Machine Learning Times
Machine Learning Times
Three Best Practices for Unilever’s Global Analytics Initiatives
    This article from Morgan Vawter, Global Vice...
Getting Machine Learning Projects from Idea to Execution
 Originally published in Harvard Business Review Machine learning might...
Eric Siegel on Bloomberg Businessweek
  Listen to Eric Siegel, former Columbia University Professor,...
Effective Machine Learning Needs Leadership — Not AI Hype
 Originally published in BigThink, Feb 12, 2024.  Excerpted from The...

11 years ago
Numbers Game: Will Predictive Analytics Decide September’s Federal Election?


A week after November’s United States presidential election and Nate Silver was everywhere. Already an established member of the American political commentariat via his blog, Silver’s name went into the stratosphere after he correctly predicted the winner of all 50 US States and the District of Colombia in the 2012 race for the White House.

Silver went from being a political analyst on news television to a regular guest on mainstream talk shows across the United States. How had his predictions been so accurate? And what did it mean for future elections in the US? Silver had become the country’s political oracle, in the process doing a bunch of entrenched pundits out of a job, and his profile was rammed home with the release of his own book, The Signal and the Noise.

But as the northern autumn turned to winter and the United States settled in for the January swearing in of Barack Obama for a second four-year term as president, a new story began to emerge. Silver had become a lightning rod for the statistical community and ‘big data’ believers the world over. But by January it was clear his statistical predictions were actually a secondary story – a sidebar to the most phenomenal harnessing of information in electoral history.

“He was doing forecasting,” says Eric Siegel. “So for an entire state of voters – especially swing states – it was interesting. Whereas the Obama campaign’s use of predictive analytics was making predictions for each individual voter in the state.”

Predictive analytics. Persuasion modelling. Machine learning. These are the terms that have emerged out of the fog of numbers documenting the election. Siegel, Ph. D., is the president of Prediction Impact, a US-based analytics and data mining firm, and the author of Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie, or Die. He watched with fascination as the real story behind the presidential vote became apparent.

“Which is more powerful? Winning at a prediction, which Nate Silver competed publicly to do. Or winning the election itself, which by its nature of course is going to be done less publicly. Predictive analytics: a defining characteristic is that it’s predicting for each individual customer or person or patient or voter. That means it gives it the power not only to predict but to influence the future.”

Predictive analytics has already been widely used throughout the commercial sector, in situations ranging from the curious to the profound. Prison parole boards predict which prisoners might be repeat offenders. HP predicted employees who were likely to quit their jobs. In one particularly infamous case, Target used purchase history and other customer information willingly disclosed to the company to predict pregnancy among their female customers.

In each of these instances, the value of predictive analytics to change the future is clear: don’t release the prisoner, adjust working conditions for the disaffected employee, lift sales with the targeted marketing of baby products. But the Obama campaign took such techniques to a whole new level. Using a particular brand of predictive analytics known as persuasion modelling, a team of 300 data analysts went to work augmenting data sets from the Democratic National Committee with reams upon reams of polling data. A poll on its own is a method of forecasting, but combined together with other data – both from the DNC and in some cases acquired from private sources – they allowed the Obama team to make predictions per voter and run models on whether each particular person was persuadable.

Now, the Democrats wouldn’t be wasting time talking to people who had no chance of voting for Obama. Working from one single data source, they could utilise all of the channels available to them – email, Facebook, TV and mail drops – to target their list of ‘persuadables’.

Mitt Romney had his own analytics operation – Project ORCA – which he’d used to give himself a clear edge in the Republican primaries. But it was nothing compared to the scale of data collection being undertaken by Obama’s campaign team.

Siegel tells of shooting the breeze with a colleague months before the presidential election: “We were speculating, ‘Gee, I bet you within four or maybe eight years we’re going to find out one of the presidential campaigns has used this stuff,’” he recalls. “Then, a month after the election, that’s when I really got to interview the campaign’s chief data scientist and found out, ‘Oh my God, they actually did this!’ And they did show that it was a factor that got them significantly more votes.”

Extrapolating the data

Obama’s chief data scientist was Rayid Ghani. Ghani neatly encapsulates predictive analytics as the ability to personalise at scale. “There’s no other way,” he tells TheVine. “This is the only way you can do it for millions of people and still keep it very personalised.”

