Donald Trump’s win surprised many around the world, but none more than the modelers and big league prognosticators who were calling for a likely Clinton victory. That outcome doesn’t mean data-driven forecasting died on Tuesday. In fact, the best performance went to an artificial intelligence able to crunch more data than its human rivals.
The takeaway? Forecasters need new ways to talk about the uncertainty of their models and need to expand beyond “polling data.” The Defense Department and the intelligence community, who have also grown fond of machine-aided prediction, would do well to heed that lesson.
There’s an important and overlooked distinction between a prediction (which suggests certainty) and a forecast, which acknowledges more than one possible outcome and then weighs multiple outcomes in terms of their relative probability. But humans, and particularly media types who report on polling information, like certainty. So forecasts are cast as predictions.
Nate Silver, the famous prognosticator, forecast Hillary Clinton’s win as likely, giving it an about 71 percent probability. Importantly, that does not mean he was technically wrong. It just means the less likely outcome occurred.
The site Election Analytics, maintained by computer scientist Sheldon Jacobson of the University of Illinois, outputs a number of possible outcomes for elections. In the Clinton-Trump matchup, the site had had 21 possible scenarios. In only one did Donald Trump win.
“We didn’t report it widely,” said Jacobson of the winning scenario. “I woke up on Tuesday and wondered, ‘Should I report the most extreme scenario? Or should I report the average?’ It’s like buying a basket of stocks versus buying an individual stock.” Bottom line, he says, “People expect a level of precision that the data cannot provide.”
Jacobson also does forecasting for Transportation Security Administration, part of the effort to screen for terrorists effectively while still keeping lines moving at airports. He says a key difference between forecasting elections and forecasting for airport security is that the risk of getting an election prediction wrong is small. “If we had 21 scenarios and one came up red” in a national security policy, “We would not implement that security policy. That risk was too great.”
Jacobson’s method, like Silver’s, employs Bayesian statics. That means both models employ the following formula.
P (A | X) = p ( X | A ) p ( A )
p ( X )
P means probability, A is the answer and X is something that will influence the probability. The more information you provide, the more times you run the formula, the greater your confidence level about A. But Bayesian statistics never tell you exactly what will happen, only what might happen based on available data.
That means your output is only going to be as good as the data you’re putting into the model. The pollsters who failed Tuesday all relied on conventional polling information.
“I don’t think the polling data was the problem,” Jacobson said. “It was the uncertainty in the data because of the undecided factor. In a few polls, it was 11 percent. That’s like trying to drive 80 miles an hour in fog. We just had an accident.”
University of Maryland computer scientist V.S. Subrahmanian creates forecasting methods the intelligence community uses to predict the activities of terrorist groups. He said the Clinton-Trump matchup presented three “Black Swan” factors traditional polling data could not capture.
“The first was a candidate with a background totally different from major recent presidential candidates—one who had held no political office, and who spoke his mind freely, unlike the more guarded speech of seasoned politicians,” Subrahmanian said. “The second was the unprecedented disclosures by the FBI during the final few days of the election. A third was the influence campaign run by a foreign state through the drip-by-drip disclosure of internal campaign deliberations. I believe that existing predictive models were unable to account for these three types of variables, leading to predictions based on other historical factors, which were eventually proven wrong.”
Kalev Leetaru used big data to accurately discover the location of Osama Bin Laden in Abbottabad, Pakistan, using not polling information but open-source published news articles. Leetaru chalked up Tuesday’s failure to bad judgment about which data was relevant.
“When it comes to the kinds of questions that intelligence personnel actually want forecasting engines to answer, such as ‘Will Brexit happen’ or ‘Will Trump win,’ those are the cases where the current approaches fail miserably," he said. "It’s not because the data isn’t there. It is. Is because we use our flawed human judgment to decide how to feed that data into our models and therein project our biases into the model’s outcomes.”
Beyond Conventional Polling: The Future of Forecasts
Today, of course, researchers can use more than just phone data to make forecasts, including broad social phenomena. In 2012, Virginia Tech computer scientist Naren Ramakrishnan was able to correctly predict social protest movements in Mexico and Brazil with nothing but Twitter data, part of a research program sponsored by the intelligence community through the Intelligence Advanced Research Projects Activity..
In his own theory about Tuesday’s outcome, Ramakrishnan also cited Black Swan high-impact, low-probability events. Bottom line, many forecasters did, in fact, predict it. They just predicted that outcome alongside many others. To the average person, that works out the same as not predicting the actual outcome.
“There are a lot of best practices that one could use, the important thing is to rely on multiple sources of data, the other [important thing]… quantify uncertainty. You have to ask the question, ‘When will I be wrong?’” he says. Forecasters need to make that self-analysis as public as their forecasts.
Predata, a predictive analytics company established by former members of the intelligence community, also said a Clinton win was more likely than a Trump win. They crunch hundreds of thousands of data points a day from over a thousand websites, such as YouTube.
In a note to subscribers of their newsletter, Predata officials explained three things led them to their conclusion on a Clinton win being more likely based on the complexity of the electoral college system, overweighting polls and “other experts,” and Trump generating a lot of online content and excitement that should have made it into their analysis, but didn’t, for the same reason pollsters and elite liberal newspapers wrote off Trump as a joke until his victory was a reality.
“Trump did not win because he uploaded some ads to his official YouTube channel. Rather, his victory was a function of the support that thrived in the seamier, populist corners of the internet—the corners, that is, we made an early decision to exclude from our curation of the signal sets.”
In other words, it came down to precisely the sort of human judgment error Leetaru highlighted.
It’s thus not surprising the big winner in the polling contest last night was not a human at all but an artificial intelligence named MogIA. Developed by Indian entrepreneur Sanjiv Rai, it takes in 20 million data points from a variety of open public websites such as YouTube, Google, Twitter and Facebook to uncover trends in “user engagement.”
Rai said his exclusive arrangement with CNBC prevents him from talking about how the system works or crunches the data points.
“While most algorithms suffer from programmers/developer’s biases, MoglA aims at learning from her environment, developing her own rules at the policy layer and develop expert systems without discarding any data,” Rai told CNBC reporter Arjun Kharpal.
Big data and computer-aided projection are hardly dead. They're just getting started