It’s easier to sell something if you can predict the future. But the crystal ball does not work. The Ouija board is a bust. There are a few cards missing from the tarot deck.
Marketers want to know: will this customer buy that product?
Propensity modeling is not about divination, but can show the likelihood of a sale, based on the data the marketer already has about the customer. But note that propensity modeling is not the same as predictive modeling. There is a subtle, but important difference. Anticipation, not prediction
“Predictive modeling is any algorithmic process using data and probability to forecast an outcome,” said Wilson Raj, director of customer intelligence at analytics and BI leader SAS. “Predictive modeling is a much broader category. Propensity is a sub-set of predictive.”
Predictive modeling will take into account web data and device data to determine what product a consumer is interested in, and the likelihood they will purchase it. Propensity modeling will look at CRM data instead, using that information to target a message that will most likely prompt a purchase, Raj further explained.
If a consumer is looking to buy a digital camera, predictive modeling will flag past web searches and site visits. Propensity modeling will look at the customer’s record of past purchases, using that data to figure out how likely they will buy something, and if they will buy again. If the customer bought that camera, can you sell him a camera bag? A tripod? A flash unit?
Ready, aim, compute
Understand that propensity modeling is part of a larger assembly to reach the customer. Laura Patterson, president of VisionEdge Marketing, explained this via a borrowed archery metaphor: “The arrow’s feathers are the data (it is the balance), the shaft is the intelligence, and the analytics is the point. The only way to get the arrow to accurately hit the target is with your bow.” she said. “The bow is what determines the force and direction. The farther you need the arrow to go, the harder you have to pull the string back, which means you need more data. So you need data related to these variables: who has bought, what have they bought, when did they buy it, in what order did they make the purchases, and what combination of products were bought by which set of customers?” Patterson continued.
The model must change as the data changes. In a business handling lots of transactions, that means frequent updates. “ Whatever the sales cycle is for the company is the frequency in which the model updated.” Patterson said.
Seeing the data for the trees
The gigabytes of data marketers mine for insight is simply too vast for mere humans to process, much less comprehend. To this end, propensity modeling is heavily dependent on big data, machine learning, and artificial intelligence. “The data is often dirty. It’s not formatted correctly. It has duplicate values, missing values and different distributions,” said Gabriel Mohanna, director of data science at Clarity Insights.
Once the data is cleansed, the fun begins. How do you sift the data to craft your propensity model? Tree-based modeling fits the bill, as it is fast, easy to deploy, but does require an extra measure of training , Mohanna explained. Analysts also need to experiment with making decision trees, so that they can later craft propensity models that can yield insights on messaging, cross-selling, upselling, and repeat sales.
One question, one model, one answer
“Each model should serve one specific purpose,” said Omer Artun, CEO of the enterprise CDP AgilOne. “If you try to mix too many things in the model, you get a meatball.” Adding a question to the question increases the complexity of the model exponentially, and will also require an exponential amount of data. “Nobody has that much data,” Artun said.
One should not use a model to seek the meaning of life, the universe and everything. But you should have different models to gauge different factors of the propensity model. What is the likelihood of a person to buy something? What is the best channel to sell that item? What should be the next best offer? What is the right frequency of contact? The models can be stacked and used in combination to build a multi-dimensional portrait of the customer’s propensity to buy. “The art of the marketer is how you play the stack of models,” Artun said.
How frequently one reaches out to the customer, and how often, is the marketer’s art, reinforced by the insights gained from the model. “The model tells you who and what, not how and when.” added Patterson, as the model is always separate from the marketing strategy.