Customer Loyalty, Competition Analysis, Propensity to Defect and Ad Media Selection
Kenneth L. Reed1 and Joseph K. Berry2Presented at the GIS'99 Conference, Vancouver, British Columbia, March 1-4, 1999
The geographic location and distribution of customers is a critical piece of information that is usually missing from customer relationship marketing (CRM) and data mining applications. Often, data miners develop models for propensity to buy, attrition, risk assessment, and the like with no consideration of where the customer lives and works. It is common sense, however, that people in certain neighborhoods are more likely to take out home improvement loans. Other neighborhoods may be more risky for insurance. People tend to shop where it is convenient, which usually means close to home or work, hence travel time is important for retail response to promotion. These spatially related factors usually do not find their way into models because most businesses are only now starting to store geocoded customer data, and few realize what can be done with that information. The approaches discussed in this paper include spatial information, providing models that recognize spatial correlation and use spatial inference for increased power and utility. Specific examples illustrate the use of spatial modeling and analysis for understanding customer loyalty, assessing competitive threat, identifying customers likely to defect, and targeted print media promotion choices.Introduction
The geographic location and distribution of customers is a critical piece of information that is usually missing from customer relationship marketing (CRM) data mining applications. Often, data miners develop models for propensity to buy, attrition, risk assessment, and the like with no consideration of where the customer lives and works. But it is common sense that people in certain neighborhoods are more likely to take out home improvement loans. Other neighborhoods may be more risky for insurance. People tend to shop where it is convenient, which usually means close to home or work, hence travel time is important for retail response to promotion.
These spatially related factors usually do not find their way into models because most businesses are only now starting to store geocoded customer data, and few realize what can be done with that information. HYPERparallel's CRM approaches (see authors note at the end of this paper) include spatial information, providing models that recognize spatial correlation and use spatial inference for increased power and utility.
In this paper, we illustrate the use of spatial modeling and analysis for understanding customer loyalty, assessing competitive threat, identifying customers likely to defect, and targeted print media promotion choices.The Situation
The town is Smallville, whose most famous son is Superman, known to locals only as Clark Kent, son of Ma and Pa Kent. Pa Kent died of a heart attack, and loyal son Clark provided a retail outlet for Ma, called Kent's Emporium, founded in 1954. Kent's has become a landmark in Smallville, typifying a small town general merchandiser. As Smallville grew, Kent's grew with it and now enjoys a hefty market share. Kent's has been building and maintaining a customer and transaction data warehouse. They are contemplating the use of data mining to improve their customer relationships.
Now trouble looms. The evil Lex Luther is plotting to build a huge superstore on the outskirts of town: the feared Colossal Mart! Colossal Mart specializes in destroying small town retailers - providing goods at prices the locals cannot match, and selling them in fancy stores with ultra modern fixtures, state of the art planogramming, and aggressive promotion strategies. Lex gloats as he contemplates the destruction of the retail structure in Smallville, anticipating the time when proud former retailers will be working for him at box boy wages. What a pleasure, and totally legal, too! Now if he could only get rid of Superman!
In panic, Kent's turns to trusted retail consultants, who employ HYPERparallel data mining and spatial modeling recipes. Fortunately, the data warehouse contains enough information to get started. The consultants followed the Where's My Customer recipe developed by HYPERparallel, customized it to match the information contained in the Kent's data warehouse, and began to analyze the situation. Here are the steps they followed.Where's My Customer Recipe
A recipe is HYPERparallel's way of encapsulating knowledge discovery procedures and methodologies into a reusable framework that can be adapted to a business requirement. The Where's My Customer recipe provides methods for understanding the relationship between customer distribution and their behavior. The basic steps are outlined below.Define Customer Behavior
Behavior is defined specifically as measures pertinent to the business question and is a customizable part of the recipe. Often we define behavior in terms of recency, frequency, or monetary measures (RFM).Customer Behavior Variables
Recency measures can be defined as time since the last visit. For a grocer, time would be expressed in days. A general merchandiser or department store might express time in days or weeks. An auto dealer might express time in weeks, months or years depending on whether the behavior tracked is service visits or new car purchases. The consultants write a SQL script to compute recency and add that as a column in the customer table. Kent's analysts defined recency as days since last visit.
Frequency is a measure of the frequency of visitation, usually expressed in visits per unit time. The time frame is usually the same as that used in the recency calculation. It could also be expressed as a probability of visitation, derived from statistical analysis of the data. This is what Kent's did - they used frequency information combined with other factors to develop a Propensity of Visitation model using the HYPERparallel rule induction algorithm. The propensity model was built from historical customer purchase data and visitation history. The model gives each customer a score representing the probability of response to a promotional event. A separate model was built to give the probability of visitation without promotion. The results of these models were used as input variables to the spatial models as discussed below.
