Cobra Effect · Weather and climate

The rounded number that changed the forecast

How a rounded number revealed why the weather can't be forecast far ahead.

7 cards, read aloud in 2:59, with a test and sources.

A large old computer cabinet with a long strip of printout paper spilling from it

In 1961, Edward Lorenz was running a simple model of the weather on a computer.

Lorenz was a meteorologist at the Massachusetts Institute of Technology. His model boiled the weather down to twelve numbers, such as temperature and wind speed. The computer printed out how those numbers changed, step by simulated step. One day, he wanted to look again at part of a run.

To save time, he restarted the run halfway, typing in numbers from a printout.

The printout showed three decimal places, though the computer kept six. So instead of 0.506127, he typed 0.506. Then he went to get a cup of coffee while the machine ran.

When he came back, the new weather looked nothing like the old.

At first the two runs matched closely. Then they drifted apart, and after about two months of simulated weather they were completely different. The tiny rounding had grown, step by step, into a different forecast. The computer was working perfectly.

Lorenz realised the real atmosphere might behave the same way.

In a system like the weather, tiny differences in the starting state can grow until they swamp everything else. Nobody can measure the whole atmosphere perfectly. So there must be a limit to how far ahead the weather can be forecast in detail. He published the idea in 1963, in a paper titled Deterministic Nonperiodic Flow.

In 1972, he gave a talk with a famous question for a title.

It asked whether the flap of a butterfly’s wings in Brazil could set off a tornado in Texas. The title was suggested by another meteorologist, Philip Merilees. Lorenz had earlier used a seagull as his example. His point was not that butterflies cause tornadoes, but that tiny changes make long forecasts uncertain.

Forecasters now run many forecasts at once, each starting slightly differently.

When the runs agree, forecasters can be confident. When they spread apart, less so. Detailed forecasts are now useful for about a week or more ahead. Beyond about two weeks, the limit Lorenz described still holds as a rough guide.

So when a long forecast looks precise, ask how much a tiny error would change it.

Some systems are steady, and small errors stay small. Others, like the weather, magnify them. In those, the honest forecast is a range of outcomes, not a single answer. Knowing the limits of prediction is a discovery in its own right.

Sources

Nearby ideas

All stories

Cobra EffectIdeas worth knowing
Streak 0 daysRead 0 of 0Known 0Kept 0
Every story is written to a rule: framing can be invented, facts cannot, and a tale that is probably a legend says so. Narration and pictures are generated once and kept. Your progress is saved on this device and in a record of its own on the server, so you can pick it up on another one. Nothing else about you is collected, and there is a button to delete the lot. Privacy.