About Customer Targeting
Customer targeting is the business process that defines which customers to market to. For each direct marketing campaign, be it email or direct mail, there is a decision to be made on who will, and who will not receive the campaign. Influencing factors related to that decision include:
- The type and cost of the media
- Past Customer Performance, most commonly referred to as Recency, Frequency, and Monetary – RFM for short.
The objective is to select the “best” collection of customers to mail a particular campaign to, among the entire inventory of available customers.
Often, since the marketing campaign is a fixed cost, the definition of best is defined as maximizing the return on marketing investment, which is often restated as maximizing revenue under fixed circ.
Customer Performance Detailed
To maximize revenue under fixed circ, we need to select the best subset of customers from all the customers we can mail to. We want to know who will, and who will not buy from us if we market to them, and if they buy how much they are likely to spend. Essentially, we want to sort the customers from most to least profitable.
Often we describe customer performance in simple shorthand, by using the hackneyed RFM acronym. Customer performance in reality is a much more detailed and complete view of each customer. It integrates well honed measures, such as: days since last purchase (Recency), the number of times the customer has purchased within given time ranges (Frequency), and revenue stemming from those purchases (Monetary), along with key business concepts. Examples of business concepts include:
- Seasonality and Holiday
- Club and loyalty programs
- Private Label Credit
- Overlay Data
- Demo and Psychographics
Examples of Measures include:
- Days since last event
- Lifetime Frequency
- 0-12 Month Frequency
- Lifetime Revenue
- 0-12 Month Revenue
- Lifetime Average Order Value (AOV)
- 0-12 Month AOV
These measures are created for each key business concept.
Consider a Women’s Apparel Direct Marketer, who has a yearly swim campaign that mails in the spring. For sake of simplicity, let’s limit the measures we will consider to Recency (days since last event), Frequency, and Average Order Value (AOV). For each business concept for which we have data, we develop a set of variables to be used in our analysis:
Business Concept Specific Concept Measure Variable Name Merchandising Swim Recency Swim_Recency Merchandising Swim Frequency Swim_Frequency Merchandising Swim AOV Swim_AOV Seasonality Spring Recency Spring_Recency Seasonality Spring Frequency Spring_Frequency Seasonality Spring AOV Spring_AOV Discount Deep Discount Recency Deep_Discount_Recency Discount Deep Discount Frequency Deep_Discount_Frequency Discount Deep Discount AOV Deep_Discount_AOV
In this simple example we defined three business concepts and three RFM measures: which results in 9 variables. If each business concept had multiple levels (Swim, Sport, Poolside,…), and we developed all sensible measures, the resulting set of variables can quickly grow into the thousands.
To maximize revenue, we need to rank order our customers from best to worst. There are a number of methods for doing this. The simplest is probably to just sort the file, based on marketing instinct. The issue with this approach is the order of the sort greatly changes the select, and we have no idea which sort maximizes revenue. We can do better by incorporating historical performance from already complete Swim Campaigns, by creating each of our variables as of the day before last years inhome for each customer mailed, and add response to that mailing to the customer performance dataset.
Typically we regard people who responded to the mailing as a 1, and people who did not respond as a 0, and an additional monetary measure defining how much she spent.
With the addition of the response variable(s) to customer performance, we can now start to understand which set of variables are most useful in separating buyers from non-buyers.
Once we have our variables aligned with customer responses, we can attempt to learn which customers are more likely to respond to our upcoming Swim Campaign. We might notice that people who have purchased Swim products before are much more likely to buy in the future. Within swim buyers we might notice that customers, who purchased in the last 12 months, outperform customers who purchased 13+ months ago.
By hunting around for customer groups that outperform other groups, we can group customers into homogeneous segments. By reviewing past performance for these groups, we can now sort from best performing to worst performing segment.
Tools of the Trade
Einstein famously quipped,” Everything should be made as simple as possible, but not simpler.” In our case we should continue to add complexity, as long as the added complexity is profitable, and operationally tractable.
There are several mathematical techniques which can be used to improve the sorting of our file:
- Decision Trees
- Regression Analysis
- Neural Networks
These techniques are ordered by complexity, but deliver the same output – the ability to rank order the file from best to worst customers for a given campaign.
Over the years we have found regression analysis to be the most controllable, most tractable, and most robust technique in ranking the file. In most cases we develop a three tier model:
- Propensity to buy
- Expected Order Size
- Expected Return on Marketing Investment
We use Logistic Regression for our propensity to buy model, Linear Regression for the expected order size model, and non-Linear Regression to combine the two prior models.
This appears to be a good balance between precision, adaptability to new concepts and reliable and robust operational attributes.
Developing a customer targeting business process can range from a simple segmentation, to a complicated neural network. The key is to provide a framework which reliably integrates key business concepts with logical measures into a structure that can dependably be used to maximize revenue under fixed circ.