The Fifth Risk(33)

Here was yet another illustration of chaos in life: even slight changes in our ability to predict the weather might have fantastic ripple effects. The weather itself was chaotic. Some slight change in the conditions somewhere on the planet could lead to huge effects elsewhere. The academic meteorologists around DJ knew this; the question was what to do about it. The Department of Meteorology at the University of Maryland, as it happened, had led a new movement in forecasting and spurred the National Weather Service to change its approach to its own models. Before December 1992 the meteorologists had simply plugged the data they had into their forecasting model: wind speeds, barometric pressure, ocean temperatures, and so on. But most of the planet’s weather went unobserved: there was no hard data. As a result, many of the model’s inputs were just estimates—you didn’t actually know the wind speed or barometric pressure or humidity or anything else at every spot on the planet.

An idea pursued at Maryland and a couple of other places was to run the weather model over and over, with different initial weather conditions. Alter the conditions slightly, in reasonable ways. Vary the wind speed, or barometric pressure at 10,000 feet, or the ocean temperature, or whatever seemed reasonable to vary. (How you did this was its own art.) Do it twenty times and you wind up with twenty different forecasts. A range of forecasts generated a truer prediction of the weather than a single forecast, because it captured the uncertainty of each one. Instead of saying, “Here’s where the hurricane is going,” or “We have no idea where the hurricane is going,” you could say, “We don’t know for sure where the hurricane might go, but we have a cone of probability you can use to make your decisions.”

“Ensemble forecasting,” the new technique was called. It implied that every weather forecast—and not just hurricanes—should include a cone of uncertainty. (Why they don’t is a great question.) “There’s a storm coming on Saturday” means one thing if all the forecasts in the ensemble say the storm is coming. It means another if some of the forecasts say there is no chance of rain on Saturday and others say that a storm is all but certain. Really, the weather predictions should reflect this uncertainty. “Why is the newspaper always giving us a five-day forecast?” asked DJ. “It should be a two-day forecast sometimes. And it should be a fourteen-day forecast other times.”

By the time DJ discovered the security hole in the government’s database, the National Weather Service had taken to ensemble forecasting and was generating a dozen or more forecasts for each day. On some days the forecasts would be largely in agreement: slight changes in the estimates of current weather conditions did not lead to big changes in the future weather. At other times they varied radically. That is, sometimes the weather was highly chaotic and sometimes not. DJ quickly saw that instability was not in any way linked to severity: a Category 5 hurricane might keep on being a Cat 5 hurricane without a whole lot of doubt. Then, other times it wouldn’t. “Why in the case of one storm are the forecasts all the same, and in the case of another they are all different?” he asked. Why was the weather sometimes highly predictable and other times less so? Or as DJ put it, “Why does a butterfly flapping its wings in Brazil cause or not cause a tornado in Oklahoma?”

With the government’s data he was able to contribute a new idea: that the predictability of the weather might itself be quantified. “We all know the weather is chaotic,” he said. “The question is: how chaotic. You should be able to assess when a forecast is likely to go seriously bad, versus when the weather is stable.” In the end his thesis created a new statistic: how predictable the weather was at any given moment.

When he defended his thesis, in the summer of 2001, he was surprised by what the U.S. government’s data had enabled him to do. “As a grad student you’re just like, I hope I have something that doesn’t suck. You don’t actually expect your stuff to work.” He wasn’t a meteorologist. Yet he’d found new ways to describe the weather. He’d also found, in himself, a more general interest: in data. What else might it be used to discover?

The relevance of that ambition became a bit clearer after the terrorist attacks of September 11, 2001. “There was a sense that this was, among other things, a failure of data analysis,” he said. “If we had known how to distinguish signal from noise we’d have seen it and prevented it.‘Hey, why are all these guys suddenly taking flight lessons?’” The assassins’ use of credit cards alone, properly analyzed, would have revealed they were up to no good. “The image of a good network is messy,” said DJ. “It’s really hard to fake messiness. It’s hard to fake being an American with a credit card.”

The big question now in DJ’s world was: How, using data, do you identify threats to U.S. interests? By this time a young postdoc at Maryland, he attended a talk by a guy who ran something called the Defense Threat Reduction Agency. The agency, inside the U.S. Department of Defense, was charged with defending the country against weapons of mass destruction. It was trying to understand terrorist networks so it might disrupt them. “I hear the talk, and I was like, Wait a second,” DJ recalled. “The idea that if you push a network a certain way it might collapse. Is the network stable or unstable? It’s a lot like the question I was asking about weather forecasts.” A terrorist network, like a thunderstorm, might be chaotic. Terrorist networks, along with a lot of other security matters, might be better understood through chaos theory. “If you pull out a node in a terrorist cell, does it collapse? Or the opposite: How do we design our electricity grid so that if you take out a node it does NOT collapse?”