Predicting the Presidential Election is practically a national sport. However, traditional predictors – especially the talkshow hosts on Fox News – have historically been terrible at calling the next set of numbers. It took Nate Silver’s exceptional statistical skill to show us that with public data, you could accurately predict the election down to the last winning percentage – if the mind doing the calculations was good enough.
Artificial Intelligence has evolved exponentially over the years. We’ve gone from Deep Blue beating Gary Kasparov to DeepMind mastering Go. A Japanese AI just wrote a novel that almost won a literary prize. We may not have another Nate Silver, but the world is in a position to create his machine analogue.
Which is why we at WSO2 have constructed a system designed for the sole purpose of election math. While Google and Microsoft have been happy to use their gifts to play board games and embarrass themselves on Twitter, ours, powered by WSO2 Machine Learner, has been set the task of picking the next POTUS.
Deep Huge, as we’ve called the system (a nod to Deep Blue) predicts that Donald Trump will almost certainly win. That is, if he picks the former Governor of California, Arnold Schwarzenegger, as his Vice President.
To state it in numbers: there is a 52.3% chance that Donald Trump will win by himself, regardless of his choice of VP; with Schwarzenegger, there is a 99.4% chance that Trump will defeat all others and become the next POTUS.
How it works
Like Nate Silver’s FiveThirtyEight itself, Deep Huge’s predictions are probabilistic. We use poll data from the Associated Press, historical records. earlier elections, news articles, secret NSA surveys and Twitter for exploring sentiment and secondary issue mapping. This data is then fed into WSO2 Machine Learner, which computes the prediction model.
Since sources other than polls are not representative, we have paid more attention to trends rather than absolute numbers, and extrapolate the poll predictions while using other sources as calibrations. The current model analyzes the win probability of presidential candidates and then runs this against an array of potential vice presidents.
At the start of the elections, the probability matrix was far too diffuse for any prediction to be useful. However, as the candidates dropped out and campaign tactics solidified, the predictions become more accurate. Deep Huge has successfully modeled the key pitfalls such as shifts of public opinion and the problems with running your own email server.
Every model shows that the choice of a VP is critical, as the second most powerful player in the game brings their own voterbase with them.
In this case, the former Terminator not only solidifies Trump’s position in California, but Schwarzenegger offsets concerns—particularly among men—about Trump’s small hands.
The two men also share strong similarities, including a desire for closing the Mexican border. Schwarzenegger also has a track record of what one might consider Trump-like, politically incorrect statements, such as in 2007, where he urged Hispanic journalists to “Turn off the Spanish television set” and “Learn English”.
Other potential Republican VP candidates provided small gains or even losses to Trump’s odds of winning the 2016 US presidential election. Notably the probability of Trump winning the general election was 53.4% with US Senator Ted Cruz, 46.2% with former Alaska Governor Sarah Palin, and 65.1% with Fox journalist Megyn Kelly.
“According to the latest research, in today’s connected world there are just three and a half degrees of separation between an also-ran and America’s next president,” said Dr. Sanjiva Weerawarana, WSO2 founder, CEO and chief architect. “Our Deep Huge project demonstrates the power of combining streaming, batch and predictive analytics to take a pulse on American voters’ sentiments and provide insights into the winning combination of presidential and vice presidential candidates in 2016.”
Deriving secondary insights
We noted while digging into the model that, in the GOP, the divide between campaign position in terms of key issues between Trump and Cruz is semantically closer, while in the Democratic party, the semantic divide is much larger. This makes it harder for the party to rally the voters who are divided in primaries. Further analysis revealed that divide proved to be a major turning point in earlier election outcomes.
To train itself to this level of accuracy, Deep Huge has to date run 11,302 simulations on available prediction data from the previous years, comparing it against the actual results to dynamically build a prediction model using Random Forest Regression.
While it may bear some passing resemblance to 538’s model, it has not been taught the concepts of weighted polling averages and state fundamentals. Its prediction model has been learned and built by the neural network itself, using features from social media sentiment, news articles, poll numbers in terms of campaign issues, and to compute a constantly evolving prediction model.
In the process of building Deep Huge, we’ve gained valuable insight to the uncertainty inherent to elections. While we’re thrilled to have created the machine analogue of Nate Silver, we hope that one day we will be able to scale Deep Huge to predict any election throughout the globe – one bot to predict them all.
We’re also heartened by the fact that after hearing of this prediction, Donald Trump has reversed his stance on outsourcing and decided to have his campaign planning computed by WSO2 Machine Learner running in Sri Lanka.
And by the way… April Fool.
Deep Huge isn’t real, but WSO2 does keep an eye on politics via our Election Monitor project. This offers a real-time window on the US Election unfolding across Twitter and across mainstream media – mapping influence, sentiment, popular opinion and so much more. Visit wso2.com/election2016/.