The outbound marketing landscape, especially in the retail industry, is a challenging one.

Customers are receiving multiple communications from multiple businesses, and for marketers it’s sometimes hard to understand which activity is driving results. Many retailers have a complex multichannel strategy of internet, email, direct mail, telemarketing, and catalogues. These are aiming to increase product sales.

In a nutshell, the problems are:

• Which activities should I do more/less of?
• Who are my most valuable customers?
• What can I do to increase sales or reduce costs?

The problem is there are too many customer communications to model individually. And there is too much cannibalisation and too many halo effects involved that reduce the accuracy of predictive models.

Attribution: making simple decisions in a complex environment

We investigated an innovative solution with a UK-based retailer to solve the problem. Usually, predictive models use customer data to predict a target variable, such as sales.

But what if you used the marketing activities alone to predict the sales?

For example, if we know that people who received a particular email purchase more than people who did not receive the email, it becomes possible to build a picture of the incremental value of that particular email. If you extend that to other marketing communications, you can start to build a picture of the incremental value of each marketing communications.

Using a decision-tree approach, a picture like this one can be created:

Using industry expertise and analytical acumen, marketers can start to identify opportunities. The tree above tells you that receiving more than two catalogues provides no benefit (in terms of increasing sales). You can attribute an increase of 50 (150 – 100) to sending the second catalogue. So, the marketing decision here would be: Stop at 2 catalogues.

Similarly, one can attribute a value of 60 (180 – 120) to sending an email and 100 to making a phone call (230 – 130). But don’t forget this is in the instance where two catalogues and an email have already been sent (so the halo effect of these was considered in the scoring). The marketing decision here would be: only follow up customers that have received two catalogues and an email.

This tree identifies the customer journeys that are more likely to be successful. The next step would be to apply a customer-level estimate – that is, building a predictive model for each end-node of this tree to identify and optimise the customer journey for all customers.

New techniques for attribution coming soon

You can see that attribution covers a lot of territory, but traditional analytics can go a long way to solving a variety of attribution-related problems. Newer methods can be used to help identify the next step on a path, and there are some exciting developments coming from SAS that build on these traditional and newer techniques to truly drive an optimised customer journey.

The first step for us on our journey was to build a more analytical approach to attribute value to the source of traffic to a website domain. The traditional rules based approaches work well (first click, last click etc.) but there is potential to understand behaviour more with a more analytical approach. Below is an example of how this capability is presented to the marketer, and I’ll be sharing more details on this soon.