{"page":"/blog/bid-shading-explained","modified":"2026-11-10","license":"https://creativecommons.org/licenses/by/4.0/","columns":["Family","What it learns","Feedback needed","Strengths","Limits","Source"],"rows":[["Exchange-side shading service","The exchange lowers high bids to a price that keeps the same win rate","None from the buyer","Easy for buyers who cannot shade","The exchange's yield and the buyer's surplus pull in different directions; examples include AdX Bid Translation Service, Rubicon Estimated Market Rate, AppNexus Bid Price Optimization","Gligorijevic et al., 2020"],["Segment-based parametric shading","A shading function per inventory segment (exchange, domain, device, layout), tuned by recursive least squares","Won and lost outcomes per segment","Simple, fast at bid time","Segments must be chosen by hand; sparse segments learn slowly","Gligorijevic et al., 2020"],["Winning-price distribution forecasting","The distribution of winning prices for a set of targeting attributes, often log-normal","Historical winning prices","Gives a full curve for planning and pacing","A fixed distribution shape can miss real data","Cui et al., KDD 2011"],["Censored winning-price regression","The winning price, treating lost bids as censored observations","Wins, plus bid price on losses","Works when the exchange hides the winning price","Assumptions break on some real traffic, hence mixture models","Wu et al., KDD 2015"],["Win-rate curve with surplus maximisation","A modified logistic win-rate curve per request, then the bid that maximises surplus","Win or loss only","No minimum-bid-to-win needed; reported 7% more profit than bidding the most likely winning price","Accuracy depends on calibration of the curve","Pan et al., AdKDD 2020"],["Machine-learned shading on rich features","Factorization machines or deep distribution networks predicting the shading ratio or the full price distribution","Minimum bid to win (open auctions) or censored outcomes","Shares information across sparse segments; deployed at scale","Heavier models must fit a 20 ms bid budget","Gligorijevic et al., 2020; Zhou et al., KDD 2021"]],"sourceUrls":[["https://arxiv.org/abs/2009.01360"],["https://arxiv.org/abs/2009.01360"],["https://doi.org/10.1145/2020408.2020454"],["https://doi.org/10.1145/2783258.2783276"],["https://arxiv.org/abs/2009.09259"],["https://arxiv.org/abs/2009.01360","https://arxiv.org/abs/2107.06650"]]}