels: to a system it means tolerating local failure; to an individual it may mean being the local failure. An antifragility claim that does not name its level is usually the upper level demanding sacrifice from the lower one.
Third, the falsifiability really is weak. This is the one that matters most. "Survived and got stronger" is easily declared antifragile after the fact, while everything that died stays out of the denominator — textbook survivorship bias. The risk researcher Terje Aven, reviewing the concept in 2015, reached a measured verdict: it does add something traditional risk analysis neglects, namely the dynamic side of risk and performance, but it has to be reconciled with existing risk and resilience concepts and given operational definitions before it can enter practice. The operational half — convexity, second derivatives, the two-sided test — comes from option theory, where it already existed and could already be computed. The non-operational half is rhetoric.
Fourth, convexity usually has to be bought. Inspirational readings skip this entirely. Options carry premiums, and in financial markets index put options have historically been expensive on the whole — many buyers of tail protection, few willing sellers, and the resulting premium has been documented repeatedly. So "buy convexity" is not a free law of nature; it is a claim about mispricing: you have to argue that this particular tail is underpriced. Fail to argue it and you are simply paying, continuously.
Before using the word "antifragile", fill three blanks: which stressor, what dose range, which level benefits. If you cannot fill them, drop the word and use something computable instead — convexity, two-sided asymmetry, premium paid, number of redundant copies. This is not pedantry: the three blanks map exactly onto the concept's three standard failure modes (wrong stressor, out of range, swapped level), so filling them in is itself the check.
Yes, but from a narrow set of sources worth listing individually. One is mispricing — others systematically underestimate that tail. The other is structural advantage: for the very same undertaking, your cost of quitting is lower than theirs (a small team can kill a project; a large company often cannot, because commitments, headcount and face are attached to it). The second is the one to watch, because it does not require you to be smarter than the market, only genuinely freer to leave. Conversely, if your abandonment costs are rising — more people, more public commitments — your free convexity is being eaten even though the business has not changed.
Technically they do not conflict; budgetarily they do. Redundancy (resilience) and premiums (convexity) draw on the same money and the same people. A workable split: buy resilience for the irreversible parts and convexity for the reversible ones. The test is not which is more fashionable, but whether there is a next round after this piece breaks.
The middle layer — of assets, of roles, of projects — loses its buyers and thins out, while both ends get crowded: returns at the safe end are pushed to nearly nothing and the risky end gets bid up. Worse, once everyone treats the same thing as the safe harbour, correlation inside that harbour rises — so the strategy manufactures a new tail risk of its own. Any strategy that depends on others not doing it should carry that dependence in its stated assumptions.
An old methodological problem in hormesis research. A rise in the group average has two entirely different sources: the same individuals each got stronger, or the weak ones disappeared and the survivors merely look stronger. There is only one way to separate them — track the same individuals over time rather than comparing group averages at two moments. Every claim of the form "the industry came out of the crisis stronger" has to clear that bar first.
Psychology has a concept called post-traumatic growth, and the popular version runs well ahead of the evidence. The core difficulty is measurement: most instruments ask how you feel you compare to before, which may capture the rebuilding of a narrative rather than a gain in capability, and longitudinal studies typically find effects considerably weaker than self-reports. This does not say growth never happens; it says the evidentiary standard here is nowhere near that of muscle or bone, where the change can be measured directly. Putting the two in the same sentence is this issue's easiest overreach.