PAW: Predictive Modeling for E-Mail Marketing

Live from Predictive Analytics World

Arthur Hughes (author of Strategic Database Marketing) and Anna Lu of e-Dialog.com presented on predictive modeling for e-mail marketing. Arthur has been developing databases for database marketing for 30 years or so. Initially he focused on databases but found that people could not use them to make money and that led him to his focus on database marketing.

E-mail marketing is often considered as only a way to drive online sales but it also drives offline sales. E-mail is not recognized as the potential marketing powerhouse it is. E-mail produces far better returns than other direct mail and 4x the offline sales than online but remains siloed as part of online sales. In addition, data shows that multi-channel customers are more profitable than single channel customers and e-mail helps with all channels. Yet too many e-mail marketing budgets are departments are just lost among the online marketing department.

Predictive modeling does not get used in emails because of the cost difference – where direct mail is $600 or more CPM, email is $8 CPM and so the value of analytics seems lower. However, email open rates are down to 16% and a recession will s…

Live from Predictive Analytics World

Arthur Hughes (author of Strategic Database Marketing) and Anna Lu of e-Dialog.com presented on predictive modeling for e-mail marketing. Arthur has been developing databases for database marketing for 30 years or so. Initially he focused on databases but found that people could not use them to make money and that led him to his focus on database marketing.

E-mail marketing is often considered as only a way to drive online sales but it also drives offline sales. E-mail is not recognized as the potential marketing powerhouse it is. E-mail produces far better returns than other direct mail and 4x the offline sales than online but remains siloed as part of online sales. In addition, data shows that multi-channel customers are more profitable than single channel customers and e-mail helps with all channels. Yet too many e-mail marketing budgets are departments are just lost among the online marketing department.

Predictive modeling does not get used in emails because of the cost difference – where direct mail is $600 or more CPM, email is $8 CPM and so the value of analytics seems lower. However, email open rates are down to 16% and a recession will simply cause still more email to be used (because its cheap) so open rates will continue to fall and subscription rates will do likewise – and unsubcribing costs money. E-mail marketing must get smarter to avoid this fate.

Anna then introduced a case study – a frequent flyer program sending out semi-monthly emails. Opt-out members were about 18% but accounting for nearly 30% of revenue – they were better customers. Furthermore these were mostly recent unsubscriptions.

To solve the problem they built a model using CHAID. They had all the member data plus all the email behavior information like opens, clicks (and what kind of link they clicked). They found some good predictors among the attributes. Next they looked at the revenue from the members and found that the top 10% generated 67% of revenue with the top 20% generating 84%. Clearly retaining the top 20% was going to be key while the bottom 50% contributed almost nothing.

They developed a risk-revenue matrix, mapping the lifetime value and the likelihood to defect. Clearly the ones who are valuable and likely to leave are the highest priority. Now these folks can be targeted. When they did this with Cingular in the past they were able to reduce churn by 26% and retaining millions of dollars.

The second example was an e-tailer, specifically an off-price name-brand retailer. Email is their single largest channel and their most important retention tool – linked to 40% of their revenue. Have lots of data like attrition, sales, seasonal purchases, departments shopped, number and value of items purchased etc. 50% of their revenue comes from their loyalty club (that has a fee) even though it only had 1.8% of the email list as members. So how to find potential loyalty club members in the other 98.2%?

Used logistic regression to build a model based on about 10 predictor variables such as lifetime purchases, email source (internal or third party), opens, clicks etc. For instance, total purchases and months since first purchase contributed positively but more recent email acquisitions were better. With this information could afford to spend money getting these high-potential folks to join the loyalty program. Of course you also do some testing with a random sample to see how much better the results are than random.

Besides the improved positive results, you need to value the lower unsubscribe rates. For many companies a subscriber has a value of $5-$15 – they must be replaced and this can cost say $14! If a mailing not only has less positive effects but also drives people away (because the customer feels they are not understood by the company) then the cost of replacing the subscribers could exceed the value gained making the overall email campaign a net value destroyer! As I like to say, people respond to decision (emails in this case) as though they are personal and deliberate so make sure they are!

Arthur is a great presenter and this presentation was terrific.

More posts and a white paper on predictive analytics and decision management at decisionmanagementsolutions.com/paw