<b>Win notices bias your training data, and the bias is systematic</b>

<b>Win notices bias your training data, and the bias is systematic</b>

A win notice is the callback an exchange fires to tell you that you won and at what price. DSPs learn clearing prices largely from these notices. The problem is that win notices are a biased sample of all auctions, and the bias is predictable.

1. You only get a clean price signal on auctions you <i>win</i>. Auctions you lose return, at best, a loss code — rarely the actual clearing price.
2. This is censored data: the expensive auctions you lost are systematically missing from your observed price distribution.
3. A model trained naively on wins alone will underestimate true clearing prices, because it never sees the high-priced auctions it couldn't afford.

This is the same censoring problem that haunts bid shading, viewed from the data-science side. The fix is survival-analysis-style methods that account for the missing losses, or deliberate high exploration bids to occasionally observe the expensive tail.

<b>Why it matters:</b> any pricing model that learns only from won impressions will drift optimistic and gradually price you out of competitive inventory without anyone noticing. When evaluating a DSP's bidding intelligence, ask specifically how it corrects for censored losses — it's the difference between a model that holds calibration and one that slowly hallucinates a cheaper market than exists.