Marketing Data: Using predictive analytics to make sense of big data
One buzz word/phrase that became very popular in business circles this year was “big data.” And, even though the term is trendy and probably overused, the overall concept has major implications for marketers.
Marketers are awash in campaign data, more so now than ever before. Email marketing campaigns produce data about open rates, clickthroughs, unsubcribes, and more. Visitor activity on company websites can be tracked, and in the case of registered users or leads flagged for scoring, that activity is not only tracked but also attributed to a particular individual.
Elements tracked can include the website visit itself and activities such as downloading Web content or watching embedded video. That tracking can get pretty granular, such as combining a series of website activities, or exactly where in an embedded video the viewer stopped the playback.
Taken as discrete pieces, all these data points are essentially meaningless. Taken together, they can provide insight into the tracked individual. Furthermore, subjected to deeper analysis, they can provide insight into what the most promising prospect or customer with the most long-term value looks like for the company.
This is where predictive analytics come into play. To provide more insight into predictive analytics and big data, I interviewed Omer Artun, CEO and founder of AgilOne, a cloud-based predictive marketing intelligence company. Omer also has an academic background in pattern recognition, data mining and complex systems.
Defining predictive analytics
Omer says predictive analytics should be discussed in the context of “exploding data sets,” or “big data.”
He defines predictive analytics as, “About detecting patterns in the data where you have a lot of data, but the information in that data is limited.”
In effect, the data is very “noisy.” He adds that one way to look at predictive analytics is to imagine jewelry hidden in a mound of sand.
“Predictive analytics is about sifting through that noise and that pattern in that data,” states Omer.
Omer also says that this process is important to marketers today.
“Ten or 15 years ago, somebody would come and buy something from you, and you would know what they bought, when they bought it, and how much they paid for it,” he states.
Omer continues, “Fast forward to today. You know (what) people are searching for with keywords, and they are coming back and clicking on something, and they are viewing a banner ad. They are clicking an email, and they are doing all sorts of things that are not specifically revenue-generating transactions. But, (these actions) have information in them about (the prospect’s) intent.”
He says 98% of these interactions are noise, but within that noise is information on those prospects’ needs and behavior.
Big data can be used to improve customer acquisition
Omer explains why big data, particularly when combined with predictive analytics, can improve marketing strategies and campaigns.
“The marketer’s goal is to maximize the lifetime value of customers, and minimize the spend that they do it with,” he says.
“[Predictive analytics] allows them to look at all this massive data and break this problem down into finer chunks and ask, ‘How do I spend a minimum amount of money to create the maximum value to a customer,’” Omer states.
The answer is to efficiently acquire high-value customers.
But, he adds, this doesn’t mean cost-per-acquisition for each customer is the beginning and end of looking at the data. Predictive analytics might find that you have two customers. The first costs $20 to acquire as a customer, versus the second who costs $50 to acquire.
Looking at the entire picture with a deeper dive into the data might show the predicted lifetime value of the higher cost-per-acquisition customer to be $500, whereas the $20 cost-per-acquisition customer might be predicted to only spend another $5 over that customer lifetime.
The reasons could be the lower cost-per-acquisition customer came to the company through a sweepstakes or shopping comparison engine, and did not become a customer with any sense of loyalty.
“It completely changes the way you would invest your [customer acquisition] dollars,” Omer says. “That is where predictive analytics help optimize the acquisition part of the equation.”
Predictive analytics also applies to retention and reactivation
Getting deep into your data is a great way to determine the most cost-effective customers across the entire lifespan, and to help determine the best customer acquisition marketing efforts. However, predictive analytics can also be applied to two other marketing pain points: keeping existing customers, and retargeting those customers who are no longer actively buying from you.
Retain existing customers
Omer says one way to retain an existing customer where predictive analytics can be very effective is in product recommendations.
He explains that understanding the products a customer is looking at on the website, the keywords they are typing into search engines, and what they purchased allows a marketer to both predict what that customer might do in the future and make product recommendations based on that knowledge.
Another aspect of customer retention is looking at the entire customer database to find personas that evolve, and Omer adds to not forget that individuals can belong to more than one persona.
