Predicting the future, from bones cracked in fire to computers: how far ahead each method can see

Does your brain predict the future?#

Before any oracle, there was the body. Cricket batsmen do not follow the ball with their eyes: they jump their gaze to the spot where it will bounce and wait for it there (Land and McLeod, 2000). Your own touch hardly tickles because, on the leading explanation, the brain has already predicted what your movement will feel like and turned the sensation down (Blakemore, Wolpert and Frith, 1998). A school of neuroscience goes further and describes the whole brain as a prediction machine. The theory is influential and contested: its reviewers say the evidence has not caught up with it.

Further ahead, people imagine. An amnesic patient described by the psychologist Endel Tulving could recall no episode of his life. Asked what went on in his mind when he tried to think about tomorrow, he answered: “Blank, I guess.” Brain scans show that remembering and imagining use much of the same network. Whether the hippocampus is needed to picture the future is still argued: in one study four of five patients with hippocampal damage struggled to imagine new scenes; in another, six patients managed normally. Children of four already choose today what they will need later. Scrub-jays hide food where they will be hungry the next morning, though whether that is the same faculty is disputed.

When you picture your own future, you often see yourself from outside. In seven studies, Emily Pronin and Lee Ross found that people describe their past and future selves the way they describe other people, and picture them more from an observer’s point of view. In four experiments, from drinking an unpleasant liquid for science to tutoring in exam week, people making real decisions decided for their future selves the way they decided for others. Part of what the mind does with the future is run a model of someone who is not quite you yet.

When did people start trying to see the future?#

Divination turns up in so many societies that the anthropologist Donald Brown put it on his list of human universals, next to “dream interpretation” and “future, attempts to predict”. The oldest traces are bones. In northern China, from the fourth millennium BCE, people put the shoulder blades of animals in the fire until they cracked and read an answer in the cracks, thousands of years before anyone wrote the question beside them. Around 1250 BCE the Shang kings at Anyang had the question carved beside the crack, often twice, once each way: “Ban will have misfortune”, “Ban will have no misfortune”. Well over 100,000 fragments survive, by most estimates: the earliest large body of Chinese writing.

In Mesopotamia diviners read the livers of sacrificed sheep. The palace of Mari kept 32 clay models of livers around 1800 BCE, and in one month there between three and seven hundred lambs were slaughtered for divination. When the sky threatened an Assyrian king, a substitute was crowned to take the omen’s sentence while the real king stood aside. Babylonian scholars watched the sky for omens night after night and wrote down what they saw. Their Astronomical Diaries run from 652 BCE into the first century BCE, and, on the best reading of the evidence, that patient watching is how the sky became the first future anyone could calculate.

Greek oracles spoke in riddles. Herodotus tells how Croesus of Lydia asked Delphi whether to attack Persia and was told that he would destroy a great empire; he did, his own. It is a story, written down a century later, and it shows the device: an answer that cannot be wrong. Romans watched birds and sacred chickens. Cicero quotes Cato, who wondered how one soothsayer could look at another without laughing.

If divination cannot see the future, why has it lasted?#

A famous answer is that it works by not seeing. In 1957 Omar Khayyam Moore suggested that the Naskapi of Labrador, who chose where to hunt by reading the cracks in a burnt caribou shoulder blade, were in effect rolling dice, so the herds could not learn their habits. The idea is elegant and often retold. Revisiting it in 2025, the anthropologist Ze Hong concluded that Moore’s version is “probably wrong”, citing critics who found that caribou do not learn hunters’ routines and that the Naskapi turned to the bone mostly in hungry years.

A second answer is about people. Pascal Boyer argues that divination persuades because the diviner is visibly not the author of the verdict: the cracks, the chicken or the cards speak, so nobody’s interests seem to be talking. Among the Azande of Sudan in the late 1920s, E. E. Evans-Pritchard watched the poison oracle: poison was fed to a chicken and a question put to it, and a second chicken had to confirm the answer. When it failed, he counted eight ready explanations, among them the wrong poison, a broken taboo, witchcraft and angry ghosts. “Let the reader consider any argument that would utterly demolish all Zande claims for the power of the oracle,” he wrote. Translated into Zande ways of thinking, “it would serve to support their entire structure of belief.”

Ze Hong and Joseph Henrich call divination a technology people use to get information they believe is accurate, kept alive by strong prior belief, by misses nobody reports, and by seeing others rely on it. Magic and magical thinking seem to flourish where the stakes are high and knowledge thin. Bronisław Malinowski noticed that Trobriand islanders used no magic in the calm lagoon, where fishing was safe, and a great deal of it on the open sea. During the 1991 Gulf War, Israelis in places hit by missiles showed more magical thinking than those in places that were spared. In a 2016 experiment, people’s stress peaked when a shock was hardest to predict.

