Rank and Impression Estimation in TAC Ad Auctions

Existing empirical bidding research on ad auctions often aims to
estimate model parameters (e.g., click-through-rates) from aggregate
data (e.g., average position, total number of impressions, etc.).
This is not surprising, because aggregate data are what the search
engines reveal.  But it has recently been pointed out that aggregate
data can produce biased estimates.  We construct a
disaggregated model of positions and impressions from aggregate
statistics.  Previous work has demonstrated that the disaggregation
process is feasible using tree-based search techniques.  We extend the
previous approach by formulating the disaggregation process as two
intertwined problems, rank and impression estimation.
We solve these problems using a mathematical programming formulation,
and a problem-specific tree search.  We evaluate the merits of these
two solution techniques in a simulated ad auction environment, the Ad
Auctions division of the annual Trading Agent Competition (TAC AA).

Joint work with Jordan Berg, Carleton Coffrin, and Eric Sodomka.