He says an example might be a health products retailer that finds one persona cluster including moms buying products for their children and another of people buying sports nutrition enhancements. If a customer was a mom buying for her children, she might also be training for a marathon, meaning she would also be part of the sports nutrition cluster.
Omer explains when you are equipped with this knowledge and developing a relationship with this particular customer, you can target both the content of the email and offers you send her and the landing pages or areas of the website you direct her to visit.
“That is where predictive technology comes into play — whether it is product recommendations, clustering or understanding propensity by behavioral attributes,” says Omer.
Reactivate old customers
One other place predictive analytics can improve marketing efforts is getting those customers back who have stopped buying from you for one reason or another.
Omer describes a hypothetical company where 50% of the customers are either buying for the first time or have been reactivated, and the other 50% are ongoing regular customers.
He says at this business, after six or seven years of customer attrition, there are one million regular customers with an additional five million in the database who have not bought anything for some time.
At what point is it more effective to attempt to reactivate these parts of the database measured against just acquiring a new customer altogether?
“Predictive technology comes into play in helping you tradeoff between reactivating an existing customer versus acquiring a new customer in the sense that you can predict who is likely not to come back,” states Omer. “There is a tail end to their lifecycle, and they stopped buying from you, and (predictive analytics) gives you the ability to (have) more surgical pricing, or surgical offers, to reactivate those customers.”
He explains that a 10% discount offer might not be enough to bring that customer back, but a 50% discount offer might be more expensive than just running a PPC ad campaign to obtain new customers. Predictive analytics can help find that discount level where it makes sense to retarget lapsed customers.
Omer adds one wrinkle to this process – whenever there is an issue with inventory where there is an overstock with excessive storage costs, or maybe a product with a looming expiration date, this inventory can be used with deep discounts to target the non-active portion of the database because you already know what they might be likely to buy.
He says, “With this predictive thing, you are basically optimizing your inventory, and at the same time, you are reactivating your customer. You are basically killing two birds with one stone.”
Of course, bad data will produce bad results
It’s the old garbage in, garbage out argument. Omer reminds us that data quality and data cleansing are important issues when engaging in predictive analytics, “because you need clean data to do better predictions.”
With good data, marketers can begin a model building process.
Building a model for predictive analytics
Omer describes five major steps in model building for predictive analytics:
- The first step of a well-designed system is bringing all your data together and performing data cleansing.
- Ask questions of the datato create predictors. These question include:
- What and when did the customer purchase? When was their first purchase?
- What did the customer search for and when?
- When did the customer view this specific banner ad?
- How many categories have they purchased from?
- What is the average time between purchases?
- Does this person buy seasonally?
- Do they use a coupon?
- Do they visit through a specific channel?
Omer explains these questions help uncover what data points are relevant in understanding the customer to generate features.
- Feature selection requires understanding what you are trying to accomplish — the problem you are trying to solve. This step involves choosing features based on what customer behavior or attributes you trying to predict.
- Classified design is the step where the actual predictive analytic algorithm comes into play; these include both supervised and unsupervised learning algorithms.
- The final step is system design and involves calculating errors as the model runs.
Omer says, “Periodically check your model to make sure its accuracy continues to stay the same, and if it degrades, you need to do the whole cycle again to make sure the model stays accurate.”
He adds at this stage, the model should be tested against the data on an ongoing basis to “make sure your predictions don’t go out of whack.”
“So, there are two stages,” says Omer. “There is a model building process [steps 1-4], and then there is a model deployment process [step 5], which are two separate things. You need to automate the process if you want to be able to do this at large scale with large amounts of data.”
The value in this complex process
Predictive analytics is not a one-and-done marketing effort. It involves a very deep dive into the database, and a certain sophistication of data handling to be effective. But, it can also be very valuable for today’s marketer awash in data – lots of noise, with actionable information buried inside that noise.
Omer states, “Predictive analytics becomes much more powerful because it gives you what I call amplified intelligence. It basically tells the marketer, you know, you can make meaning out of this data, and predict persona and behavior.”
He continues, “The other aspect of it is ROI. I have an ROI on the long term of customers, and now with predictive analytics I can start understanding that. If I can predict long-term behavior, I can affect long-term behavior, as well. That’s what predictive analytics brings.”