Consumer Insight: Driving profitability with retail analytics
How to make choices and investments that deliver on expectations.
By Sally Taylor-Shoff (left) and Shalini Raghavan
The most successful retailers today are increasing response rates to their offers and driving profitability by using Big Data and predictive analytics to make relevant, personalized, and precisely timed offers to customers. Predictive analytics provides a concrete means of realizing the long-standing exhortation to “know your customer.” In the era of Big Data, analytic tools have sufficient information to enable retailers to treat every customer as an individual based on insights into their preferences and future behavior.
Retailers who know their customers are making smarter decisions that maximize consumer loyalty without leaving money on the table by offering unnecessary or excessive rebates, discounts and special offers. These retailers are able to hit the sweet spot where customer behavior intersects with what the goals of retailers and their suppliers.
What Problem does Analytics Solve for Retailers?
For retailers, one of the greatest values of analytics is providing decision points for determining how to treat each customer. For example, will it be profitable to offer Customer X free delivery? Or, would that offer be wasted because the customer is going to buy the product anyway?
Consider what happens when consumers visit an online or brick-and-mortar store for the first time. The retailer knows nothing about these potential customers and thus treats them all the same. At some point, the consumer may click on a product or category, or make a purchase. This consumer behavior is likely to trigger a business rule that initiates an action by the retailer. Coupons might be printed at checkout. An e-mail might be sent offering a discount on a product in an abandoned online shopping cart.
In such instances, consumers are differentiating themselves by their behavior. The retailer, however, is still treating them all the same – everyone who exhibits the same behavior receives the same offers. Inevitably the offers will be more relevant to some recipients than to others, and responses will vary accordingly. Because the retailer doesn’t know anything more about these consumers beyond their click stream or the product they purchased, there’s no reliable basis on which to make a more relevant offer.
However, if analytically targeted offers can be made to consumers on a large scale, there is now a foundation on which to perform rapid test and learn. Both relevancy and speed to relevancy are critically important; they enable retailers to differentiate themselves in a crowded market to increase loyalty and margins.
Understanding When and Why to use Various Analytic Approaches
Let’s look at what different types of analytics can tell retailers about their customers, giving them various decision points to consider when determining which actions to take.
Collaborative filtering enables retailers to take a degree of targeted action even with first-time customers. This type of analytics is often behind the product recommendations offered on e-commerce sites and the printed coupons generated at in-store checkout.
The form of collaborative filtering most often used in retail is sometimes referred to as an “affinity model” or “lookalike model.” It identifies relationships between customers and items that have either been purchased or viewed, and infers how an individual will behave based on how other individuals who look similar (i.e., share one or more characteristic) have behaved.
Collaborative filtering doesn’t have to be triggered by a current transaction. It can be used to target ongoing campaigns and other kinds of promotions. Still, this analytic technique is fundamentally transaction-oriented. The algorithms used are best suited to modeling data about items purchased or viewed. They’re not effective for modeling purchases with the wide range of “bigger” data retailers have in their databases (e.g., attitudinal data, seasonal purchase patterns, natural product adjacencies, basket builders) or can access from external sources (e.g., demographics, public records, third-party marketing information).
Clustering algorithms enable retailers to differentiate between customers in broad ways such as customers who like premium brands or customers who prefer organic food. One of the benefits of painting customers with this kind of broad brush is that it can help direct and justify large-scale expenditures on store design, new merchandising schemes and promotional programs. This approach to analytics moves closer to the use of Big Data. In fact, this approach can utilize exceptionally large data sets and a wide range of data types.
Using analytics in this way (often called behavioral segmentation) enables retailers to make far more accurate decisions than can be achieved through traditional methods of database querying on customer attributes such as recency, frequency and monetary value of past purchases. Analytics is more accurate in large part because it can handle greater data complexity. While query-based segmentation generally involves no more than three to six customer attributes, analytic-based segmentation can encompass dozens or even hundreds of attributes, greatly expanding the range of possible segmentation schemes down to a nearly individual level.
Most clustering methodologies try to make sense of these hundreds of attributes analytically. A select few clustering algorithms use a hybrid approach of balancing the analytic approach with inputs from the business user. This ensures that the algorithmic approach is one that is more usable by the retailer.
With many more ways to group customers, and the ability to try lots of alternative groupings quickly, retailers can make better strategic and resource-allocation decisions. One large national retailer, for example, has used analytics-driven segmentation to better understand and serve customers who account for the bulk of the company’s revenues. Using their characteristics to define population segments, and using these segments to guide decisions from store layouts to how staff interacts with customers, this retailer increased same-store sales in the first quarter of implementation alone by 8.4 percent –resulting in a 15 percent increase in total revenue.
Propensity models enable retailers to predict how individual customers are likely to behave. With such specific insights, retailers can differentiate between customers to a much greater degree, further increasing the granularity of segmentation and the relevance of offers. This analytic approach is particularly effective in Big Data environments. In fact, large retailers who serve tens of millions of customers, each of whom has many attributes and preferences, can go as far as to essentially create segments of one.
