Bid shading and the adverse selection trap

Bid shading and the adverse selection trap

There's a subtle failure inside bid shading that punishes the impressions you most want to win. It's adverse selection (winning disproportionately the auctions you mispriced upward).

The mechanism, carefully:
— Your shading model predicts a clearing price and you bid near it to win cheaply.
— On auctions where competition is genuinely low, your shaded bid wins easily — good.
— But on auctions where you happen to over-estimate value, you bid high, and you win those too — because high bids win.
— So among your wins, the mispriced-high ones are over-represented. You systematically pay more than the impression was worth on the very impressions you secured.

This is the winner's curse adapted to programmatic: winning is itself evidence that you may have bid above the true market value, because everyone with a more accurate, lower valuation lost to you.

The correction:
— Calibrate your value model on won-impression outcomes (post-click or post-conversion), not on bid-time predictions, so you measure what wins were actually worth.
— Apply a deliberate downward adjustment proportional to how contested a segment is: the more competitors, the more a win signals you may have overbid.

Why it matters: the act of winning carries information that your bid may have been too high. Shading models that ignore this overpay precisely on the inventory they capture — the opposite of the impressions you never see.