## April 17, 2013

### Weighting

I can't stress the importance of weighting transactions.  So many of the tactical mistakes we make come down to over-estimating the importance of prior customer activity.

Example:  Customer spent \$100 via phone 5 years ago, spent \$100 online thirty months ago, and spent \$100 via a tablet today.

• Best Practice Thinking:  This is an "omnichannel" customer, one that does everything, one that should be marketed to via old-school and modern techniques.
The reality is that the transaction from 5 years ago has almost no meaning.  In many of my projects, there's a natural progression, one that goes something like this:
• 0-12 Months Ago --- Weight = 100%.
• 13-24 Months Ago --- Weight = 50%.
• 25-36 Months Ago --- Weight = 25%.
• 37-48 Months Ago --- Weight = 15%.
• 49+ Months Ago --- Weight = 10%.
Using this version of reality:
• Tablet Transactions = \$100 * 1.00 = \$100.
• Online Transaction = \$100 * 0.25 = \$25.
• Phone Transaction = \$100 * 0.10 = \$10.
From a weighted perspective, this customer is 74% tablet-focused.  Your marketing strategy for a customer 74% tablet focused is very different than your marketing strategy for a customer that is 33% tablet focused.

Thoughts?

#### 2 comments:

1. This makes total sense, but how would you adjust it based on the product? For instance, take a product like mattresses. How would you adjust the sliding scale based on the long period between purchases?

Also, does the channel from 5+ years ago deserve any credit? It seems to me like the revenue should be credited to the customer for analysis, but not the channel. Let's say the \$ figures were different. Let's say the 1st purchase was for \$500, the second purchase was for \$100, and the third purchase was for \$25. In that scenario, how would we look at that client?

2. You don't have to adjust the weighting - run regression models, they give you the weights to use. I run models that have a dependent variable (say demand in the past 'x' months), with independent variables being channels/products in time intervals (0-3 months ago, 4-6 months ago, 7-12 months ago, 13-24 months ago, 25-36 months ago etc). The coefficients of each independent variable will make logical, intuitive sense to you, they represent the weights you will use.

On the channel from 5+ years ago and the dollar amount ... the regression model tells you the answer, no reason for you to make that determination. In my models, a social media tweet has a half-life of four hours, a website visit has a half-life of a month ... an order from a retail store has a half-life of 18 months ... an e-commerce purchase has a half-life of 24 months ... an order from a catalog placed at a call center has a half-life of 36 months. On dollar amounts - yes, \$500 usually counts for five times a much as \$100 - but you can test this via log or square root transformations.

In other words, there's no need to guess or speculate. Your customer data is there, all you have to do is mine it and the answers are readily available to you!!

Note: Only a member of this blog may post a comment.

### Items per Order

Take a look at this graph. As the customer becomes more loyal, the customer (not shown here) orders more often. However ... however ... as t...