How Startups Could Have Predicted Trump’s Stunning Upset

Last Tuesday’s election results took most Americans by surprise, including Donald Trump’s campaign. Most of the poll data available publicly pointed to a Hillary Clinton advantage up to the very end.

Given the stunning upset, which showcased the obvious failures of most pollsters to predict the correct outcome of the election, many are questioning whether polling is dead. An obituary may be premature.

A handful of companies are working on resolving these fatal flaws, using different combinations of human analytics and technology to better anticipate votes. That way, by the time the next election rolls around, the nation won’t be left similarly flat footed.

First, a note about traditional polling: Most polling companies highlight their inherent margin of error, which is typically around three percentage points. After all, these predictions represent a statistical probability. Traditional polling tends to further rely on live or automated landline calls to generate a sample of the voting population and discern the likely voters for both Democrats and Republicans.

This latest election proved strange because so much of the voting public was relatively new to voting. Whereas polls reach out to likely voters, they didn’t gauge the unlikely voters, who, as it turns out, came out in droves to support Trump. Additionally, telephone response rates have fallen to single digits in recent years.

Queue the startups.

“Four years ago we didn’t have the computing power for the type of modeling that we’re doing,” said David M. Rothschild, an economist at Microsoft Research who started PredictWise, a forecasting site that uses big data to model likely outcomes. “It took 24 to 36 hours to pull data to then model the data; now it’s just a matter of minutes,” he told Inc.

Mobile First

Back in September, Rothschild took part in a New York Times experiment which gave the same raw data to different pollsters to predict which presidential candidate would take the state of Florida. The four other participants showed Clinton winning from one to four percentage points.

By contrast, Rothschild’s model gave Trump an edge over Clinton. Unlike traditional polls, which rely heavily on public polling data, PredictWise uses a computer-based model that relies on private polling, as well as public data. For the Times experiment, PredictWise used data gathered from private surveys with both MSN and mobile-first survey company Pollfish. Rothschild’s model also aggregates online, social media and election-betting data to make predictions.

“Public polling has done well in the past, but the bottom line is that with 100 publicly available polls interviewing 1,000 people each, all you get to see is a collection of a few data points,” Rothschild explains. He added that polling might be made further vulnerable, should an error pop up in even a single region. One wrong prediction could lead to errors in other states.

“Ultimately the existing frontier is in individual [state] level data and spending less time aggregating other people’s top line,” Rothschild said. “You’re probably better off combining as much data as you can.”

Additionally, because of the reliance on landline calling–a practice exacerbated by regulations limiting automated calls to cell phones–much of public polling misses the nearly half of the American households that only uses mobile phones.

“It’s wasteful not to take advantage of mobile, Rothschild says, “and it’s one of the things we want to explore further.” He adds that mobile-first polls may be reaching a more diverse and wider audience than landline polls. “In four or eight years the vast majority of people will be mobile first.”

Social Media

On the heels of having accurately predicted the victory of the “leave” vote in Britain, forecasting company Predata entered this U.S. election on a high note. The company uses social media to show directionally which candidate’s digital campaign is inviting more engagement. Although the company, which was founded by ex-CIA officer James Shinn in 2015, was ultimately wrong about which candidate would seal the deal, it did accurately indicate a lead for Trump in 12 of the main battleground states. The overall forecast was wrong–in part because of the human error that goes into discerning which information is the relevant to input into a model.

“The U.S. presidential election is more complicated than the simple ‘Remain’ and ‘Leave’ vote,” Aaron Timms, Predata’s director of content and digital strategy, told Inc. “The electoral college introduces a massive element of complexity.” Nonetheless, he highlighted the benefits of having digital engagement data for future U.S. elections on the company’s blog: “people may have stopped talking to pollsters, but they are talking online–?to their political leaders and to each other?–with increasing frequency.”

Other Factors

Timms’ point is strengthened by the success of MogIA, an artificial intelligence and deep learning model created by Indian entrepreneur Sanjiv Rai that accurately predicted a Trump win. Similar to Predata, MogIA looks at online engagement in relation to each party’s campaign message but without any sort of human curation. Its method corrects for human bias in data selection but is unable to distinguish between negative and positive online engagement. Timms cautions that there’s no indication that the MogIA model will work in all situations since it can’t discern if the data captured is really representative of the population.

Moreover, some human-led approaches also successfully predicted the results. The USC/LA Times poll, although its methodology is questioned by some, consistently (and correctly) placed Trump in the lead. Without using technology, American University professor Allan Lichtman correctly predicted a Donald Trump win since October. He uses a system he calls the 5 keys to the White House.

So perhaps the future of election polling lies in a combination of human analytics and artificial intelligence. And although the right balance might provide more accurate predictions, human error–although mitigated–will still be present.