Monetary indicators can be a variety of measures: average transaction value, customer value, etc. Kent's analysts chose to define the monetary indicator as the total sales value per customer per year. All customers were then placed into monetary deciles and ranked.Segment Customers by Demographics Kent's then enriched the customer data with demographic information, and identified naturally occurring customer segments using //Cluster (pronounced HYPERparallel Cluster). They then analyzed the purchasing behavior of the segments in order to understand the relationship between behavior and customer demographics. The demographics of Kent's top 30 percentile in terms of the monetary measure are illustrated in Figure 1.
Given the customer segmentation above, Kent's can now analyze propensity of visitation scores for each group. The demographic clusters for the top 30 percentile customers are also active for the next 40 percentile of customers. Figure 2 presents a comparison of visitation scores for the clusters of top 30 percentile customers to the mid 40 percentile customers
In Figure 2, we see that the top 30 percentile in sales have greater propensity to visit than the next 40 percentile. For example, the "Older marrieds with 1 teen" segment has an average propensity score of 43% for the top 30 percentile bracket and a score of 25% for the next 40 monetary percentile. This gap of 18% is an opportunity measure - an index of a potential wallet share increase if we can devise incentives that work.
|Figure 1. Major demographic segments of the best Kents customers.|
Ordinary data mining approaches stop here. Kent's developed a propensity to visit scoring model and a segmentation of their customers. This is a useful start, but these models are not responsive to environmental change, such as the advent of new Colossal Mart, which will change all of the propensity scores. Nor are these models good for customer acquisition purposes by themselves, since only existing customers can be scored.
If we use only the analysis above, we cannot predict the effect of Colossal Mart, nor do the data give us much insight as to how to close the opportunity gap. We need answers to questions like the following:
|Figure 2. Comparison of propensity of visitation scores by rank and segment.|
How will Colossal Mart impact our top customers? Where are our good customers? Will they be close to Colossal Mart? Our best demographic segment is "Married with 2 to 4 Children." Where are they? If they are close to Colossal Mart, how many will we lose? We can advertise through several print media. Are there differences in effectiveness? Which paper would reach a higher percentage of our good customers? Which customers will we lose to Colossal Mart?
These questions can be addressed by spatial modeling. The spatial models will be used to assess the potential impact of Colossal Mart, to identify customers at risk, and to help develop a strategy to cope with the changes brought on by Colossal Mart.
Develop Spatial Models
It is well known that travel time from home or work to store is an important retail concept. All things being equal, a person will shop at the most convenient store. A major component of convenience is travel time. Neighborhood factors are also important. Socioeconomic and demographic factors are spatially correlated. The non-spatial models above assume that all the demographic factors are uniformly spread over Smallville. But we know that is not true. With a GIS and //Spatial (pronounced HYPERparallel Spatial) program, we can build models that account for these spatial realities. The recipe for building the spatial models is outlined below.
The spatial models are developed from geographic information. We used a popular desktop GIS (geographic information system) to house the spatial database. We geocoded each customer's address and the location of our store. We also included the location of the Colossal Mart and all the streets, streams, and lakes in the Smallville vicinity. Streams and lakes act as barriers to travel. A map of Smallville showing Kent's Emporium and the evil Colossal Mart is shown in Figure 3.
There are two basic approaches to computing travel time. A commercially available approach uses the line and polygon GIS format (vector format). The algorithm follows the roads out from the store, computing distance at intervals on the roads. These models sometimes include traffic densities and speed limits to compute travel time estimates. This approach is good, but can only compute along roads or map features.
|Figure 3. Base map of Smallville.|
HYPERparallel's approach is somewhat different. The base map is converted from a line and polygon map (vector format) into grid maps, similar to the bitmap rendering of Figure 3. The grid format supports powerful analytic operations that are very difficult or impossible on vector format maps. //Spatial operates only on grid maps, but which usually are created from GIS vector maps.
