the second hundred · metaphor 132
How do you budget for a thing that has never happened yet — the flood that tops every gauge, the grief heavier than any you have carried? The worst case is not just a bad version of the ordinary day. It is its own kind of event, with its own shape.
Most planning quietly assumes the future looks like the past, only more so — a hard year is a normal year with the dial turned up. So we prepare for the average, pad it a little, and call it caution. But the events that break us are rarely the average turned up. They are the maximum of a long run of tries: the highest river in a century of springs, the deepest loss in a lifetime of ordinary losses, the one customer surge that outruns every capacity plan.
And the strange, useful fact is that the maximum has a life of its own. Collect the biggest value from each long stretch and those record-holders do not scatter like ordinary days; they pile into a lopsided curve all their own, sitting far out past the crowd, and they keep drifting outward the longer each stretch runs. Learn that curve and you can budget for a flood you have never seen.
The idea
Take a run of ordinary draws — rainfall days, quarterly losses, request spikes — and keep only the single largest from each long block. Those block maxima have their own distribution, and here is the theorem that makes planning possible: almost regardless of what the ordinary day looks like, once the blocks are long enough that distribution collapses onto one of just three shapes, the generalized extreme value family. This is the Fisher–Tippett–Gnedenko theorem: the extreme-value analogue of the bell curve.
Which of the three you get is decided by the tail of the ordinary day — the part almost no single draw ever reaches. A light tail (heights, most weather) gives the Gumbel shape. A heavy, power-law tail (wildfires, market crashes, city sizes) gives the Fréchet shape, whose records have no ceiling and grow ferociously. A hard upper bound (a percentage, a finite basin) gives the Weibull shape, whose records crowd up against a wall. Same theorem, three temperaments of catastrophe.
Two things never stop being true. The record curve sits far to the right of the crowd of ordinary days — the worst case is a genuinely different event, not a bad Tuesday. And it keeps moving: double the length of each block and the typical record climbs further out. The 500-year flood is not the 100-year flood plus a margin; it is drawn from a curve that has already walked downstream.
What to try
The instrument draws real blocks from the ordinary distribution you pick, keeps each block's maximum, and heaps the maxima into the live histogram — nothing scripted, every bar counted. Press Draw one block to watch a single season of draws splash across the top strip with its record flagged in red. Then let it run: the maxima pour into the teal histogram, and the violet extreme-value curve — fit live to the records you have collected — settles onto their skewed shape.
Now drag block size. As each block lengthens, the whole record-curve slides right and its left flank steepens: bigger blocks make bigger, more predictable records. Switch the ordinary day between a light tail, a heavy power-law tail, and a bounded one, and watch the limit shape change character — the heavy tail flings its records far out with a long right smear; the bounded one jams them against a ceiling. The crowd of ordinary days barely moves. Only the records travel.
The mapping
A prudent life plans on two clocks at once. There is the ordinary distribution — the spread of normal days you live inside, and mostly plan around. And there is the distribution of maxima — the record flood, the worst diagnosis, the peak grief — which is not on the same curve at all. Confusing the two is the classic planning error: sizing the levee, the savings buffer, or the emotional reserve for a bad-but-normal year, when the thing that arrives is the maximum of many years.
Extreme value theory says: give the worst case its own distribution, and let three facts govern the budget. It lives far out past the crowd, so plan for it as a different event, not a scaled-up ordinary one. Its size depends on your exposure window — how many draws you are exposed to — so a longer life, a wider portfolio, a bigger city faces a heavier record. And the tail you live under sets the character: light-tailed risks have records you can nearly bound; heavy-tailed ones have records that keep outrunning your last plan. The record flood really does have its own distribution — and you can budget for it before you have ever seen it.
Read as life lessons
The record flood is not a rainy day with the volume up; it is drawn from a different curve, sitting far past the ordinary. Plan for it as a separate kind of thing, with its own budget line.
Records grow with the number of draws you face. A longer life, a wider net, a bigger system meets a heavier maximum. The block size — your window of exposure — is half the answer.
A light tail gives records you can nearly cap; a heavy tail gives records that keep escaping your last estimate. Know which tail you live under before you trust any single number.
In the wild
Hydrologists fit extreme-value laws to annual peak river levels to size levees and spillways for the 100- or 500-year flood — a maximum, by construction, no single year has shown.
Reinsurers and banks model the tail of losses — the once-a-generation quake or crash — with GEV and its cousin, the generalized Pareto, because the mean loss tells them nothing about ruin.
Wind loads on bridges, gust limits on aircraft, thermal ceilings on chips: safety cases are written against the extreme a structure will meet over its life, not the load it feels on an average day.
The mapping, exactly
| Mathematics | Life |
|---|---|
| the parent distribution | The spread of ordinary days — the weather, moods, losses you live inside and mostly plan around. |
| a block of m draws | A window of exposure — a century of springs, a lifetime of losses, a season of demand. |
| the block maximum | The record for that window — the worst flood, the heaviest grief, the peak surge. |
| distribution of maxima (GEV) | The worst case as its own kind of event, with its own shape — far from the crowd of ordinary days. |
| the shape parameter ξ | The temperament of catastrophe — bounded, light-tailed, or ferociously heavy-tailed. |
| the rightward drift with m | Longer exposure meets a bigger record — the 500-year flood already walked past the 100-year one. |
The honest model
There is no scripting and no faked curve. You choose a parent distribution — a light-tailed Gaussian, a heavier exponential, a heavy power-law (Pareto), or a bounded uniform. The instrument samples a block of m real draws from it, keeps the largest, and drops that record into a live-counted histogram. Every bar is a tally; the mean record and its spread are computed from the collected maxima.
The violet overlay is the generalized extreme value density, GEV(μ, σ, ξ), with the shape ξ fixed to the value the chosen tail is known to converge to (ξ=0 Gumbel for Gaussian and exponential, ξ=+1/α Fréchet for Pareto(α), ξ=−1 Weibull for uniform), and its location μ and scale σ matched by moments to the records you have gathered. It is a fit to your live sample, not a drawn cartoon — which is why it wobbles when few records are in, and locks onto the histogram as the count climbs. Widen the block and both the histogram and its fit slide right together, exactly as Fisher–Tippett–Gnedenko predicts.
Where the metaphor tears
Estimating an extreme-value law takes a long record of independent maxima — decades of annual floods, thousands of simulated years. A person meets the true catastrophe once, without replicates and without a second draw. The metaphor sharpens how you think about the worst case — its own event, exposure-driven, tail-shaped — long before it hands you an honest number for your own life.
The clean limit assumes draws are independent and the world isn't shifting underneath them. Real extremes cluster — floods follow floods, crashes breed crashes — and the climate, the market, the body are non-stationary, so yesterday's tail is not tomorrow's. The record curve you fit can itself be walking, and a stationary fit will quietly under-budget.
When ξ is large and positive — the Fréchet world of true black swans — the mean or even the variance of the maximum can be infinite, and any single "expected worst case" is a mirage that the next record dwarfs. Here the theory's honest message is a warning, not a plan: no finite budget bounds it, and prudence means capacity and optionality, not a point estimate.