Propensity models deliver this level of specificity and accuracy because of the ability to handle enormous amounts of internal and external data. Further, they are able to pinpoint the specific customer attributes in the data that are most predictive of a future behavior.
Relationships between numerous attributes and other variables are examined to see how a change in the value of one variable affects the value of another. Attributes that prove to be highly predictive of a behavioral outcome are incorporated into a predictive model. For example, a propensity model can be built to predict a customer’s propensity to make a specific purchase or to discontinue using a premium service.
Because these models can make predictions for individual customers, they open up the possibility of unique treatment. Moreover, by using multiple propensity models, retailers can gain a much clearer picture of the customer. Knowing that Jane is not only likely to buy a 48-inch TV, but that she tends to like the Sony brand, but she doesn’t tend to pay premium prices for cutting-edge technology, enables the retailer to greatly increase the relevance of individualized offers and interactions.
Here are examples of how leading retailers are applying such insights:
- A large retailer is using this type of analytic approach to increase the ROI from promotional campaigns. When a popular new movie or video game comes out, for example, the retailer sends offers only to those likely to buy the product within the offer redemption period. Response rates are two to three times higher than when the same offer is sent to everyone. And because the retailer is not wasting customer time with irrelevant offers, future promotions are likely to be considered by the customer.
- Another large retailer is improving its ability to predict when customers are about to make a big purchase by incorporating customer clickstream data into propensity models. Many consumers do extensive online research before making a major purchase. By analyzing billions of clicks across millions of customers, along with each customer’s purchase histories and historical behavior patterns, the retailer can pinpoint the right moment to make an offer.
If a propensity model predicts when a customer is likely to buy a given product, why should the retailer go to the expense of sending a promotional offer? Uplift models help retailers determine if an investment is likely to be worth the result.
Often used in conjunction with propensity models, uplift models predict the amount of change likely to occur in customer behavior as a direct result of a particular retailer action.
Uplift models can save retailers millions by enabling them to avoid offering discounts to customers who will purchase without them. For example, such a model might predict whether or not a 20 percent discount is likely to increase a particular customer’s propensity to buy a pair of designer jeans within the next two weeks. The retailer can then send the coupon only to customers whose behavior it’s likely to change.
When uplift modeling indicates a customer’s behavior is likely to be affected by a promotion, it can also help retailers determine which promotion will have the most impact. Will 20 percent off be any more effective than 10 percent off? Will any discount be more effective than free shipping? Is offering 12 months of interest-free credit necessary, or will six months be nearly as enticing? Uplift models provide the analytic insights retailers need to make precise decisions about where to put marketing spend for higher ROI.
Uplift models are based on Big Data analytic techniques that can predict individual customer sensitivities to price incentives, redemption terms and even promotional package design. For instance, one company that helps to make markets for new products by spending heavily on promotion, is using uplift models. The analytic insights enable the retailer to accelerate the purchasing behavior of so-called “laggers”—customers who historically haven’t been among the first to purchase. By targeting these customers with offers that are likely to change their historical behavior, the retailer is increasing the concentration of sales in the first two months of the product lifecycle – its critical period before competitors can draft off of their momentum.
As retailers add more data to their analytic efforts, they increase the number of decision points for differentiating between customers and making more targeted decisions. But just as having lots of data can be overwhelming and of little value in and of itself, so it is with the analytic predictions drawn from Big Data. Their business value depends on the retailer’s ability to operationalize them.
Retailers of all sizes are bringing analytics into their operations. The key to making choices and investments that deliver on expectations is to understand what various types of analytics do and how they fit (or don’t fit) what the business is trying to accomplish.
It’s also important to think about analytics as an incremental process rather than a packaged solution. No one approach serves all purposes. Wherever a company is in the spectrum of analytic sophistication and experience, there’s a next step to take to achieve even greater benefits.
Sally Taylor-Shoff is a vice president in FICO’s Marketing Practice. Shalini Raghavan is director, analytic product management, at FICO. Together they have nearly 40 years of experience in retail and marketing analytics.
- 44Cathy O’Neil’s provocative book, “Weapons of Math Destruction: How big data increases inequality and threatens democracy,” created quite a stir in the analytics community when it was released last fall.
- 43Cathy O’Neil, an industry insider and experienced expert, thoroughly covers the sociological downside of data science in her New York Times bestseller and first-of-its kind book, “Weapons of Math Destruction.” In the world of big data, there’s a lot of music to be faced. With all its upside, data science’s…
- 42September/October 2012 Opportunity to deliver real-time customer insights by harnessing the power of structured and unstructured data. By Rohit Tandon, Arnab Chakraborty and Ganga Ganapathi (left to right) The rise of the Internet has created terabytes of data or “Big Data” that is available to consumers and enterprises alike. This…
- 42Manufacturer and retailer data from across the supply chain is a powerful source of information, providing a picture of product movement in near real-time. Access to the data is just the first step, because the volume of data available each day is overwhelming and much of it is noisy.
- 41November/December 2014 Big data needs advanced analytics, but analytics does not need big data. By Eric A. King Thanks big data! Now we’re even more data-rich … yet remain information-poor. After staggering investments motivated by an overabundance of buzz and hype, big data has yet to produce cases that reveal…