Predictive analytics have proved a powerful tool for overcoming many of the pitfalls of traditional customer segmentation efforts. By defining the characteristics of segments or groups, and then predicting the value of those customers, marketers have a tremendous opportunity to focus limited marketing resources to the customers with the largest strategic and return on investment benefits.

This white paper explores the topic of predictive and adaptive customer segmentation in three parts:

• The Case for a New Approach to Customer Segmentation
• Building a Predictive and Adaptive Segmentation Model
• A Case Study Demonstrating the Power of Predictive Analytics

 

It is no secret that retailers have more data about their customers than ever before. Purchase histories are easily accessible. Demographic and descriptive data are also obtainable. This wealth of available data should enable companies to more accurately cross-sell and up-sell their customers. But it doesn’t always work out that way. A recent survey by Bain & Company found that only 25% of companies believed they were using such data effectively.

More and better data should also lead to improved customer segmentation and more persuasive communications with customers. Powerful customer segmentations can serve as a blueprint for everything from product development to organizational design. But this isn’t happening. Technology, budget, and cultural obstacles are significant contributors to the failure of many segmentation efforts. However, even when there is strong leadership and appetite to address these structural issues, segmentations often fail to achieve ROI expectations. Why? It is because most customer segmentations are backward looking and static. In other words, they use past data, assuming it will have the same impact regardless of subsequent changes in the marketplace or the customer.

 

On the surface, traditional marketing segmentation sounds straightforward. Bucket customers into unique groups, and make the people in each group as similar as possible. Because the members of each group have the same needs, wants, and values, each person in a segment is supposed to feel like they are being spoken to as an individual. From an ROI perspective, low value segments get low cost marketing communications while the higher value customers get more marketing dollars dedicated to them.

As straightforward as they sound, these conventional segmentations fail to accurately account for future changes in customer behavior and desires. With details about price, product, and place readily accessible on the internet, consumers no longer limit their purchase decisions based on information pushed to them by marketers or what they have done in the past. They know they have options and maximize them. Coupled with the fragile economic recovery, the future behavior of consumers is less connected to the past than ever before. Segmenting based solely on past behavior to steer a future marketing strategy doesn’t work. It is like driving a car by only looking in the rearview mirror.

What conventional segmentation does do well is tell us that not all customers are the same. At the same time, scarce resources mandate that we focus on those customers who are inclined to respond. Combining these two concepts, success is contingent upon aligning the cost of promotions with the value of the customer.

 

One way to do this is to use past data to build a flexible predictive framework that tells how a customer is likely to behave in the future. Typically, this involves a) the use of factor analysis to establish key drivers of behavior and, b) two nonlinear out-of-time regression equations – one to eliminate non-buyers and the second to estimate buyer value. A highly predictive model will have a correlation of .8 or more. And, a highly predictive segmentation guides marketing efforts that are both measureable and profitable.

 

Less complicated than building a predictive model is making it adaptive. Applying an actionable segmentation enables tracking how customers change segments each month, week, or day. Adaptive segmentations regularly re-score and re-segment customers based on future value. The results should enable marketers to speak to individual changes and truly communicate one-to-one with their customers. While acknowledging that there are technology and personnel constraints to how adaptive a model can be, failure to build an adaptive feedback cycle is one of the primary causes of unmet ROI objectives.

 

There are five fundamental steps to building a forward-looking, value-based segmentation model.

1. Identify the Key Drivers of Customer Behavior
   One way to do this is to use factor analysis to determine which attributes load on the same factor(s) as the outcome being measured. For example, key drivers to purchases might include:
   • Available income after rent/house payment, insurance, food, and medical costs
   • Need for product: for example, the more generations in the household, the greater the need for a retailer’s products
   • A fraud score to exclude fraudulent buyers from the analysis
   • Past purchase behavior: holiday only, once a year not holiday (i.e., presents for a birthday), regular buyer, binge buyer, and a category for those who buy a lot but return most of it

   When initially building this type of descriptive model of customers, hold out a control group to evaluate segmentation and campaign success.