The base map in Figure 3 is converted to several grid maps: one containing only the roads, one containing the lakes and streams, and one for each store. The road grid map is encoded with relative speed factors. These factors are good enough for computing a travel time index applicable to any time of year or hour of day. Then we use these maps in //Spatial to develop a map of travel time. In this approach, travel time from any point on the map to the target, in this case Kent's Emporium or Colossal Mart, is computed. This creates a monotonically increasing surface, like a convoluted bowl, from the store to any point in the map. To compute travel time from any customer's residence or work place to the store, one simply has to identify the grid cell wherein the starting point is located, and look up the travel time value on the travelshed map.
|Figure 4. Travel-shed map for Kents Emporium|
The travel-time map for Kent's Emporium is shown in Figure 4. A similar map is computed for Colossal Mart. These maps are "sliced" into contours for display purposes. Each contour or color band represents approximately 5 minutes of travel time. Notice that the travel time contours are not circular or uniform. Some places close to the store as the crow flies are actually quite far away in terms of travel time. Notice also that the travel time model covers the entire map, not just along the roadways. It is possible to constrain the travel time operation to the roadways, but for many applications, an unconstrained travelshed is better.
In Figure 4, we can see that Colossal Mart is about 15 minutes away from Kent's Emporium (the purple zone includes area 10 to 15 minutes from Kent's).
A travel-shed map is used to develop a new column in the customer table, travel time to store. Kent's analysts added travel time to Kent's and travel time to Colossal Mart to each customer record.
These data elements are used to compute the next maps, Region of Influence, which is an essential requirement for assessment of the potential impact of Colossal Mart on Kent's business.
Region of Influence
Region of Influence is an extension of the trading area notion. It is defined as the probability of visitation from any grid cell to the target store. Probability of visitation is computed by aggregating all customer behavior information to the grid cell. That is, we average the behavior measures of each customer who lives within a particular cell to obtain a grid cell average. Each measure is stored as a separate grid map - one for average recency, average propensity to visit, average monetary, and so on.
It is important to remember that the region of influence is defined spatially, not by individual customer. Each grid cell needs to have a Region of Influence (R-of-I) value. A database table is created by //Spatial that contains the grid cell identifiers (usually row and column numbers), and all the aggregate behavior data, travel-time values, etc., from each of the grid maps in the spatial database. This database table is input to //Induction, which will compute the R-of-I score for each grid cell. The output of this model is then imported back into the //Spatial database, and after some spatial statistical analyses are performed, the region of influence map is produced.
The model is also run for Colossal Mart using the same customers to model the Colossal Mart travel-shed. Because Kent's analysts expect that Colossal Mart would have a stronger "pull", the individual customer's propensity to visit scores were adjusted up - reflecting the belief that people might be more willing to travel to Colossal Mart than to Kent's Emporium. The two Region of Influence Maps are compared in Figure 5.
The greater R-of-I for Colossal Mart is obvious. Kent's analysts notice that Colossal Mart's region of high and very high influence is much larger than Kent's and their total influence reach covers nearly the entire town.Using the Where's My Customer Recipe
The data and models produced in the recipe can now be used to better understand Kent's customer base. We can now begin to assess the potential damage from the evil strategies of Lex Luther. Even as the Colossal Mart ground breaking begins, Kent's analysts are planning their defense.
At the core of the defense is the customer. Clearly, Kent's Emporium must retain their loyal customers. We have already seen that if we define loyalty in terms of RFM, we have identified many loyal customers, now known by name and location. But RFM is not the only definition of loyalty. Consider the elderly widow, who loyally shops only at Kent's, but spends relatively little, and because her consumption rate is low, falls into the second Frequency tier. Is she not loyal, if not greatly profitable? How can we define a meaningful loyalty score? Which customers are likely to defect? Can we provide a propensity to defect score? How can we use that knowledge to our advantage?
What about targeted marketing? For which customers should we craft special promotions? Indeed, to which customers should we promote? Profitable customers only? Profitable customers with low propensity to defect? Profitable customers with medium propensity to defect scores? There are several opportunities to advertise selectively in Smallville. Would this be effective?
The Where's My Customer recipe also provides suggestions for the use of the newly created spatial database in answering the questions outlined above.Discovering At Risk Customers
Subtracting Kent's "R-of-I" map from Colossal Mart's R-of-I map (a //Spatial operation) gives an assessment of the impact on Kent's Emporium once Colossal Mart opens. While this assessment is based on simulation with the spatial models, the assumptions that went into the models are reasonable.
|Figure 5. Region of influence maps for Kents Emporium (left) and Colossal Mart (right).|
The results of the analysis are sobering (Figure 5). The influence of Colossal Mart has greatly reduced the overall extent of the Kent's Emporium R-of-I. Most of Kent's customers are at risk. Those who fall into Colossal Mart's High and Very High regions are probably lost. The battle zone between the two stores is clearly identifiable. Customers falling into this zone are "on the fence" and could perhaps be lured to Kent's by targeted promotions.
