How Delivery Hero Beat Purchase Churn in the UK

One of the UK’s most popular food ordering apps leveraged predictive analytics to forecast and preempt purchase churn.


About Delivery Hero and hungryhouse


With more than 10,000 restaurants on their platform, Delivery Hero’s UK business hungryhouse is the UK`s premier online platform for takeaway and food delivery. From Birmingham to Glasgow, and from London to Manchester, hungryhouse partners with the finest takeaway restaurants to provide the best experience for finding and ordering delicious takeaway in minutes.

As hungryhouse has ramped up to millions of mobile users, they’ve been forced to become more focused in using data to drive their customer engagement strategy. As Erik Kubik, marketing executive at Delivery Hero related to us:

“The wealth of user data we have is so vast, it can be hard to find the little golden nuggets of info you might not think to proactively seek out. By leveraging predictive analytics, we’ve been able to serendipitously drive these insights at a scale that we can use to automate our marketing communications.”

From our very first conversation with Erik and his team, the opportunity to plug the inevitable retention leaks that result from running an ecommerce business was clear. The churn problem cannot easily be tackled through traditional descriptive analytics; once you’ve observed users to churn it’s already too late to rescue them. And in the context of a non-contractual product and service like mobile food ordering, churn can be even more challenging to identify and and beat back.


Asking the Right Questions


At Localytics we often hear that user churn is one of our customers’ biggest challenges. But when asked how they specifically define churn, we’ve found that the actual metric used to measure churn can vary widely.

For hungryhouse, a user is considered to have churned when they have not completed a food order in 30 consecutive days. This means that only ‘active’ users who have placed an order in the past 30 days are eligible to churn, while any users who had placed their last order over 30 days ago have already churned. It also means that active users may have ‘churned and returned’ -- meaning they might have experienced some 30-day period without placing an order before coming back and being active in the past 30 days.

With a specific definition of churn in place, we can now go about answering the questions:

  • Who is unlikely to complete an order in the next 30 days?
  • Why are those users unlikely to order?
      • Are there any behaviors which indicate that a user is unlikely to order?
  • Are there any user attributes which indicate the same?


Getting to Meaningful Answers


Once the challenge of asking the right questions and defining churn has been resolved, going about finding the answers is much easier. Using Localytics, Erik and his team were able to easily create a new prediction in their dashboard the reflected their granular, custom churn definition:


Using a unique, two-step process that combines statistical modeling and decision trees, we were able to segment all of hungryhouse’s current active users into groups with different likelihoods of churning: High, Medium, and Low. This provides us an answer to our first question of Who is unlikely to order in the next 30 days.


We grouped current active users into one of three churn likelihood segments: High, Medium, and Low.

Answering this first question was a boon to Erik and his marketing team, who were able to use this new layer of segmentation to offer up more targeted, relevant, and contextual communications. Perhaps as importantly, however, was answering Why users are likely to churn -- and specifically what if any behavioral or other indicators there might be to identify users’ churn likelihood.

By scanning hungryhouse’s users’ historical data and calculating the churn rate for each group of users that performed each event a unique number of times in each of the first 1/3/7/14/28 days of a user’s lifetime, we were able to surface the early user behaviors that resulted in the greatest decrease in churn. Some of these insights included:

  • Users who performed 1+ Checkout Success events in their first 14 days were 16% less likely to churn; and
  • Users who performed 3+ Restaurant Tapped events in their first 3 days were 13% less likely to churn.

These new insights armed the hungryhouse team with new information that they can use to drive their communications strategy, both with existing active users but also with new users as they push to onboard them through these key gateways to retention.


From Insight to Action


The marketing team at hungryhouse decided right away that they wanted to be able to leverage these new insights at scale. Instead of setting up a one-time blast campaign to their High churn risk segment, they opted for an approach to design an evergreen, preemptive churn campaign  that would run in the background and recurringly catch users as they were flagged for a certain level of churn risk.

Specifically, the team setup a recurring push campaign that would run daily at 5pm (GMT), but limited to only allow newly flagged High risk users to qualify for one message.


Thinking deeply and empathizing with their users, the hungryhouse marketing team created two different messaging copy variations that they then A/B tested in the campaign. The first variation highlighted to at-risk users how easy it was to order food delivery through the hungryhouse app, while the second variation offered a 20% discount on a new order:



Evaluating Campaign Performance


Working together, Localytics Data Science team and hungryhouse took an experimental approach extending beyond the A/B test by setting aside a control group of 10% of randomly sampled High risk users. By setting this control group aside and hiding them from the campaigns, we were able to isolate the effect on purchase retention in the experimental group that resulted exposure to an intervention:


Thanks to the many monthly active users that hungryhouse attracts to their app, gaining enough statistical power to achieve a significant result was quick and easy. The experimental group was retained at a higher rate, representing an 11% increase in purchase retention over the control group.


The Business Impact


Given an average revenue per transaction of roughly $28 for the month of October, the 11% increase in purchase retention from the test directly saved hungryhouse over $11,000 in what would have otherwise been churned revenue.

The commercial value of this campaign is only partially explained, however, by the directly rescued revenue. Most ecommerce businesses realize downstream benefits of retention worth multiples of the original retained purchase value. This is precisely the reason why, for example, eBay offers its’ affiliates 200% commission on purchases sent from reactivated buyers.

In the end, the greatest benefit came from the additional insight predictive analytics afforded the hungryhouse team when planning their retention strategy:

“For some time, we’ve been wondering what key factors lead to churn besides the obvious. Being able to see and fully understand these factors lets us plan effective campaigns without leaning on our business intelligence team to crunch numbers.”

--Erik Kubik, Marketing Executive @ Delivery Hero (hungryhouse)

With this success, Erik and the team plan to continue to test different approaches towards rescuing High churn risk users and moving them towards greater engagement and retention.