Predictive marketing: It’s no crystal ball

There’s a scene in the Netflix series Billions, where upstart hedge fund manager Taylor Mason stands alone in a dim empty low-ceilinged office lined with rows of empty desks with little but the signage of Mason's new budding ‘quant’ fund firm illuminated in the background.

It captures the shift in power, where Mason's old bosses, the hedge fund managers, once financial disruptors theselves, are now being disrupted.  We know Mason’s new firm does not need those desks to be filled, it really just needs the one guy with the laptop in the back room, who has built a new algorithm that can more ably detect, through quantitative analysis, which stocks are about to drop or uptick and accordingly execute ‘faster, better and cheaper’ than a room full of traders.

There are parallels here to the new real world of marketing. When we reach the chief strategy officer of Mutiny, the Australian startup billing itself as “the world’s first predictive consultancy”, Henry Innis explains he’s still finding his way around the company's new office in Melbourne as it works towards filling it up. Led by former agency leaders, Mutiny aims to do to the marketing industry what ‘quant’ funds did to the financial industry, Innis believes.

Actor Asia Kate Dillon playing Taylor Mason in 'Billions'. Source: Netflix

“The media market is now being digitised in the same way the stock market was,” he explains, noting that even a level of television as well and radio can be bought programmatically. “What that allows you to do is to turn every piece of media that is traded on programmatic networks into a data point and provide analysis. The strength of that is you can start to actually predict trends in the media market, forecasts in the media market and forecasts against clients’ data and activity. We take a ‘quant approach’ to understanding a lot of the marketing problems based off media, financial and sales data, rather than looking quite executionally. And I think agencies come from a very executional background.”

Just like in the financial space, the explosion of digitised data, coupled with cloud-computing capabilities, has opened up marketing to all sorts of predictive analysis. Virtually all outcomes can be modelled and predicted now with a higher degree of accuracy than ever before. Algorithms have the potential to sift through reams of data to predict which sales leads are most likely to convert, which customers are receptive to upselling, what the outcome will be to changes in brand spending, which type of message is more likely to resonate at any given time, which creative campaign is likely to be more effective, which tweaks to a campaign will trigger an action, and the list goes on.

In essence, much of the guesswork and pseudo-science is being kicked out of forecasting. Predictive marketing is anything but a specialisation in the dark arts of reading crystal balls. It’s the opposite, it’s a science and it’s far more rigorous…and yes, boring.

“The heart of predictive marketing is data," says Anurag Gupta, COO at data-driven digital agency and consultancy Ada in Kuala Lumpur. "It is not magic. It’s a lot of data analysis.”

Right customer, right time, right message

Despite all the potential predictions data can make, Gupta says, marketers largely want one thing.
“At end of day what are we trying to predict? Predicting that this person is the right person to buy my product. If we cut all the jargon it’s predicting that this person is the right person to buy my product and he’s ready to buy it now.”

So in practical terms, it means a handset maker like Samsung or Oppo wants to figure out which people are about to upgrade their phones, Gupta explains. More than that, they want to know where they can reach them, if they’re on Facebook, watching videos or playing games. The preponderance of data means firms like Ada can “go deep into their behaviours, whether they’re photographers, gamers, surfers, to know what they do with the phone so you can position your phone in the language so that he or she tries to buy your phone.” Naturally, Gupta says, it’s the industries where the cost of acquisition or lead generation is highest, such as the financial sector, where Ada’s clients, like FWD Insurance or DBS, will most want to use predictive analytics to isolate the leads most likely to open accounts.

“I would like us to be able to predict financial results far more effectively to our CFOs and empower marketers to have a proper seat at the table.”
—Henry Innis, Mutiny 

Meanwhile, digital media agency Essence, which is also looking to target prospective buyers at the right time and place, has also been using predictive analytics to power programmatic buying. Using effectiveness data from marketing campaigns over the past decade, explains Aarti Bharadwaj, APAC vice president of client analytics in Gurgaon, the agency has captured an understanding of how marketing can change behaviours or attitudes towards products and services.