Ghani is talking from Sydney, where he’s travelled as a guest of CeBIT Australia, a global business meeting, exhibition and networking event. Ghani’s not the only member of Obama’s analytics team who’s been in the country recently – Joe Rospars, the president’s principal digital strategist and adviser, appeared at last month’s Sydney Writers’ Festival – and he says there’s been plenty of Australian interest in the big data techniques used during the presidential election. “Not just Australia, but pretty much any country where there are elections happening – they’re all interested in learning about how we did certain things and how it transfers to their electoral system and how they can apply some of these techniques.”

One particularly interested audience member at the Rospars panel was Rosanne Bersten, national digital communications coordinator for the Australian Greens. “The 2012 campaign bumped analytics up to another level again,” Bersten says. “I was fascinated to hear [former chief of staff to John Howard] Grahame Morris and [Labor’s 2007 election advertising advisor] Neil Lawrence on the same panel effectively saying that Australian parties don’t need American-style campaigns.

“I think there’s an understanding that you watch what’s going on in different campaigns around the world and ask, ‘OK, how can we use that? And how can we use that?’”

So how might predictive analytics apply to the upcoming federal election? And how should the different political parties be using it?

To Ghani it comes down to whether a political party is taking a grassroots approach or focussing on high level campaigning, such as television ads and broadcasts. “If you’re doing personalised interaction, you have to figure who are the persuadable voters,” he says. “That’s where two things are critical: being able to figure out, at least at a small level, who these people are, and then you’re going to have to run predictive models that predict for the rest of the country how persuadable they are.”

Once a campaigner has their list of persuadable voters, they focus their efforts from the most pliant to the least until they’ve run out of resources.

“The second thing: if you have a list of persuadable people you have to use all the mediums and channels you have access to and use that data,” Ghani says. “So if you’re sending emails to people, if you’re using Facebook, if you’re putting TV ads out, or you’re dropping mail, or anything – it all has to come from the same data source that has this list of persuadables. If you spend any money, you want to see how many people you’re affecting and get a sense of the unit cost and really unify that across the board.”

For Siegel, it’s all about persuasion modelling. “Presidential campaigns are different to normal marketing campaigns in the sense that the stakes are so high that you leave no stone unturned,” he says. “With a marketing campaign, if you’re making a bigger profit, you’re happy. With a political campaign, improving the situation to a certain degree is not good enough.”

Persuasion modelling is more analytically complex than other types of predictive analytics, and essentially requires a second set of data – a control set – and some special techniques to leverage that data. “Persuasion modelling is the key analytical method to do better with prediction,” Siegel continues, “because it literally tells you who you should contact and who you shouldn’t, and that’s what you’re ultimately trying to determine … So on a very detailed, granular, actionable level, that’s what persuasion modelling does and that’s the game changer.”

Lost in Translation

But how well do lessons from the United States really apply to Australia?

Nick Drewe illustrated the value of big data earlier this year when he and his Warmest 100 team blew open the national youth broadcaster triple j’s Hottest 100, correctly predicting 92 out of 100 entries on the final countdown. Like many others, Drewe watched with interest as the data-driven story behind Obama’s election win started to seep into the press. But he’s not convinced predictive analytics will have such a profound effect on the Australian federal election. For starters, he says, there are numerous technical limitations that blunt the effectiveness of micro-targeted campaigns.

“As far as we could see, the granularity just isn’t there to the same extent in Australia,” Drewe says. “Even things like geographic targeting on social media isn’t available to the same extent. In most cases we can only pinpoint someone’s location to their nearest capital city via the IP address. We can’t get it to their suburb, or even their region.”

According to MaxMind, an industry provider of IP intelligence tools, Australia has the second lowest Geo-IP accuracy of all the OECD countries. MaxMind can correctly resolve an IP address to within 40 kilometres in 81 percent of cases in the United States, but only 58 percent of cases in Australia. Practically speaking, this limits the targeting capacity of networks such as Google AdWords. “As an online marketer, its quite frustrating when ad networks roll out all of these awesome targeting features only to be let down by Australia’s telecommunications infrastructure,” Drewe says. “[The US] rely a lot more on cable, whereas we’re a lot more reliant on DSL connections here, and the US run at a neighbourhood level … and that allows your IP address to be targeted to that specific neighbourhood.

