Dean Abbott, Co-founder & Chief Data Scientist, SmarterHQ
Putting Predictive into Perspective
Let’s say a brand recently created a series of category-level browse abandonment campaigns, all based on a “gut feel” of what amount of engagement was enough to warrant nurture campaigns. But they believed they could better optimize the customer experience by only sending messages to those who were actually interested in the items they were browsing, versus sending to those who were casually window shopping. Could they identify which customers were likely to buy in the future (say, within seven days), even if they didn’t have the rules to find this audience?
Their data science team went to work, and the models they built told some obvious stories: Those who buy a lot of things are more likely to buy more things. But the models found other trends that weren’t as obvious: Some customers were moderately likely to transact but had never transacted before, even though they were highly engaged on the website.
The model indicated this window of opportunity would fade, so the brand built a new campaign and sent messages to propel those who showed the most interest to finalize their purchase—which resulted in converting 33% of the audience.
How did the models find these patterns without knowing exactly what to look for?
Defining Today’s Predictive Analytics
Predictive Analytics (PA) automates pattern discovery using mathematics and heuristics in the form of machine learning (ML) algorithms or statistical models. For the companies we work with, PA leverages behavioral data and ML algorithms to power customer messaging and personalized recommendations based on predictions of their future behavior, their propensity to convert, likelihood to churn, or potential future spend.
To break this down in the opening scenario, the data team created a data set containing rows and columns, in the same form as you would load into an Excel spreadsheet. Each row represents the “unit of analysis”—such as a shopper, a shopper on a particular day, a web session, or any other level a decision is being made.
Significant amounts of training and experience are needed for a data analyst to gain the data science experience necessary to understand how to properly build models and interpret model findings
Columns in the data represent the characteristics of a shopper at the same level of aggregation. There could be dozens or even thousands of descriptors of each customer’s behavior, demographic, loyalty history, etc. Such descriptors are often called the “input” variables.
If a customer transacted at any point within the seven-day time frame, they were coded as “1”, and those who did not as “0” to label which customers transacted and which didn’t. This column is typically called the “target” variable. The task of the ML is to find which input variable patterns correspond most strongly with the target variable. If we only had a few inputs, the task would be fairly easy—but we may have numerous, which quickly outstrips the ability of analysts to examine all combinations manually.
Enter: Mathematical algorithms that automatically sift through combinations to find the most important predictors of the target. The key difference between Business Intelligence (BI) reporting and Predictive Analytics is the former uses only the inputs and describes what happened. PA associates the behavior in one time period (history) with other behaviors in a different time period (the target variable).
Implementing Effective Predictive Models
Successful predictive modeling requires a team approach of what I often refer to as the “three-legged stool” of successful analytics. First, we need good predictive modelers or data scientists. Second, we need strong database skills so that we can access and build the data needed for predictive modeling. Third—and perhaps most overlooked—is domain expertise, usually provided by a project or program manager, or a stakeholder. This person or persons provide the information that guides the unit of analysis, what the target variable should be, and how the model will ultimately be used.
Most organizations don’t have a team of data scientists at their disposal like large ones often do, which brings us back to the age-old question of build versus buy. Moreover, significant amounts of training and experience are needed for a data analyst to gain the data science experience necessary to understand how to properly build models and interpret model findings.
For companies that desire to scale and remain flexible, investing in a platform that provides their customers the best in predictive marketing techniques allows them to achieve their modeling objectives. But even for companies that outsource their modeling, if they understand data and modeling well, they can manage the outsourced teams effectively and, as a result, increase their likelihood of success.
Marketo reports 79% of consumers say they are only likely to engage with an offer if it has been personalized to reflect previous interactions the consumer has had with the brand. Whether anticipating customer intent to power real-time personalization, or identifying ways to better forecast product demand, the impact of PA is perpetual and imperative for brands looking to increase their ROI and win over customers time and time again.