And when a prophecy fails? For seventy years psychology students learned from When Prophecy Fails (1956) that a small flying-saucer group whose world did not end on 21 December 1954 doubled down and began to preach. In 2025 Thomas Kelly published what Leon Festinger’s newly opened papers show: the leader recanted, the group dissolved and the preaching stopped; one of the researchers, posing as a spiritual authority, had helped bring the climax about. Most later studies of groups whose dates passed, Kelly notes, did not find the famous doubling-down either.

What does your star sign say about you?#

Type your birthday. The reading below is written for your sign. Read it, then say how well it describes you.

Skip the test: show what the tests found

What was the first future people could calculate?#

The sky. Babylonian astronomers, working from centuries of omen-watching, could tell a year ahead in which months an eclipse was possible, and for eclipses they could see, their predicted start times were right to within an hour or two. Many predicted eclipses were “watched for, but not seen”: they forecast possibilities, not certainties. Herodotus says Thales foretold the year of an eclipse, now dated to 585 BCE; historians of astronomy doubt that any method of his time could have done it. In 1715 Edmond Halley printed a map of a coming eclipse over England; in London it began within a minute of his time.

In 1066 Halley’s comet was an omen: the Bayeux Tapestry shows men pointing at it before the Norman conquest. In 1705 Edmond Halley worked out that the comets of 1531, 1607 and 1682 were one object and wrote: “I dare venture to foretell, That it will return again in the Year 1758.” It was found on Christmas night, 1758. A refinement of Halley’s prediction that allowed for Jupiter and Saturn missed its closest approach to the Sun by a month; for its 1835 return the error was four days. The next return is in July 2061 (where the comet is on its orbit now).

From 1872, the tide machine of William Thomson, later Lord Kelvin, drew a whole year of tides for a port in about four hours. NASA’s canon of eclipses runs from 2000 BCE to 3000 CE, nearly twelve thousand solar eclipses; the doubt is where on Earth the shadow falls, because the Earth’s spin slows unevenly: by the year 3000 the uncertainty grows to about 8 degrees of longitude. The orbits themselves can be computed for tens of millions of years. Beyond that, the solar system is chaotic: its exact future on that scale cannot be calculated, only its odds.

How far ahead can the weather be forecast?#

On 1 August 1861 Robert FitzRoy, a naval officer who ran the new meteorological department, published the first public weather forecast in The Times, for the next day or two. “Prophecies or predictions they are not,” he wrote in 1863: a forecast is an opinion that results from “scientific combination and calculation”. No measure of how often he was right has been found. The forecasts stopped in 1866, a year after his death, and returned in 1879; the shipping forecast followed in the 1920s. Folk forecasting kept going alongside: Old Moore’s Almanack sold 560,000 copies in 1839, by the historian Bernard Capp’s count.

In 1922 Lewis Fry Richardson tried to calculate the weather by hand, solving the physics for one place over central Europe. His forecast for six hours: a pressure change of 145 hectopascals. The barometer barely moved. In 1950 a team from Princeton ran the first computer forecasts on ENIAC, four re-runs of past days, each forecast taking about a day to compute.

Since then the horizon has moved steadily. The European centre’s forecasts of the large-scale flow stayed above a high bar of accuracy (a correlation of 0.8 between forecast and observed departures from the usual upper-air pattern, not an 80% chance that a rain forecast is right) for about 4.4 days in 1986 and about 7 days in 2024. Forecasts stay useful for about ten days, against seven thirty years earlier: roughly one day gained per decade. An Atlantic hurricane’s position three days ahead is now forecast as well as its position one day ahead was in the 1990s. Machine-learning models joined in 2023–25 and match or beat the physics models on many measures, within the same window.

That window has a limit. In 1963 Edward Lorenz found that tiny differences in a weather model’s starting state grow until the forecasts part company; in 1969 he showed that better observations cannot push that limit out forever. For day-to-day weather in the mid-latitudes, recent experiments put the limit at about two weeks.

Where does foresight stop?#

Seasons can be forecast in outline. Computer models of El Niño, the warming of the Pacific that shifts weather around the world, keep a strong correlation with what happened about eight months ahead, though most of that skill was measured on re-runs of past years, and forecasts that cross the northern spring are much weaker.

Climate models make a different kind of statement: if the world emits this much, it will warm about this much. They are projections, not forecasts. Of 17 projections published between 1970 and 2007, 10 matched the warming that followed; 14 did once differences between the emissions they assumed and those that happened were allowed for (Hausfather and colleagues, 2020).

Some futures have no horizon at all. No one can predict earthquakes, says the US Geological Survey. In 1975 a county government around Haicheng, in China, ordered an evacuation on the day of a large earthquake, after a swarm of foreshocks; the next year, at Tangshan, there was no warning, and 242,769 people died by the official count. Earthquake warning systems today detect a rupture that has already begun and send a few seconds’ notice ahead of the shaking. That is detection, not foresight.

