Why Astrea? She is the virgin goddess of justice, innocence, purity and precision, which is exactly the goal of the proposed algorithm: We do not want to pay more or less but just the right price.

This algorithm is especially powerful when exploring first price auction requests. 


In RTB, most DSPs suffer from what we call the Data Cycle Vicious Wheel. Machine learning algorithms are known to be very sensitive to the data used during learning; the main problem resides in the fact that this data acquired with the bids depends on the data learnt by the algorithm, whose knowledge is obtained with the bids performed in the previous day… and so on.

In few words: we obtain data from those requests we already know and train new models with this knowledge aiming to explore new fields but ending up acquiring the same type of data.


To overcome this, we present a new algorithm whose function is to model the WinRate curve with respect to the bid value for each request. With this, our exploring data stops being acquired by algorithms whose target is obtaining the optimum CPA; instead, the new one focuses only in managing data volume thus exploring new unseen requests.

Given a request A, if we are able to model the Win Rate curve with respect to the historical bids performed for this type of request.  It is only a matter of deciding the volume of impressions we want to obtain for a given number of bids (i.e. the Win Rate) and extracting the optimum value to bid in the next auction with the mapping function learnt by the algorithm.


What if the learnt curve is too flat? There are two reasons that can explain it. 

On one hand, the request type might belong to a very expensive request, from which we have historically bid in a very low range of values making the algorithm learn only the left part of the curve.

On the other hand, the algorithm might have seen little data of this request type, a volume not high enough to learn the curve properly.

In any case, we want to protect us from bidding really high values when dealing with flat curves. 

The algorithm consists of three strategies. In the first one, called specific, we deal with very granular features such us the app. The second one is a fallback model which is based on the same logic but being trained with more general features. This strategy is used only when the algorithm hasn’t learnt well the Win Rate curve for a given request with the specific one. If it hasn’t learnt neither the curve with the fallback strategy, then it bids randomly across a bid spread determined by the price_floor and a maxBid.

The following picture shows perfectly the transition among all strategies at the start of a campaign:

The more data we acquire, the better we can learn the Win Rate curves. As a result we should see a decrease in the use of the random strategy in parallel with an increase in the use of the other two strategies (the increase in the specific one should be higher like in the image above).


One important part in our campaigns is to keep exploring to discover new unseen opportunities. Astrea is the perfect solution to explore new inventories while avoiding to pay extra price which will penalize us especially in the current first price auction scenario.