“[In Australia] they all go back to an exchange, and then back to your ISP, where your IP address is assigned – rather than at that really neighbourhood level.”

As an example, Drewe provides two Google AdWords screenshots. In Australia, only the closest capital city can be targeted.

But in the United States, it’s possible to target down to an individual suburb, Nielsen ratings region, and even congressional district.

“I think it’s probably still valuable at a more a qualitative level,” Drewe says. “What people are saying, what they’re talking about, but in terms of pinpointing down to a really close seat and targeting those people specifically, I’m not that convinced that’s quite at the level that we’ve seen in the US elections.”

For chief Liberal Party strategist and pollster, Mark Textor, the limitations have more to do with privacy. “I just think it’s the way you can’t merge data,” he says, referring to the Commonwealth Privacy Act, which forbids data collected for one purpose being sold onto a second entity for a different purpose. “We cannot buy Coles customer data and we can’t buy customer data from the online networks, because that would be illegal. So a lot of the stuff they do in the States … just isn’t possible here.

“They are exempt from the Privacy Act, the political parties, but that doesn’t mean other third party sources are exempt. So you’re limited in what you can source elsewhere, which is the point of big data – you merge databases to create a rich dataset.”

On this point, Eric Siegel disagrees: “For most – but not all – predictive analytics projects, an organisation in fact uses only their own internal data, without acquiring data from a third party. There’s often plenty of available data to develop effective predictive models without going outside the organisation.”

But perhaps the most elemental difference between the United States and Australia is one of electoral systems. In the US voting is non-compulsory – of a population of over 300 million, only 126 million voted last November – so political parties have to first mobilise a support base. Also, individual campaign donations are considered of vital importance in the United States. By comparison, voting is compulsory in Australia and there’s not nearly the same attention paid to ground-level donations. The focus in elections therefore shifts instead to capturing the swing vote.

“In America the challenge is to get people out to vote,” says former Labor strategist, Bruce Hawker, who is currently writing a book on political campaigns and elections. “They knew the people they were calling with a reasonable amount of certainty how strong their intention to vote would be – if they could get them there. That is an important distinction to make in this area. With compulsory voting we’re focusing very much on the marginal seats and swinging voters.”

Current Labor chief pollster and campaign strategist John Utting agrees: “The American system is fundamentally all about mobilisation,” he says. “It’s less to do with persuasion. The recent Obama-Romney election: basically, support broke down to a 50-50 split, but the winner was the candidate who got a better degree of mobilisation. In Australia, we’re already mobilised. So our target is people who are essentially disengaged from the political process and not that interested and not particularly committed to following it.”

Even locking down somebody’s political stripe is much harder in Australia compared to the US, simply because voters don’t declare their allegiance by being one of those who chooses to make his or her way to the polling booth on election day. “Exactly,” Hawker says. “Here people are dragged kicking and screaming to the voting booth. In America they’re not afraid to have a point of view, whereas in Australia it’s almost something to be ashamed of. Don’t mention the war, don’t mention your voting intentions,” he laughs.

Ditch the Calculator?

But besides the question of whether predictive analytics could be applied to Australia, there’s the idea that it doesn’t even need to be applied.

As far as Mark Textor’s concerned, Australia’s compulsory system long ago put large tracts of data in the hands of the major political parties, negating the requirement to understand the nature of people who are likely to vote. “The need to use big data as a predictive model isn’t there,” Textor explains, “because we have a much greater handle on the voting patterns in marginal seats. Whereas they’ve got 300 million people and you’ve got to know where they’re living and who’s going to turn out, in Australia, everyone turns out. So you don’t have that requirement to understand the nature of the people who are likely to vote, because everyone’s likely to vote. And therefore other data sources – the usual data sources – are sufficient to predict their behaviour.

“The predictive part of it, you have it with normal polling analytics anyway. Not to say there isn’t a matching up of polling data with concurrent internet polls and other things. But there’s just not the same requirement of data and need for data as there is in the US.”

John Utting is more encouraging, but he too thinks predictive analytics will very much be a secondary tool at this year’s election. “There have been developments,” Utting says. “But I don’t think it’s gotten to the point yet where it’s a real game changer or whether or not it ever will be. Not that I want to talk it down, but just putting it into perspective.” Utting does concede, though, that Labor modelling has become a lot more sophisticated, moving away from traditional, regression-based systems towards areas of machine learning and computer science-based approaches.

