### Email Relevance: Weighting Transactions

When I work on email marketing projects, I like to assign weights to historical transactions.

I will create monthly or quarterly variables (i.e. demand spent 1 month ago as a variable, demand spent 2 months ago as a variable).  A dependent variable (next month spend or next quarter spend) is also created.

I run a regression analysis.  I document the coefficients for each variable.

In the graph above, I build a curve as a function of the coefficients for each recency variable.

Then, I compare the fitted function against the most recent variable I have in the analysis.  The most recent variable is assigned a weight of 100%, then each subsequent variable is assigned lesser weights (49%, 39%, 35%, 31%, etc.).

This yields the weighting scheme I use when deciding which version of an email marketing campaign to send to a customer.  If a customer spent \$100 on Womens merchandise last month, and \$250 on Mens merchandise 36 months ago, I have to make a decision whether the customer should receive a Womens message or a Mens message.
• Womens Weight = \$100 * 100% = \$100.00.
• Mens Weight = \$250 * 20% = \$50.00.
In this example, the customer spent more on Mens merchandise, historically, but the purchase is 36 months ago, and therefore, less relevant to future activity.  The customer should receive a Womens message.

Clearly, these strategies should be tested, because your mileage will vary.  Regardless, in email marketing, it is important to come up with a weighting scheme, so that the most relevant messages are sent to a customer.  I've yet to run across an instance, in twenty-four years, where we make a mistake by weighting older transactions as being less important than recent transactions.