On an October morning in Tokyo, meteorologist Haruka Arai noticed something strange.
The official forecast on her phone said there was only a 30 percent chance of rain.
Yet, as she walked from the station to her research institute, she counted umbrellas everywhere.
One person carried a transparent convenience-store umbrella. Another wore a compact rain jacket. A third had tucked an umbrella into the side pocket of a backpack. At the crosswalk, she counted twenty pedestrians.
Fourteen had some form of rain protection.
Haruka stopped.
“Thirty percent?” she muttered.
Her colleague Sato, walking beside her, laughed.
“You’re trusting the crowd more than the forecast?”
“Not exactly.”
She looked down the street.
“I’m treating the crowd as another sensor.”
That idea had become increasingly interesting to meteorologists and data scientists in the age of ubiquitous smartphones, weather apps, radar networks, and machine learning.
A modern weather forecast is not produced by one person looking at the sky. Numerical weather prediction models combine observations from satellites, radar, weather stations, aircraft, ocean buoys, and other instruments. Those observations are fed into computational models that estimate the evolving state of the atmosphere.
But there was another source of information that was rarely described as a weather instrument:
human behavior.
Every person carrying an umbrella had, consciously or unconsciously, performed a tiny calculation.
They might have checked a weather app before leaving home.
They might have watched the television forecast.
They might have noticed dark clouds.
They might have remembered that rain was predicted in the afternoon.
Some might simply have seen other people carrying umbrellas.
Each decision was imperfect.
But that was precisely what interested Haruka.
“One person’s umbrella tells us almost nothing,” she explained during a seminar later that afternoon. “Ten thousand people’s umbrellas tell us something much more interesting.”
She called it the Umbrella Index.
Suppose 70 percent of commuters entering a station between 7:30 and 9:00 were carrying rain gear.
That did not prove that rain would occur.
People could be unnecessarily cautious. Yesterday’s rain might have influenced them. A weather app could have been wrong. Some commuters might carry umbrellas every day.
But the percentage contained information.
The crowd had effectively performed thousands of independent—or partially independent—judgments about the probability of rain.
This resembles a principle known in statistics as the wisdom of crowds: under appropriate conditions, aggregating many individual estimates can produce a surprisingly accurate collective estimate.
There was, however, an important catch.
The crowd could also be wrong.
If everyone looked at the same erroneous weather app, their decisions would no longer be independent. Instead of thousands of separate signals, Haruka would effectively be observing the same signal thousands of times.
“That,” she told Sato, “is the dangerous part.”
The crowd was not automatically intelligent.
It became informative when its members possessed somewhat different information and made decisions independently.
The following week, Haruka encountered the opposite situation.
Rain began falling during the morning commute.
She expected umbrellas to bloom across the station plaza.
Instead, most people kept walking without opening theirs.
Five minutes passed.
The rain became heavier.
Still, only a few umbrellas appeared.
Sato looked at her.
“Your crowd sensor is broken.”
“Maybe.”
Haruka checked the radar.
A narrow band of precipitation was moving rapidly eastward.
Then, almost as quickly as it had arrived, the rain weakened.
People who had decided not to open their umbrellas had unknowingly made the correct prediction.
The crowd had detected something before the instruments became intuitively meaningful to a human observer:
the rain was temporary.
That evening, Haruka wrote a sentence in her notebook:
A crowd is an analog form of big data.
Unlike digital big data, it did not arrive as rows in a database.
It arrived as gestures.
Umbrellas opened.
Umbrellas remained closed.
Raincoats appeared.
People ran.
People walked slowly.
People stopped looking at the sky and continued toward the station.
Each behavior was a tiny data point.
Modern artificial intelligence was teaching machines to recognize patterns in enormous digital datasets. But humans had been generating another kind of dataset for centuries simply by living their lives.
A marketplace revealed demand through what people bought.
A train platform revealed delays through where people gathered.
A restaurant revealed popular dishes through what disappeared from tables.
A city revealed approaching rain through thousands of umbrellas.
The crucial insight was not that the crowd was always right.
It was that the crowd contained information.
The smarter question was therefore not:
“Should I trust the crowd?”
It was:
“What information caused the crowd to behave this way—and how independent are those judgments?”
Haruka closed her notebook.
Outside the window, Tokyo was turning silver under another brief shower.
On the street below, umbrellas began opening one after another.
She smiled.
The official forecast had predicted rain.
The crowd had simply arrived at the same conclusion—without ever speaking to one another.
All names of people and organizations appearing in this story are pseudonyms
Comments