Kent's analysts immediately moved to identify those specific customers at risk. A new model is built, using the risk factor computed in the analysis shown in Figure 6. Each customer record in the database is scored by this risk factor, which is simply taken from the grid cell in which they live.
|Figure 6. Competition analysis model output. Positive number indicates regions where Colossal Mart dominates. Negative numbers favor Kents Emporium.|
The risk factor score ranges from 1 (very high probability of loss) to -1 (very high probability of retention). Kent's analysts ranked each customer with the risk score and were able to assess potential revenue loss. Fortunately, 70% of Kent's revenue came from customers living within the areas having a risk range of 0 to -1, shown in the battle zone, pink and red areas on the map.
Kent's might be willing to concede the brown and green areas (risk factor scores from 0.2 to 1.0) if they could devise a strategy for targeting the battle zone to increase visitation from that area. The also need to devise strategies to enhance the loyalty of their good customers in the pink and red areas (-0.4 to -1.0).
The Where's My Customer Recipe provided Kent's Emporium with a previously unobtainable understanding of their customer base. It supported a critical what-if analysis that was helpful in devising strategies for dealing with Colossal Mart's incursion. By adding the spatial dimension, they were able to go far beyond mere analysis of their customer and transaction database. They were able to build models.
Validation of the models and the analysis will of course have to wait until the Colossal Mart opens for business. Nevertheless, the framework for analysis exists.Advertising Trial
One suggested strategy was selective advertising, employing print media distributed only or primarily within the battle zone. While Kent's would continue to place ads in the Smallville Daily Rag, it was suggested that some of the neighborhood flyers might be a possibility worth investigating.
|Figure 7. Map showing distribution of various local print media where Kents Emporium might wish to advertise.|
A map was quickly produced that identified distribution of a number of neighborhood flyers with the battle zone superimposed on the map (Figure 7).
The P&S Tidings paper looked to be a good choice for a trial advertisement. Its distribution is clearly in the battle zone. The question is whether an advertisement in the P&S flyer would yield measurable results.
Kent's purchased an advertisement in the P&S Tidings. Sales responses were compared to average customer response to previous promotions. After the promotion, the results were tallied and mapped. The sales results were summed up by map grid cell and spatially averaged in //Spatial. Before and after comparisons of total sales are shown in the expanded boxes of Figure 8. The differences are subtle but significant. The promotion yielded an overall increase of 17%.
The response to the promotion can be seen in the enlarged areas of Figure 8. The differences are subtle, but significant. Sales increased (shown largely in the green areas of the enlargement). The distribution of the sales figures tended to follow travel-shed properties: in essence, the promotion tended to provide some increased incentive for people to travel to Kent's.
|Figure 8. Maps of sales densities before and after the test promotion in the P&S Tidings.|
This was extremely encouraging. A battle zone medium was chosen, and a selective effect was observed. Trials were immediately designed to test the effectiveness of other battle zone media.
Kent's Emporium used knowledge discovery technology in combination with geographic information and spatial modeling to set up a sophisticated analysts workbench. The purpose of the workbench is to better understand their customers. They developed the following models:
1) a customer loyalty segmentation using demographic, geographic and sales transaction data;
2) a propensity to travel model that helped to explain the geographic distribution of their customer base;
3) a propensity to defect model based on the known distribution of current customer behavior;
4) an a priori competition analysis using spatial modeling to help predict the effect of competition on their business; and,
5) a targeted media model that allowed evaluation of the impact of promotions in selected media.
Plans call for the development of future models of:
1) a propensity to respond to promotion both all-store and specific theme promotions;
2) a propensity to shop within strategic departments segment and locate the customers that do only a portion of their shopping at Kent's;
3) a propensity to cross-sell and up-sell models; and,
4) a customer lifetime value models.
But that is another story.
Authors Notes: Since the writing of this
paper, HyperParallel has been purchased and outside consulting services are longer
supported. The procedures described in the paper are under consideration by the Customer
Insight Practice, Andersen Consulting.
A slide set containing color versions of the figures in this paper plus other graphics describing the approach in more detail is available on the Internet at www.innovativegis.com/basis, select Papers and Presentations, "Spatial Modeling and Data Mining in Retail" paper.
1Kenneth L. Reed, Consultant, Customer Insight Practice, Andersen Consulting
2Joseph K. Berry, President, Berry & Associates // Spatial Information Systems, Inc.