Using information collected during a customer’s first purchase, a new marketing tool that leverages machine learning technology can provide firms with valuable predictions about the customer’s future behavior, says Eva Ascarza, a marketing researcher and associate professor at Harvard Business School.
By incorporating data most companies discard, Ascarza and her co-researcher devised an algorithm capable of quickly analyzing more than 40 variables to create a “first impression” of the customer after the initial transaction. The algorithm then predicts which customers will become repeat buyers and which will be most responsive to email campaigns, information that firms can use to improve marketing strategy and return on investment.
“Companies are leaving money on the table,” says Ascarza, because they don’t know how to use most of the customer data they collect, especially when trying to manage newly acquired customers, for whom there is no historical data—something known in computer science as “the cold-start problem.”
People who bought more products at the first transaction, especially those who did so in stores, were more likely to be repeat buyers in the future.
In their working paper, Ascarza and coauthor Nicolas Padilla, a doctoral candidate at Columbia Business School, show that a statistical tool known as a deep exponential family model can effectively sort the most relevant variables and account for complex relationships between them.
The algorithm creates a first impression of each customer much the same way two people ascertain each other’s character on a first date—by interpreting certain cues to infer important traits.
“Let’s say you’re on a first date and you want to figure out whether this person is going to be a good partner in life,” Ascarza says. “You pick up cues from the person’s behavior, including what the person is wearing, how they treat the other people in the place, how the person treats you, what comments they make, etc. Similarly, the data we use from the first transaction, we believe, provide cues as to how the customer will be.”
Testing data on 13,000 shoppers
To show how the model works, Ascarza and Padilla fed it some customer data supplied by an international retailer that sells its own brand of beauty products exclusively in its stores and on its website. This included about four years of transactions by more than 13,000 customers spread over six markets.
The researchers extracted six variables to represent first-time customers’ acquisition behavior, including whether the purchase was made online or offline, number of items purchased, prices, discounts, whether the purchase was made during a holiday period, and whether the customer purchased a newly released product. With only those six variables, the authors show how the algorithm can better identify future heavy spenders and those who will be most responsive to future email promotions after just one transaction.
In particular, they found that people who bought more products at the first transaction, especially those who did so in stores, were more likely to be repeat buyers in the future. Likewise, the algorithm tagged people who bought newly released products as potential high-value customers who would keep coming back.
On the other hand, those who bought discounted products, especially those who did so on Black Friday or during other holiday periods, were “lower-value,” often one-time customers.
Although these findings might seem obvious, Ascarza says, the research provides empirical evidence for the usefulness of forming first impressions of customers, and with richer data, companies could learn even more. For example, companies that use cookies to collect browsing data could incorporate the type of device and search engine a person used to shop online, or the number of products they searched before buying, to generate predictions of what the customer would do in the future.
As with other artificial intelligence tools, the drawback to this technology is that the model acts as a sort of black box, so researchers can’t determine the exact factors that trigger predictions. The same holds true for the predictions Ascarza’s model makes about which customers will respond well to marketing emails. Based on the researchers’ analysis, the beauty retailer’s email marketing campaign would have been 39.7 percent more effective if it had targeted customers based on their first impressions.
“The good news is that you can add as many variables as you want, and the model runs fast and is very efficient,” Ascarza says. “The downside of doing this kind of modeling is that, at the end of the day, it’s very difficult to pin down exactly what makes people sensitive to emails. I can tell you who is sensitive to emails, but I can’t identify one or two variables alone that can explain that.”
What’s inside the black box?
Unlocking the mysteries of the black box is something Ascarza has started to explore. She understands that providing evidence for the effectiveness of new marketing methods only helps if the companies who need them can access and trust them. For that reason, she works to find ways to reverse engineer the prediction model to gain some insight into the relationships between variables and the purchasing motivations it discerns from the data.
“The way you’re going to get more firms on board to apply these things is if we’re able to open the box,” she says.
Retailers and other organizations, such as charities, that seek to segment high-value patrons for promotions and appeals should be thinking about what data they collect, Ascarza says, and how they could learn more from the first contact by leveraging that data and advances in machine learning technology.
Ascarza expects her research will be published early next year and at that point plans to make the code for her model available online for free. With the help of a trained data scientist, any organization that has a few years of clean historical customer data, including repeat customers, could then train and calibrate the model to offer firm-specific predictions about new customers based on their initial purchases, she says.
In the future, Ascarza expects that more and more companies will begin to invest in algorithmic tools to augment decision-making. That investment will in turn facilitate increased understanding about the potential applications and value of such tools both in marketing and in business more broadly, she says.
“I think the potential of the model is way higher than what we show, because firms will have access to even richer data on the first transaction, so they can extract even more information than what we are currently doing.”
About the Author
Kristen Senz is a writer and social media creator for Harvard Business School Working Knowledge.
[Image: Julien Viry]