Our client Ströer distributes video ads of customers. A customer approaches our client and provides targeting criteria (age, gender, and websites) as well as a price per view. When a user clicks on a video, the customer's advertisement campaign then competes with many other campaigns in a millisecond auction. The campaign with the highest bid will then be shown to the user. Of course, customers are highly interested in the potential reach of their campaigns, i.e. how many times their advertisement will be viewed. Thus, it is vital for our client to provide predictions on how many views a campaign will be able to generate.

As a first step, we gathered traffic and tracking information of all websites to which our client distributes ad campaigns. Moreover, we collected all tracked auction data, in order to find winning campaigns, which were able to show their advertisement to the user.

Besides campaigns administered by our client, there are many further external campaigns that compete with all the campaigns our client distributes. Therefore, we have to find a model for internal and external campaign interactions.

Furthermore, we need to deal with human interference with respect to campaign properties. If a campaign does not perform as desired, the sales department of our client usually adjusts targeting criteria and prices. This leads to frequent changes in the internal campaigns.

Since internal mechanisms behind auctions are known and fixed (internal campaigns are ranked by their bidding price and their priority and individual frequency caps are applied if users would see the ad multiple times), we generated ground truth data by a full simulation of real traffic, but with static campaigns. Once ground truth data was available, we could build and test our model. From auction data, we could train a boosted decision tree in order to determine the winning probability of an internal campaign against external campaigns. The interaction of internal campaigns follows a clear set of rules and was thus thoroughly simulated. In order to predict traffic and user behavior, i.e. how often unique users within a specific targeting group view the same videos, we generate histograms of views per unique user for each relevant targeting criterion. From this aggregated information we can sample simulated traffic.

The Product

The software we delivered to our client provides a complete and accurate simulation to predict views for specific campaign video ads. Our client can tell the customers within several minutes if their desired impact is within reach. For example, if the desired amount of views is realistic for a given price per view and fixed targeting criteria.