Microsoft Bing already used its algorithms to predict outcomes of events like the 2015 Academy Awards and last month’s Super Bowl. But the search engine’s newest prediction engine may be its most robust to date.
As we noted Monday, Bing is crunching a huge amounts of data and analyzing the 9.2 quintillion possible outcomes of this year’s NCAA men’s basketball tournament to help the public predict winners — and dominate their office pools in the process — with a new tool called “Bracket Builder.”
To learn more about how and why Bing Predicts is now in the college basketball world, GeekWire caught up Walter Sun, Principal Applied Science Manager at Bing, and Bryan Saftler, Senior Product Marketing Manager at Bing, to find out about the group’s mission to offer people data-driven March Madness advice. The conversation was edited for brevity and clarity.
GeekWire: Why is Bing doing this?
Bryan Saftler: “When people want find information related to March Madness, they have sources like Yahoo, ESPN, and MSN. But those are discreet sources that maybe have a very specific feel, versus how individuals tend to find search engines like Bing and Google non-biased and start their search there to get a lay of the land. We saw that as an opportunity to use our search engine as a platform to help people make brackets, find schedules, and learn more about individual teams.”
GW: Tell me more about the “Bracket Builder.”
Saftler: “We had two intents that we saw when people are building brackets that we wanted to go solve. First off, the average person follows no more than five teams. If you have 64 teams in the bracket, how do you pick for the rest of the match-ups? What we’ve done with Bing Predicts is added that as an element and layer into “Bracket Builder” — it’s not meant to replace your selections, but there to assist you.
There will be a bunch of people who start with an empty bracket and click “auto-fill.” But where I find “Bracket Builder” to be a little more valuable is if I go game-by-game. The interesting thing we saw was that people wanted to play the scenario game. Where Bing Predicts really comes in as value is where we can understand what teams are doing past the second round. For example, what if Villanova and UNI meet in the Sweet 16? That looks like it’ll be a close game. But what if I put Louisville in that situation instead of UNI? Villanova’s chances go up by 10 percent. I can start to do scenario planning, which is another thing that lots of people felt was missing in brackets. We’re looking to serve that here.”
GW: Thanks to your partnership with the NCAA, your team had access to more than 10 years of historical data. How did this help your predictions?
Saftler: “This really helped in the predictions modeling because we were able to look back over the past 10 seasons and use that data to predict each tournament as if we didn’t know the outcome. We used that to train our model so it understands how a season goes and how a tournament turns out. The interesting thing is that over the last couple seasons, there have been a lot of upsets. We’ve been able to train our model to understand what upsets would look like and to think forward about what games would most likely have upsets and what scenarios would actually happen to engage that.”
GW: I’ve heard some people say that these data-fueled predictions and tips are “ruining the fun” of what used to be a simple fill-out-your-bracket experience. Thoughts on that?
Walter Sun: “It doesn’t take away from the fun. The idea is that you can still have your preferences — maybe you’re an Arizona fan and you think they’ll go to the Final Four — but you’re also able to see what we think is the statistical preference. Everyone has their own angle in terms of how you predict. We just want to provide statistical data to help you out.
There is no perfect bracket because there is so much variability. We aren’t at a point where we can expect 100 percent accuracy — we hope to get 75 percent. But if you aren’t familiar with some teams, you can use our predictions to trust them.”
GW: What makes Bing’s predictions different than, say, FiveThirtyEight?
Saftler: “Everyone is looking at historical data, but we’ve added a layer of web and social data. This is about taking the knowledge of the collective wisdom of all people who do searches on the web and have conversations on social media and bringing that into production and weighting that as well. Our focus is really on finding as many pieces of information as possible and making sure we are accounting for everything so what we provide is the most complete picture of a prediction.”
GW: So you’re saying that just because lots of fans on Twitter are talking about how good Kentucky is, you take that into consideration?
Sun: “It seems unbelievable, but it does have value. We use a sentiment analyzer that looks at positive and negative sentiments. The funny thing is, the feelings in aggregate have some signals in terms of adequately assessing how good or bad a team’s chance is. To be clear, there is the statistical model. But we also add the wisdom of the crowd. For example, when the brackets came out yesterday, people started complaining about certain teams and how they were seeded. That information, in aggregate, actually has value.”
GW: You guys have predicted the outcomes of The Voice, the World Cup, the Super Bowl, etc., but this seems like the most intensive project yet.
Sun: “This one has much more information available to users. We’re also trying to provide data for all possible scenarios — for the World Cup, we would just predict the outcome of each game and work forward. With this, we built a model for all possible combinations. If Hampton somehow gets to the Final Four, for example, we had to predict an outcome for that. The past models don’t have all this information.”
Saftler: “This is also the first time with Bing Predicts where we not only are predicting who’s going to win, but also have the data that explains what the scenarios that have to happen for an upset to occur. If Kentucky plays Manhattan, Kentucky has a 99.9 percent chance of winning. But what are the scenarios that need to happen for Manhattan in order for them to win? We are actually drawing up those scenarios and sharing them with users.”