In December 2015, Bing quietly rolled out an Election 2016 experience, with data visualizations powered by Bing Predicts. You’ll know Bing Predicts as the crew that’s responsible for tearing it up at the Oscars, the NBA playoffs, and the Scottish independence referendum (among others) with strikingly accurate predictions.
This time, with Election 2016, things are different. Not content to merely predict the outcomes of primaries across the country (with an accuracy of 84% as of June 14), the team is also bringing analytical tools to people like you and me, who have questions about the election. We want to know where the candidates stand on the issues. We might be curious about how an issue sits with a specific state. Or we wonder what gender and age group in Nevada is searching online for Donald Trump.
To get this tight, keyhole view into the true sentiment of the United States that’s happening right now, Bing is pulling in three signals: publicly available social data, Bing searcher data and opted-in Internet Explorer user data. (It goes without saying that all this data is anonymous – but we’ll say it anyway.) The point in time where these signals converge produces a sharp and undeniable picture of candidate and public sentiment on issues critical to this election cycle, such as LGBT rights, environmental policy and health care. The view you get is a little bit breathtaking. It’s a little bit staggering.
READING ALL OF THE SENTIMENTS
Nowcasting is the most appropriate word for what Bing is doing, combining people signals with current events to predict an outcome and to report on sentiment. Search engines have become the entry point for almost every piece of information we seek (thereby providing the machine with more data, which leads to more accuracy), and the best part is that our searches, our social commentary and our anonymous online behavior is not biased as polls often are. Which means Bing’s predictive analytics are more accurate to the sentiment of right this minute.
You might be wondering about the gray areas of sentiment. Bing’s model does gray. The model is trained to look at the entire context of a key phrase, as well as the words used to describe the issue, which is how sentiment is learned. For example, “gun control” and “gun rights” are the same issue, but word choice communicates clear sentiment on the issue. If a searcher types “Bernie Sanders” (broad and indeterminate in terms of intent) and then clicks on a news article titled “If Bernie Sanders Is Real, He Will Run as an Independent,” the Bing Predicts model can infer sentiment from that choice. The model can also take into consideration the grayest areas of sentiment, such as if a person favors 2nd Amendment rights, but doesn’t feel people should have access to assault weapons.
WHEN SIGNALS GET CROSSED
Nowcasting can’t catch everybody. It can only read Internet-engaged people; the models only read the signals that are being sent, and not everybody sends signals. And some people send more signals than others. This leaves certain populations who are not sharing their thoughts online under-represented in Bing’s model. This includes poor people without access to Internet or screens and older people without interest in Internet or screens. As of November 2015, North America had 88% of its population online. A Pew Internet study conducted in 2013 (admittedly eons ago in online time) showed that offline adults are predominantly white women with no high school diploma, age 65+ and living on less than $30,000 per year. With political nowcasting this matters more; with marketing nowcasting (more on this in a minute), it matters less because the online sentiment is being sent by the people who are also doing the online buying.
Nowcasting can also be hobbled by information volume – not enough, or too much. The model for both sentiment and prediction can only work with the information it has. In the case of the Bing Predicts team’s first incorrect prediction in the primaries, in Iowa, the problem was not enough information. Because the model didn’t have a baseline or a reference point for Donald Trump’s success (or failure) in elections, it was difficult to measure his potential outcome. The model needed to be trained up; subsequent primaries showed more accuracy.
If Bing Predicts can tell us how people in Pennsylvania feel about social security, what else can this kind of data analysis tell us? Take a minute to imagine the volume of signals around a specific topic – say, adverse reactions to a new diabetes medication – and it’s not hard to imagine the reach and power of a tool like this. It could save lives, by relaying key information to the Centers for Disease Control that activates an alert to physicians who prescribe diabetes medication.
What about a marketing pivot? Already Microsoft is using Bing Predicts to measure sentiment around Xbox and Surface sales forecasting, as an early experiment in the capabilities of this program. Predicts can look at the signals around Xbox and nowcast when a bug has cropped up in the system that needs immediate attention. Support forums, Tweets, queries – amassing information at the volume of an entire country reveals a lot, instantly. This lets the Xbox and Surface teams stay ahead of problems – which is the holy grail of brand marketing.
And then there’s this: through the mountains of anonymized data, Bing Predicts has noted that people who search for information about luxury hotels tend to favor older musical artists, like Paul Simon and James Taylor. Can you imagine other connections like this? What if Bing can see that people who are talking about “what to do after an accident” also visit auto dealer websites? Or that people who live in the central states start talking about summer clothes in April, while people in the Northeast don’t show an interest until July? That people who search for Disney vacations also talk about romantic getaways? If Netflix can recommend movies based on what you’ve watched previously, can Bing recommend products/services based on what others have searched for who also search for what you want? How much better would this make the user experience?
The rabbit hole of data suddenly takes on form and function when put into context. This is what the Bing Predicts model is doing with Elections 2016, with more complexity and specificity than ever before. It’s a fine-tuning on a massive level, a hint of things to come.
AUDIENCE BUYING ON STEROIDS
Bing isn’t offering access to its magic formula anytime soon. But it’s not hard to imagine a time when the nowcasting model will be the only marketing strategy we need. Even when analyzing time periods as full of surprises as the presidential election, nowcasting has its place. We’re talking about triangulating signals that drill into an audience, an audience that is hyper-likely to want what you have. Stay tuned.