Essence then built a model to identify data signals which predict if someone is likely to adopt the product or service in the long run. The difference from other media planning and optimisation tools, explains Bharadwaj, is that “all the signals we’ve optimised against are those which are predictive towards the final business goal”, such as sales. “That’s a layer that predictive analytics is able to give,” she adds.

Modelling school

It is the construction of models testing hypothetical future behaviour that gives the practice of predictive marketing its potency. Rather than let a brand client send a straight email to a potential customer showing some signs of interest in a certain product, predictive data scientists at Mutiny create an algorithm to make the data work harder, looking not only at all CRM conversions but also at all events in the digital and media ecosystem so they can assign scores to them. That allows them to predict when someone is likely ready for acquisition. They can then create custom segments of potential customers.

Image from Mutiny's website reflecting its 'predictive' positioning

These scoring methodologies are used to create “cast models,” Innis explains. “The real smarts of it comes in the fact that the mode is circular, so that as we get more conversions the model gets smarter as market conditions change.” Because it has deep learning, the model will retrain, constantly changing those clients-acquisition messages and sending them to different people based on their digital interactions from client first-party data.

This gives the client competitive advantages, Innis continues. “You can figure out, for example, what are the 10% of customers that are going to churn? That might be who you start launching discounting offers at. Or you might figure out the 10% of your email list that are likely to convert without you having to send an offer to them. So you might be able to get more price benefits in your acquisition. Or it might be that you can target your media mix more efficiently and effectively. It kind of unlocks a whole heap of things.”

How to get predictive marketing wrong

Just like you can’t train someone in chemical engineering and then ask them to be a software engineer, these hypothetical models have to be trained to the right business outcome. If your machine-learning model is just doing click-based attribution, Innis notes, of course it will train itself wrongly.

“Everything revolves around that hypothesis,” says Bharadwaj. “Data is a very messy thing to deal with. It’s a living, growing beast. As analysts, what we need to do is go with a laser-sharp focus—this is the business problem that needs to be solved and this is the data that needs to come in,” she says. “A lot of the time, the analysis tends to be clouded because of a lot of irrelevant data gets collected.”

That’s why both Innis and Bharadwaj emphasise the importance of focusing on data that gives you the frequency and context of customer actions rather than data solely about who those customers are, like the data you might get from a third party data broker. 

“What’s being peddled as performance marketing is just optimisation of media.  Unless and until we have data, media and content all together seamlessly, ‘predictive’ will be a fancy word which lots of people talk about but don’t practice.”
—Anurag Gupta, Ada

“That’s why I think the Axciom and Epsilon acquisitions [by IPG and Publicis] are slightly funny acquisitions because they’re buying a lot of data that’s probably pretty junky,” Innis asserts. “As I understand it, brands are getting much smarter about using their first party data and media data rather than buying it from brokers. They tend to get a better result.”

But even if predictive hypotheses are sound, backed by the right data. Gupta sees a more client-centric problem around delivery.  While consultancies like Mutiny will hand-off their predictive insights for brands and agencies to go act on, Gupta says clients also want the same shop that finds the right consumers and media placements to also provide the right creative.

“The way clients buy predictive advertising is still very siloed,” Gupta says. “What’s being peddled as performance marketing is just optimisation of media.  Unless and until we have data, media and content all together seamlessly, ‘predictive’ will be a fancy word which lots of people talk about but don’t practice.”

Predicting creative effectiveness

But introducing prediction to develop the best creative is another science unto itself. Ad-Lib is an emerging digital creative technology services firm based in London, with a growing Asia-Pacific presence largely built around localising, optimising and automating creative campaigns. Their aim is to quickly determine through testing what creative will work best. As Singapore-based APAC lead Rupert Privett describes, this can involve taking a client “on a phased roadmap so that we can begin to take learnings and start feeding them into testing approaches to scale up that predictive element.”

A good example for this, he cites, is Tesco, since the retailer has a pipeline of hundreds of thousands of products whose inventory is constantly shifting at scale. Working with Tesco’s data-science team, Ad-Lib helps to identify the biggest sellers aligning with different focus audiences and can test them in the creative. Then, taking those creative learnings, they begin to cut them by geography to see how products shift by cities and suburbs of varying affluence.

