Aftenposten Subscription Purchase Prediction
Overview of this campaign
Growing digital subscriptions is a strategic priority for Aftenposten and Schibsted’s publishers. Knowing if a user is ready to buy or not can help us in making a more relevant news experience and in running more efficient campaigns on our own sites and in other channels.
The main objectives of the purchase prediction project has been to understand how user behavior on the site relates to purchasing and to develop a prediction model to identify news readers likely to buy a subscription.
A key goal of the project has been to develop the prediction model in a site-agnostic way so as it can be scaled to all publishers in Schibsted easily.
Results for this campaign
During this project we used two main data sources covering our logged-in user base:
- Behavioural information extracted from clickstream data that users generate as they browse Aftenposten
- Data from Aftenposten CRM system on which users have purchased or churned subscription ownership, and when
By combining these datasets we can understand the differences in behaviour of users who do purchase versus those that don’t purchase a subscription, and then to create a model that can predict each users’ chance of purchasing.
- Users who have visited recently, read more articles, visit more days, visit from multiple devices, visit at weekends and come to Aftenposten from multiple sources are more likely to buy a subscription
- We used a random forest algorithm to predict purchasing with performance of 78% AUC
- This model has been automated so as scores are generated on a weekly basis. The output is a score for each user denoting their chance of purchasing a subscription and can be used to prioritise users for targeting
- The scores were used in a live user-facing experiment via Facebook targeting by showing users a paid article in their Facebook feed
- By the end of the campaign period, the target group had 22% higher conversion than the control group, with statistical significance.
- A second Facebook campaign promoted a good offer in the Facebook news feed of both high and low score users. High score users had 35% lower cost-per-order than lower score users in the target group
Other experiments are being planned for, such as an email campaign to high-propensity users and also use of the score within the Aftenposten product itself to customise content recommendations and the user experience based on their propensity score.