When a marketing analyst begins the quest to build a predictive model for a marketing campaign using past data to forecast future actions, it should be simple, right?
Use past data to predict future actions. It’s science.
But this is where the art comes in -- the past data you choose to utilize will impact your predictions. Think of your customers as Crayons, how do you sort through the box of 64 to know which crayons to use?
Your first instinct may be to use everyone who has ever purchased/subscribed/signed up for your product or service. That makes sense, because now you have a representation of every type of person that wants your product. That might look something like this:
However, there is a pitfall in doing so. How was that customer acquired in the past? Were some given a free month trial but others not? How do you know if it was the trial that got them or if they really wanted your product long-term?
How to avoid the pitfall
Consider how you intend to market your product for the next campaign for which this target model is being built. Are you planning on offering a free month trial this time? If so, you may want to build your model only on people from the past who signed up under a free month trial offer.
But what about the people who wanted the product regardless of the offer? We call them the “free-riders”. Why wouldn’t you want to include them too? Consider how including the “free riders” would impact your Cost of Acquisition (COA); essentially, you would be giving something away that you didn’t have to because they were going to sign up anyway. If you can afford it, fine, but if you need to keep your COA low, you might want to focus your past data selection on customers who typically sign up using free trials to build your model. However, don’t miss an opportunity to then market to a smaller random sample of people who “look like” your free-riders and test other messages or offers to market your program. With recognizable stars in your data and a few sprinkles of customers to learn from, your data now looks something like this:
What to think about before emptying the “past data” box:
- Start with your current marketing initiative; what is the offer? In the past, was a respondent’s acceptance of the offer due to factors related to their demographic, or was it more heavily due to the attractiveness of the offer itself?
- Test your hypothesis. For example: people who signed up for the program in the past with a free offer look different than people who paid up front. Test this hypothesis, to know if it is true or not, on the past data before making a selection of which data to use in your model.
- Finally, do not limit marketing your campaign to only people chosen from your predictive model. Throw in some random consumers to learn new marketing messages or offers that work for new groups. This will make you smarter for your next campaign!
In summary, there is no “right answer” for which past data you use to build your predictive model, but there are some wrong ones. The art of sub-setting your past data to be representative of your future market will position you for a campaign success worth hanging on your refrigerator!