Image from Ad-Lib video explaining creative optimisation methods

“It got really interesting when we started to overlay day of week, time of day, and weather patterns,” Privett says. “We could really accurately start to predict what is going to drive a better response.”
Patterns began to emerge around ad receptiveness, with barbecue and beer-cooler products selling better on sunny days heading into the weekend, and in contrast ready-made meals and household products faring better on a rainy Monday.

While the above results are intuitive and might reflect inherent buying intentions regardless of marketing efforts, the data would also reveal counterintuitive patterns.  For instance, in Ad-Lib’s work for J&J on tobacco alternative Nicorette, it found any references to smoking itself in creative negatively affected marketing effectiveness. Consumers were less motivated to buy when reminded of the habit they needed to change but more motivated by positive messages of healthier alternatives.

There’s little new about optimising creative based on research and testing, of course. But again, like in media, it’s the speed and scale that has made the feedback so rapidly applicable that it becomes ‘predictive’.

In the past, marketers relied on small focus groups, vulnerable to bias, explains Bharadwaj. Applying data and analytics then became useful to give us the average benchmark on what kind of messaging people like to see, she adds. But now, when a marketer wants to optimise that message, “with predictive analytics we’re able to test them in real time that will give us a very quick read on what kind of response we’ll get from creative concepts.”

“Sometimes predictive analysis is confused for a machine that tells us exactly what button to push and what to do. Very often predictive analytics is not seen as the creative field that it is.”
—Aarti Bharadwaj, Essence

“What we spend our time doing is trying to come up with the most efficient and effective ways of spotting trends and applying them in future campaigns and testing them to the point where we’re confident in the findings and can make predictions on performance based off the back of that,” explains Dougie Budge, a data scientist from New Zealand working in Ad Lib’s London office.

All creative projections, then, are still based on past behaviours and cannot be used to craft brand new ads guaranteed to deliver sales from the start. “Sometimes predictive analysis is confused for a machine that tells us exactly what button to push and what to do,” Bharadwaj says. “Very often predictive analytics is not seen as the creative field that it is.”

A holier grail would naturally be to automate the creation of new custom creative for every single individual, informed by predictive analytics. But Gupta notes the technology still needs a bit of fixing.

“We’re still a little nascent at whether you can produce individual creatives at scale. It is still expensive. So finding the right target audience you can do at scale because you can crunch a lot of data but producing and writing the message which is suitable for every single audience that is customised is a little far away.”

Benefits for chief marketing mathematics officers

When that time comes, and bots are able to successfully predict the best creative for you at the right time and place and automatically craft it accordingly, it will be the ultimate victory of science over art.

In the meantime, predictive analysis is already shifting the playing field in favour of science in other ways, like arming CMOs with stronger empirical data in the boardroom.

A lot of marketers don’t have mathematics backgrounds, Innis points out. He feels they don’t communicate value well enough and need better training with numbers to be able to communicate with the corporate brass.

Back at Mutiny, Innis’ firm employs a form of ‘bayesian’ or probability-based mathematics that allows them to ingest a lot of different variables and build models to simulate what a client may or may not do next year. In other words, it could allow a CMO to figure out likely impact on bottom line, sales and brand metrics if she cut brand spending by 10%.

“We can build those simulations for them to wargame what may or may not happen in a pretty statistically sound way, rather than do what we do today in media marketing, which is we put our finger in the air and go, ‘ah, I think we need to invest more in brand’,” Innis says.

This, in turn gives more fluidity to the CMO's whole budgeting process. Much budgeting today is still based off what marketers did last year rather than how they expect the next year to turn out. Moves towards zero-based budgeting in large organisations haven’t translated yet into an effective operating model for marketing, Innis argues.

“We’d love to add a lot of that firepower,” he says.“I would like us to be able to predict financial results far more effectively to our CFOs and empower marketers to have a proper seat at the table.”

Statistically sound models based on strong mathematics can do that, he asserts. “We want to add that mathematical rigour that’s a bit lacking in the industry.”

It took awhile, but the nerds, it seems, are finally getting their revenge.