Can experts predict politics and the economy?#

Between 1987 and 2003 Philip Tetlock collected 82,361 probability estimates from 284 experts, on about 27,450 questions. The average expert was only slightly more accurate than random guessing, and slightly less accurate than simple rules that extend the recent trend. Experts with one big idea did worst; those who borrowed from many did modestly better. Tetlock later warned that the popular summary, that experts do no better than dart-throwing chimps, overstates what he found.

Then came tournaments. From 2011, a US intelligence research agency paid teams to forecast events whose outcome was known, on average, within about 100 days. Tetlock’s team won, helped by three plain measures: less than an hour’s training in probability, teamwork, and putting the best forecasters together. Across the four years, these superforecasters looking 300 days ahead were more accurate than regular forecasters looking 100 days ahead. A claim often repeated, that they beat intelligence analysts with classified information by 30%, has no published study behind it; it appears to rest on an unpublished comparison of a team average with the analysts’ own prediction market. In 2024 superforecasters still beat the best AI model on a benchmark of live questions; by July 2026, by the benchmark’s own estimate, several AI systems were statistically level with the superforecasters’ 2024 forecasts.

Election polls have been checked most often: across more than 30,000 polls in 45 countries since 1942, their error falls from about 4 points half a year before the vote to under 2 on its eve, and it has not grown worse. Further out, the record thins. When Tetlock’s team scored 25-year forecasts made in 1988 and 1997, experts showed real skill on slow, rare events such as the spread of nuclear weapons, and no edge over non-experts on changes of borders. A 2018 study for Open Philanthropy found forecasts made ten or more years ahead came true about a third of the time, with no clear decline beyond ten years. Specific human events have been scored with real skill out to about a year; beyond that the evidence is thin and mixed.

Much of the work on long-range futures does not try to predict at all. RAND’s Delphi method (1950s) polls experts anonymously in rounds until their views converge. Herman Kahn and later Shell wrote scenarios, several coherent stories of how things could go. The Limits to Growth (1972) ran a computer model of the world under different assumptions and warned that its graphs “are not exact predictions”. Horizon scanning collects “weak signals” of change; the futures cone sorts what is possible, plausible, probable and preferable; backcasting starts from the future you want and works back to what must be done now. “Foresight is not about predicting or forecasting the future,” the European Commission says; it is “the discipline of exploring different possible futures”.

What did experts in 1964 expect by now?#

In 1964 RAND polled about 150 experts, of whom 82 answered, in six panels and four anonymous rounds, when each of dozens of developments would arrive. For each, the report printed the year by which half the panel expected it. Below are 8 of the 75 forecasts the science, automation and space panels expected by 2020, drawn at random. Has each happened by now, even if later than expected? Cases where only part of it happened are shown separately at the end.

Does science fiction predict the future?#

In January 1902 H. G. Wells gave a lecture at the Royal Institution called “The Discovery of the Future”. “Scientific prophecy will not be fortune-telling, whatever else it may be,” he said, and he hoped for “a systematic exploration of the future”. A year earlier his imagination had refused to see “any sort of submarine doing anything but suffocate its crew”. In 1914 he wrote of “atomic bombs” that kept on exploding.

In October 1945 Arthur C. Clarke worked out in Wireless World that three stations at about 42,000 km from the centre of the Earth, circling once a day, could cover the globe. Communications satellites now sit in that orbit. When researchers scored the non-fiction of Isaac Asimov, Clarke and Robert Heinlein, their technology predictions with a firm date were right 57%, 48% and 36% of the time; their forecasts reached about 40 years ahead.

Many writers say prediction is not their job. Ursula K. Le Guin wrote that science fiction “is not predictive; it is descriptive”: it describes the present in the costume of the future. From 2019 to 2023 the French defence ministry commissioned a team of science-fiction writers, the Red Team, to imagine threats for 2030–2060. In E. M. Forster’s “The Machine Stops” (1909), people talk through screens and are warned: “Beware of first-hand ideas!”

How well can you predict your own life?#

Worse than you think. In 1994 Roger Buehler and colleagues asked 37 students when they would finish their theses. On average they said 33.9 days; it took 55.5. Only 29.7% finished by the date they had called most likely. Observers, asked about other people’s projects, erred the other way. When students had to link their estimate to how long similar work had taken them before, the bias went away.

People also misjudge how they will feel in the future. Assistant professors expected that being denied tenure would leave them much unhappier for years; those who had been denied were happier than predicted, though only seven recently denied professors were in the comparison. Across more than 19,000 people aged 18 to 68, people at every age expected to change less in the next ten years than people ten years older said they had changed in the last ten.

That points to the plainest method: start from how long similar things took before, yours and other people’s.