But what about the other parties? Is there a case for predictive analytics being comparatively more effective for the smaller players on the Australian political scene? The Greens’ Rosanne Bersten is in no doubt: “I think it will be. We’ve always been a decentralised grassroots party. We’ve always had people on the ground. Anytime anyone talks about American-style campaigning they’re talking about grassroots organising. It’s something that comes naturally to us, whereas for the bigger, older parties it’s this big about-turn.

“As far as [Labor and the Coalition are] concerned they’ve almost got 50 percent of the vote anyway – it’s just a matter of these people they need to switch over. For us it’s a bigger question. It’s about saying, ‘Hey, what if you completely change the game?’ We’re talking about mobilising people to go door knocking because they care, and say, ‘Don’t do the same thing you’ve always done.’ And in some ways I see that as very similar to mobilising people to get out and vote, and that’s where I see what we’re doing is similar to the American-style campaigns.”

Both Rayid Ghani and Eric Siegel agree. “That’s a really good point,” Ghani says. “You need that base level of resources, but it does allow smaller parties to focus their efforts where they’re most useful. So instead of calling people who aren’t going to be persuaded, they can focus their efforts and do a lot more for less.”

Siegel adds that a predictive analytics approach can be particularly effective if you get the right experts for the right price, using existing software that is often open source and free. In that respect, Bersten is happy to back the Greens: “I can certainly tell you that we have a wide variety of young, enthusiastic, digital people,” she says. “I feel we are recognised as being ahead of the game digitally and technologically.”

The Greens’ focus on grassroots campaigning perhaps hints at the real effect Obama’s win has had on the upcoming election. Predictive analytics may be seeping in around the edges of the local political landscape, but if the US Democrats proved anything it’s the importance of communicating at the ground level with voters.

In this respect, Bersten might back her party’s ability to punch above its weight, but there are also rumblings of life from a federal Labor operation that has intimidating reach via its extensive branch structure. The party now has a large phone bank operating out of Parramatta. “Our field campaign has made 250,000 calls this year,” explains Labor’s national secretary, George Wright, “which is a ten-fold increase on 2010. Grassroots voter engagement is providing us with 90 percent of our information.”

And while the Greens claim to have a lead in Facebook and other forms of social media, John Utting disagrees: “I doubt that. If anything, if you look at the Greens vote levels, they’re in a slow decline.

“But yes, in a perverse way, this whole analytical approach is bringing us right back to the beginning of politics,” he says, “where some sort of personal contact is increasingly important. What drove the Obama campaign, despite all the analytics, was strong door-to-door canvassing, making contact with voters, logging their concerns and then feeding that information back into the system. Fundamentally, I think we’ve got a much bigger advantage because we’ve just got an extensive branch structure and lots of active members. We can just get more people on the ground in that really basic form of retail politics.”

The major parties have had the statistical means to conduct such campaigns for a number of years. Both run established databases – Electrac for Labor, Feedback for the Liberal Party – and since the late nineties have been able to break down communities into units of what is now known as the Statistical Area Level 1 (which in 2011 replaced the Census Collection District system). A current SA1 contains around 200 households, so a politician can direct a particular message to a neighborhood or cluster of households that might find it more relevant.

Regardless, predictive analytics only thrived in the United States because the Democratic Party anticipated such a close fight for the White House. Taking such a data driven approach to an election will help a politician nab a last percentage point, rather than the last ten percentage points. “If you’re at a stage where you’re losing an election by five percent, you’re absolutely right that you can’t do anything,” Rayid Ghani says. “If we’d been ten points behind there’s nothing we could have done to help us win. It only helps in the margins.”

Which is a telling observation when applied to Australia, where the latest Newspoll has Labor trailing the Coalition 42 to 58 on a two-party preferred basis. Bruce Hawker is blunt: “This analytical approach – this highly granular approach – it’s still only as good as the message you’re actually giving them … it comes down to the quality of the message and if you’re saying things that they’re going to be drawn to. You can never get past that, really.”

By: Matt Shea, journalist and writer, TheVine
Originally published at TheVine.

One thought on “Numbers Game: Will Predictive Analytics Decide September’s Federal Election?

Leave a Reply