2. Extract Past Data
   The next step is to gather data from the past to begin building a predictive model. The time span selected for extraction is generally one to two years, but this depends on factors such as data availability, sample size, and the timing of major changes in a product offering. The data must span two time periods to allow for comparisons between predicted and actual customer behavior.
3. Build the Model
   Using data from the earlier period, construct a model that predicts the future value of each customer. Then, using the later data, assess how close the prediction is to the actual results. Revise the model until it is a good predictor of future value. A correlation of .8 or greater is very good. When the model is complete, apply the same method to predict future value from the currently available data.
4. Reassess, Rescore, and Track
   Households should be re-scored with the model periodically to identify changes in their unique financial, familial, and purchase behavior circumstances. Some examples of changes that might affect scores include a sudden slowdown in purchasing behavior, unemployment, or an addition of a new child to the household. It is also important to track changes in customer scores and segments to understand how the needs of individual customers and the value of your customer portfolio are evolving.
5. Create Overlays for Each Segment
   The final step in this process is to create slices or overlays for each segment to add business value. An example of this might be looking at the average time between purchases for each segment. Then each customer’s time between purchases can be compared to the segment average to understand when to promote them or their risk for defection to a different store.

 

 

Let’s take a closer look at predictive analytics in action. We’ll start with a large national department store deciding how to fundamentally restructure their marketing spend. Up until this analysis, personal communications consisted of direct mail catalogs and advertisements. Other forms of traditional print media and broadcast advertising were also part of the mix. Direct marketing budget was allocated by tiers created from just two factors: recency and dollars spent. Unfortunately, this marketing mix proved particularly poor at identifying and retaining those customers at risk of defection. Because communications automatically decreased when a particular customer went a certain period of time without purchase, no customized marketing intervention was planned or made.

In this example, a model was developed using past data and demographics. First, data from 2009 was used to predict the 2010 segment value for each of five segments that came from examining the key drivers of purchase: Big Spenders, Average Joes and Janes, Big Families, Low Spend Customers, and Annual Buyers.

Next, the actual values from 2010 were compared to the model’s predicted values. When satisfied that the model would predict future value, the same model and methods were applied to predict 2011 value from the 2010 data. Plotted on two dimensions, with the area of each segment’s circle sized for value, the predicted value for each segment is shown in the following chart.

 

Within each segment, algorithms were applied to customer buying data to enhance characterizations and to predict buying behavior and value. In this case, the stratification identified four sub-sets within the Big Spender segment: Loyal/Accelerating, Loyal/Steady, At Risk, and Gone/Past Average Time – each with varying numbers of customers.

The following chart shows the slices and number of Big Spender customers that were identified within each of the sub-segments.

 

Based on these results, future marketing strategies were formulated. In this case, the executive team decided that the Accelerating and Steady Loyal sub-segments (sometimes called micro-segments by analysts) perform well and didn’t require anything special, except to let them know that their business was appreciated.

On the other hand, intervention was clearly required for the At Risk, and potentially required for the Gone segment. The At Risk segment was especially critical because if they were to leave they would likely shop at a competitor. While the Gone segment makes up a large group, they typically had a low response rate, and any contact with them would have to be inexpensive to keep a positive campaign ROI.

 

For the At Risk group of Big Spenders, the ROI-based marketing plan looked like the following:

Messaging

• Best place for you to shop
• We meet your needs
• We have competitive discounts

Contact Strategy

• Twelve catalogs
• Six special event notifications
• Six sale flyers
• Six special discounts delivered both as mails and emails

 

In this case study, the cost of this type of plan, per customer, was approximately $32 in outbound communications and $22 in lost margin for a total cost of $54 in marketing related expenses.

However, the sales per customer were $1,554 with a 12% margin, and the net ROI per customer was $132. Furthermore, remember that these are the highly profitable customers that are most at risk for defection or attrition.

 

Predictive and adaptive segmentations work because marketing dollars and organizational efforts go towards the strategic, profitable, customers -- and the cost of communication is aligned with the value of those customers. In addition, with rich customer profiles, marketers are able to craft compelling messages that are both calibrated in frequency and targeted to a customer’s phase in the buying cycle. Ultimately, all of this leads to the right message being delivered to the right customer at the right time, which leads to maximized marketing ROI.

For nearly 20 years, Catalysis has specialized in the digital integration of award-winning marketing campaigns that drive connected, measurable results. Our clients include Microsoft, Moss Adams, Banner Mattress, Thunder Valley Casino, BabyLegs, and WineBid.

For more information, contact info@catalysis.com or visit our website at www.catalysis.com.

 

The information contained in this publication is general and is for informational purposes only. Catalysis makes no warranties, express or implied, in this material.