Your loss data is censored — and that biases every model you build

Your loss data is censored — and that biases every model you build

Every buyer's bid log has a structural blind spot: censored loss data (you see that you lost, but rarely the price you lost to). Ignoring this censoring quietly biases your price models.

The problem, precisely:
— On wins, you observe everything: your bid and the price you paid.
— On losses, you typically observe only the fact of losing. The winning price is hidden.
— So your sample of clearing prices is truncated: it over-represents auctions that cleared low enough for you to win.

If you estimate a segment's typical clearing price from your wins alone, you systematically under-estimate it, because every auction that cleared above your bid is invisible to you. Your model then shades too aggressively, loses more, and the bias compounds.

The statistical fix:
— Treat this as a survival-analysis problem. Each lost auction is a censored observation: the true clearing price is somewhere above your bid.
— Estimate the clearing-price distribution accounting for that censoring (the same math used for time-to-event data), rather than averaging only observed wins.
— Where SSPs provide minimum-to-win on losses, use it — it converts a censored point into a known one and sharpens the estimate dramatically.

Why it matters: naive win-only analysis is not just noisy, it's biased in a known direction — downward. Correcting for censoring is the difference between a shading model that reflects the market and one that systematically misjudges it.