OTTO Display Network & COMPASS

In the following we will take a closer look at the COMPASS management component. COMPASS processes gigantic amounts of data and applies AI to determine the appropriate advertising content and the optimal bid level for every single user: this improves visibility, raises inventory quality and enables a deeper understanding of the market than was previously possible.

To select appropriate advertising content, OTTO can not only leverage its own brand, assortment and promotion campaigns, but also the marketing campaigns of OTTO platform partners. Campaign management needs to take both user-relevance and the target KPIs of the individual advertising campaigns into account. Thus, a volatile set of campaigns exists with different content, running times, daily budgets, possibly different optimisation requests, and a cookie pool of known and unknown users. Banner advertising, as described above, is auctioned in real time in digital auctions. Defining the optimal bid in order to maximize campaign performance is therefore the overriding objective.
The following sections describe how we use our knowledge of users to assign campaigns to the optimal recipient groups, and how we calculate perfect bids to optimise each campaign.

COMPASS can be broken down into three main components. Environment analysis is carried out within market modelling, assigning users to suitable campaigns takes place on the basis of KPI forecasts and assigning and determining bids are defined via bid optimization. These three areas will be explained in greater detail below.

Just like many other optimisation problems, Compass also initially concerns gathering information and that way getting to know and understand the advertising market.
Communication between the websites and the Orbidder creates enormous amounts of data in the form of auction logs. With 1.5 billion auctions and 1.5 billion bids, as well as tens of millions of winning notifications , several terabytes of data accumulate every month. The Orbidder's infrastructure is so powerful that it can process traffic peaks of 50,000 requests per second – and even higher. Here’s an extract of a dataset:

In raw form this data contains everything we need to know about the market, and as a result can be processed according to different market characteristics. We focus on the following three aspects:

1. Auction Density: distribution of auctions throughout the day
2. Auction Timeseries: itemised forecast of the number of auctions per day
3. Bid Landscapes & Hard Floor Price: winning probability forecast, depending on the bid

Auctiondensity delivers an empirical description of auctions over the course of the day. It is therefore a measure of German-language website users’ surfing intensity .

In the evaluation, one of several interesting insights is that traffic on Saturday and Sunday increases on average one hour later than on the other days of the week, due to the ‘late-riser’ effect.
The lunch-break peaks from 12:00 onwards are only evident on weekdays, and apparently many users’ weekends already start on Friday at noon. Furthermore, there is generally less web surfing on Saturdays before peaking for the week on Sunday afternoon. This helps us to estimate at any time how many auctions will still be available for the rest of the day.

Auction timeseries
In order to gain detailed knowledge of which advertising spaces will come up for auction, we model thousands of time series for auction volume to a very fine level via fbprophet, determine short and long-term trends, and log the forecasting error. Our modelling levels are:
1. URLs (netloc: e.g. m.spiegel.de),
2. PlacementIDs (clusters of AdSlots similar in format and website placement)
3. Knowledge of the auction user.

The model and forecasting are recalculated daily.

Bidlandscapes & HFP
Bidlandscapes allows us to visualise the relationship between the auction bid and the probability of winning. As in the Auction Timeseries, here we also look at each URL, placement and user status separately. In addition, the time and day of the week also play a role in deriving the probability of winning a bid. To learn these connections we use a Bayesian model and insights from auction theory. Here, we not only look at correlations at the current moment, but also predict the change in competition over the course of the day and week. We also visualise what the bidlandscapes would look like if there were 50% more or 50% fewer competitors in the market. The following examples show that it makes sense to model the relationships individually: