Previous benchmarks are said to have been shattered by the release of a new prediction engine from performance advertising company Criteo.
Decisions are now being made by Criteo’s latest improvement on whether to buy each ad impression based on the likelihood that a user will click on it and make a purchase on the advertiser’s site.
Criteo’s changes mean clients can reach more potential buyers while keeping to their cost-per-sales targets. When measured across billions of impressions, along with millions of users and clicks, it equates to a 38% rise in sales at the same cost of sale.
Describing the Criteo Engine as a ‘major breakthrough’, Jonathan Wolf, chief product officer at Criteo, believes it can now forecast even the most unlikely of actions from potential consumers.
“The Criteo Engine is now processing huge volumes of data with improved models, and as a result we are able to predict events that happen only once out of every 10,000 times – like when an ad impression generates a click and a sale,” Wolf revealed.
The overhaul, which has taken Criteo three years to complete, means the engine can now process up to 15 million predictions each second and respond inside of 20 milliseconds.
Supporting the open source community
Using open source technologies including Apache Hadoop helps the new prediction model analyse many weeks of data in a few hours. On any one day Criteo stores 20 terabytes of data, including information on purchase behaviour.
Hadoop’s ecosystem impressed Criteo to such an extent that it contributed the Hive compatibility layer to the open source Apache Hadoop storage engine, Parque.
Criteo says that its engine’s technological enhancements will be automatically deployed to existing clients across the globe at no added cost.