Our Machine Learning Platform 2.0 is now live with GPS visitation data of 150 million active users! Now, you can not only understand your customer’s commercial footprint (as introduced in our previous post) but more importantly, predict and target those who are most likely to visit your stores and/or get them to download your app.
Need proof? Here’s a case study for Dunkin Donuts that demonstrates how well we can service the QSR industry thanks to the predictive power of our data. None of this data or insights are proprietary to Dunkin; all of it was drawn from our own data.
How GPS visitation data makes our algorithm more predictive
Our machine learning begins with identifying a desired target outcome. In this example, we wanted to identify users who would be most likely to visit a Dunkin store location. Once we have defined our target outcome, we need to select a data set that we can use to train our model. To test whether or not including GPS visitation data would improve our predictive power, we used two data sets: one with 6000+ attributes (app installed, demographic, device characteristics) that did not include GPS visitation data; and another with 9000+ attributes that did include GPS visitation data.
The results show that both had incredible predictive power. The model without the GPS visitation data was 3.23 times better than random at predicting users who would be most likely to visit a Dunkin store while the model with the GPS visitation data was 8.79 times more predictive than random. The GPS visitation data lifts the predictive power by 2.72 times!
In short, the more predictive a model is, the more accurate your targeting strategy will be – which will result in less waste on the wrong people and better ROAS – whether you’re trying to advertise to Dunkin store visitors, or turn them into your customers.
Can we predict who would become a mobile customer?
In the “new normal” of social distancing, QSRs are trying to develop ways to connect with customers digitally to promote online ordering and safer, cashless payments. App downloads are critical to solving these problems for QSRs and Machine Learning is one approach that QSRs can use to drive app downloads and online ordering. We set out to see how our target outcome modeling can drive Dunkin app downloads and what attributes are most predictive in identifying who downloads the app.
Overall, this model was also hyper-predictive. Our model was 11.99 times better than random at being able to identify users who would be most likely to download the Dunkin app. We then sought to uncover which attributes were most predictive in identifying users who download the app. In order to do this, we rank-ordered the attributes in the model from most predictive to least predictive and found that location, app install, and visitation data are all super predictive attributes. For example, users who lived in Massachusetts had a 1269% increase in the odds of downloading the app.
Using the ML-generated audiences with the GPS visitation data layer, you can set up a media campaign guaranteed to yield higher ROI. Why spend money on testing audiences that might work when you can use audiences that are 11.9 times more likely than random to deliver the target outcome?
Already using modeling for your audience marketing?
Our modeling is more predictive than the lookalike modeling you will find in many DSPs, DMPs, and social platforms. Here’s why: