 |
|||||||||||||||||||||||||||||||||||||||||||||||||||
|
Live blogging: Hal Varian, Berkeley, "Nash equilibrium of Google Auction". U Michigan STEIT seminar. Summary: Analysis of the Google Ad words auction using game theory. Given a set of advertisers and their click-values, solve for an equilibrium set of bids. Invert, and given a set of bids, find a set of click-values that is consistent. The Nash equilibrium is a good fit. Aside: Berkeley is changing the name of SIMS to the School of Information. Approved yesterday. Presentation is being done in Mathematica Evolution: first ads were priced CPM model, with sales people. Didn't scale. Then, top ads were CPM, side ads were CPC; noticed big differences between negotiated prices and auction prices. Auction more efficient than sales. Simplified rules: top bidder gets top place and so on, but everyone pays the price of the bid of the advertiser below them. In reality the ordering includes an adjustment for ad quality. Nash equilibrium: no one wants to move up or down. do we have one? Google goal: solve for maximum revenue in equilibrium conditions. Individual goal: given everyone else's bid, what's the best bid for me? Not looking at bid strategies. best bid: expenditure of the guy above/below me plus/minus the value of the extra click. use recursion; start with the (value? bid?) of the first excluded bidder, add up from there. users use the traffic estimator to map out the cost/bid space, by iterating through pricing. "bid jamming", forcing someone else to pay too much - "doesn't happen so much" because of the adjustment for ad quality. heuristic: move up until the incremental cost per click exceeds your value per click. Quality adjustment. What do you really do? Actual CTR = position plus "ad quality" effect = expected CTR independent of position. "The world's biggest logistic regression". Estimate the ad-specific effect; it used to be based on 1000 impressions, now it's different. Effectively ordered by "bids per impression". Effectively you get a better position for the same price for a better ad. Are you effectively charging for impressions, or clicks? Easier to detect click-through fraud than impression fraud, a question with some news value. Top two positions on RHS worth a lot more; partly from impression values, part from top of page effects. Conversion rates are about the same independent on position. Busy auctions are fairly priced - if you are looking for popular words a lot of people are bidding. Less busy words ("long tail" - not Hal's term) auctions are regularly cheaper. Search engine optimization. Singulars convert worse than plurals....lots of other SEO optimizing people know more. "Google as yenta". Matching up buyers and sellers. VCG auction: pay cost you impose on other agents. Pay extra clicks from everyone moving up - same as lower bound on symetric Nash equilibrium. (proof in Demange and Gale 1985) If you do VCG, you just report the truth - maybe if you start all over you'd do that - but has open question on revenue impact since you are at lower bound. Practical problems - collusion, Most of the revenue comes from the fully sold auctions. |
Last modified: Friday, October 31, 2008 at 9:26 PM.
Tech resources |
||||||||||||||||||||||||||||||||||||||||||||||||||
