BidCore operates a machine learning engine that predicts the probability of an event happening (click, conversion, etc) using the historical data (from the bid requests, logs, and 3rd party data sources) and multiple variables analysis to predict events probability for bidding optimization and fraud prevention.
The engine predicts the probability of various event types:
VAST event (e.g 30 seconds video view).
Any custom event type that can be tracked can also be used by the prediction engine for predicting the optimal bidding strategy that will let the client achieve multiple goals, e.g buy maximum clicks for CPC not higher than X CPM with target CTR, fix margin, while spending the budget in time.
Typical tasks the predict engine helps solve¶
Predict the probability of a bid winning.
Predict the probability of a conversion.
Predict the viewability of an impression.
Solve problems like high CPC or CPA.
Provide a foundation for developing optimization algorithms.
A predict model calculates the probability of a target event using a set of parameters that can be taken from bid requests or 1st / 3rd party data sets. To create the most efficient set of keys for new prediction models, R&D analyses new parameters “contrast” — how much the parameter influences the accuracy of prediction.
To maximize data volumes available for analysis and increase prediction accuracy, the predict engine uses all the relevant impressions / bids from all the client campaigns (even if is counting probability of an event for a separate campaign).
What kind of results does a good predict engine produce¶
Let’s imagine that all people in the world have the CTR 0.2%. The worst predict model predicts the same 0.2% for all people. Formally it meets the criteria, predicting the CTR properly, but it won’t be precise when it comes to particular segments of people. A good predict model sees the difference between segments of people basing on predefined criteria (for example, the difference between users of different mobile OSs - 0.4% for iOS users versus 0.01% for Android users with the overall CTR being the same 0.2%). So, a good predict engine should see this difference and provide good features that allow seeing it.
What makes the BidCore predict engine different¶
It supports any custom event that can be tracked, e.g. viewability, length of video views on a domain, win price for a particular network, win price in HB environment, post-install events, etc.
It can build accurate predictions with smaller data sets, thus allowing it to provide predictions even for new creatives (that have no historical data) based on general platform results. After the creative generates actual history, the prediction estimation becomes more accurate.
It can use various external data sets to increase the quality of predictions, e.g. page visits, purchases, activities on an advertiser’s website, external segment.
It updates prediction models in real time, i.e. the predict models are always relevant as they are trained continuously using data from current buying activity and updated regularly.
There is a multitude of data variables behind any ad placement that can be used for predicting the probability of certain events. The probability is a ratio (correlation) between events used as attempts versus successes. Any event can be used as an attempt or successes, but two most common models to build are imp2click and imp2conv (where impressions are used as attempts, whereas clicks and conversions are successes). Also imp2conv can be split to two models - imp2pcc (post-click) and imp2pvc (post-view).
In general, the predict engine can estimate any attempt/successes ratio based on the client business logic. Other common existing options are: session2conv, bid2imp, bid_price2imp_price etc.
Upon requests from clients, new custom models can be built and implemented with any events used as the basis for attempts and successes.
The predict engine delivers accurate performance estimations starting with very low volumes of data and learns from any new data provided. It constantly updates and adjusts models to accomodate time sensitive performance data.
Models can be built on platform, campaign, advertiser, placement, or creative level.
The R&D team helps to develop, set up, improve and control predict models based on attributes that work best for each particular case.
The predict engine provides probability estimates for new creatives with no historical data and corrects them as data is collected.
Models can be created for any parameters, e.g viewability, optimal price, pre-conversion event, post-install event, etc.
Predict Data Privacy¶
Regardless of how a buyer accesses BidCore services (UI or API), all first party data, unique reporting parameters, custom buying strategies and prediction models are placed in a dedicated 1st party node, completely isolated from all other BidCore clients. This data is fully secure and remains the exclusive property of the originating client.