Math Meets Marketing with Statistical Models
After identifying desired outcomes, marketers will need to create a statistical model baseline, which requires an algorithm that can analyze existing datasets of known points, a process known as classification, experts say. The understanding achieved through that analysis is then leveraged as a means of appropriately classifying the data. Classification, a form of machine learning, can be particularly helpful in analyzing very large, complex sets of data to help make more accurate predictions.
Interoperability is another essential element of statistical models, according to Moran. “You need people to understand what the output means for the business. You need to train and retrain the models. Interpreting what the models mean can be difficult.”
So Moran recommends that companies use data scientists to help build the models and interpret the results they produce.
THE DETAIL’S IN THE DATA
While first-party data (the data that companies collect directly from their customers) will be the most valuable in statistical models because it’s relatively easy to validate, it typically gives an incomplete picture of prospects and customers. For that reason, experts suggest that companies should augment first-party data with additional information from outside sources.
However, there is a fine line between including too much data in the statistical model and adding enough small bits of information to lead to incremental sales, Ziech-Lopez maintains.
Some data might be more meaningful for some businesses and less important for others. Corporate inventory data, for example, might have little meaning for a chain of coffee shops, but could be very meaningful to an e-commerce clothier that needs to keep track of SKUs to know when an item is out of stock so it can keep its website up to date.
Then there are instances when only small bits of additional data can lead to additional sales opportunities. By adding location data to the statistical model, for example, a marketer can deliver a mobile message to a prospect when she approaches one of the company’s outlets. By adding weather information, the messaging from a coffee shop could change from a discount for a hot coffee in cooler weather to a discount for iced coffee in warmer weather.
Data quality is also key, because the analysis created is only as good as the data upon which it is based. Garbage in, garbage out definitely applies.
“Make sure that you are collecting the right data,” says Aadith Moorthy, founder of Radial, a marketing personalization provider. “If you don’t have the right data, you won’t get the right results.”
“You have to have some trust in the data provider,” adds Moran, who recommends testing models multiple times.
While data out of the norm should be ignored when marketing to the masses, it could have value in targeting groups that produced the anomalous data, he explains.
Moran also points out that companies need to consider federal, state, and local privacy regulations when building models and collecting customer data. They also need to avoid redlining, sex and age discrimination, and other forms of bias in building models and collecting data, he cautions. For example, a retailer can use age and gender information to help craft marketing messages, but financial service marketers have to ensure they don’t use this type of detail in loan approvals.
Audience modeling produces customer personas, which can then become actionable when they are activated across media channels, DeCandido says. “Campaign performance modeling produces customer journeys, which detail all of the major touchpoints a brand had with its customers along their path to purchase. Campaign performance models can then allow for smarter go-to-market strategies focused on achieving more and better outcomes.”
The combination of audience models and campaign performance models can be powerful, enabling companies to not only reach the right customers but also to build a one-to-one relationship with those customers, DeCandido adds.
Statistical models can produce different personas for different customers visiting an auto dealer’s website, for example, according to Amit Erande, senior director of solutions engineering and go-to-market strategy at Arm Treasure Data. Depending on those personas, the model can recommend the next best action, with a different recommendation for a customer at the bottom of the sales funnel than for one near the top.