Kevin Hillstrom: MineThatData

Exploring How Customers Interact With Advertising, Products, Brands, and Channels, using Multichannel Forensics.

November 17, 2007

Modified RFM For E-Mail Targeting

RFM is great for targeting one catalog to one customer. However, RFM is tough to manage in a multichannel environment.

This becomes clear in e-mail targeting. Say you have a Mens version of an e-mail campaign, and a Womens version of an e-mail campaign --- a customer could receive either version on the same date. Use this customer as an example:
  • Customer spent $100 on Mens merchandise in the past three months.
  • This customer also spent $200 on Womens merchandise 7-12 months ago, and spent $100 on Womens merchandise 13-24 months ago.
Which version of the e-mail campaign do you send to a customer? You could use RFM --- your customer is a 0-3 month $100 mens buyer, and is simultaneously a 7-12 month $300 womens buyer. Which "segment" carries more "weight".

This is where we apply "Modified RFM".

Have your statistician build a regression model one time --- and use the "weights" or "coefficients" for your modified RFM scheme. I realize this is statistical blasphemy, however, we aren't managing clinical trials for cancer drugs, we're deciding which version of an e-mail campaign a customer receives.

Step 1: Pick a "dependent" variable for "Mens". I like to look at the past twelve months.

Step 2: Create a series of "independent" variables:
  • Dollars spent on Mens in past three months (prior to the dependent time period).
  • Dollars spent on Mens 4-6 months ago (prior to the dependent time period).
  • Dollars spent on Mens 7-12 months ago.
  • Dollars spent on Mens 13-24 months ago.
  • Dollars spent on Mens 25+ months ago.
Step 3: Regress these five variables against your dependent variable. The "coefficients" become "weights" for e-mail targeting, as you'll see soon.

Step 4: Repeat Steps 1-3 for Womens merchandise.

Now, we can evaluate which version of an e-mail campaign a customer should receive. Let's look at our example:

E-Mail Targeting Strategy: Mens Weights





Spend Factor Weight
00 to 03 Months $100.00 1.600 160.0
04 to 06 Months $0.00 0.600 0.0
07 to 12 Months $0.00 0.300 0.0
13 to 24 Months $0.00 0.150 0.0
25 to 99 Months $0.00 0.050 0.0
Total Weight

160.0

E-Mail Targeting Strategy: Womens Weights





Spend Factor Weight
00 to 03 Months $0.00 1.600 0.0
04 to 06 Months $0.00 0.600 0.0
07 to 12 Months $200.00 0.300 60.0
13 to 24 Months $100.00 0.150 15.0
25 to 99 Months $0.00 0.050 0.0
Total Weight

75.0

For the Mens version of the e-mail campaign, the customer receives a "weight" of 160.

For the Womens version of the e-mail campaign, the customer receives a "weight" of 75.

So, you should send the customer the Mens version of the e-mail.

For your next campaign, you don't have to build models again --- remember, we're not trying to cure cancer, we're just figuring out which version of an e-mail campaign will improve response a bit. Just apply the same weights built in your prior modeling process, and decide who gets which version.

The key here is to not build separate RFM schemes. Instead, you build variables in your database that summarize purchases by 0-3 month, 4-6 month, 7-12 month, 13-24 month, and 25+ month time periods. Then you "weight" those purchases based on importance. This gives you a good targeting strategy.

Statistical purists will blast me for misuse of appropriate statistical techniques. That's fine. We're just trying improve e-mail marketing performance, while minimizing use of internal resources, or minimize expense incurred when hiring consulting statisticians. This gets you 80% of the benefit for about 5% of the work.

Labels: ,

5 Comments:

At 10:32 AM , Anonymous Ray said...

As a statistician, I do cringe a little when reading this post. Nonetheless, your point about cancer trials was spot-on (and it made me lol.)

One suggestion, consider standardizing the coefficients of your model output to control for potential differing magnitude and variability in men's / women's spend $. The 'weighted' values will be more resistant to the raw magnitude of the data (losing interpretability is clearly not important.)

I wonder if the data coops out there with product level data will offer this type of service someday (leveraging data across many cataloguers)?

 
At 10:55 AM , Blogger Kevin Hillstrom said...

Yup, you can standardize coefficients per your recommendation!

Co-ops should offer this service, they already use this information to figure out whether a customer receives a catalog.

 
At 12:32 AM , Anonymous Abhishek said...

Would it be possible for you to point us to a mass e-mail distribution software that allows for RFM-based targeting?

 
At 6:48 AM , Blogger Kevin said...

You could work with any of the big e-mail vendors (i.e. CheetahMail or Responsys), and they could help you with this.

 
At 8:32 AM , Anonymous Abhishek said...

Thanks Kevin!

 

Post a Comment

Links to this post:

Create a Link

<< Home