Customer Lifetime Value
The More You Know: How More Variety of Data Increased Accuracy in Predictions
We found a way to increase the average accuracy on predictions of Customer Lifetime Value (CLV) by 15.2% and churn prediction by 13.4%. The secret is in how much and what types of data you can draw on. Here’s a thought experiment to illustrate the idea behind it:
Pick someone in your life — a co-worker, an acquaintance, maybe a friend’s favorite pet. Now imagine that you only know their online purchases from a certain brand, and nothing else.
Does that give you a very clear picture of who they are, how much they’ll spend with that brand in the future, and if they’re likely to stop shopping with that brand anytime soon? You might be able to make a decent guess, but there’s a lot of room for error.
Now imagine you had a digital crystal ball (or advanced software) and knew everything they had done with that brand. What they browsed, what emails they opened, their nearest brick and mortar store and their purchases there, if they used the brand credit card, if they were in the loyalty program and how much they engaged, and more. The accuracy of your predictions would skyrocket, and better still, you’d be able to intercede, if you felt so inclined, in a more timely fashion (for example, by sharing a 50% off coupon for squirrel-shaped chew toys for your friend’s dog, who doesn’t own one, but clearly should, to help keep her from churning).
This is just what our team of data scientists were testing in their newest research paper: “Predicting Customer Lifetime Value with Unified Data”. They compared predictions for customer lifetime value and the likelihood of churn using only online transactions and then again using a host of different data types. The results were pretty staggering.
That’s where they found the aforementioned 15.2% improvement in CLV prediction and a 13.4% improvement in churn prediction accuracy. Was the algorithm better or more highly tuned? No. It was just more complete data. As an analyst or data scientist, what data is fueling your models and predictions? To learn more about the research, check out the paper here.