ResearchResearch
Internet Advertising Auctions
David Pennock, Yahoo! Research - New York
Contributed slides:K.Asdemir, H.Bhargava, J.Feng, S.Lahaie, M.Schwarz
ResearchResearch
Advertising Then and Now• Then: Think real estate
Phone callsManual negotiation“Half doesn’t work”
• Now: Think Wall StreetAutomation, automation, automationAdvertisers buy contextual attention:
User i on page j at time tComputer learns what ad is bestComputer mediates ad sales: Auction!Computer measures which ads work
ResearchResearch
Advertising Then & Now: Video
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
http://ycorpblog.com/2008/04/06/this-one-goes-to-11/
ResearchResearch
Advertising: NowTools Disciplines
• Auctions
• Machine learning
• Optimization
• Sales
• Economics &Computer Science
• Statistics &Computer Science
• Operations Research Computer Science
• Marketing
search “las vegas travel”, Yahoo!
Sponsored search auctions
Space next to search results is sold at auction
“las vegas travel” auction
Ad exchanges
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Outline
• Motivation: Industry facts & figures
• Introduction to sponsored search– Brief and biased history
– Allocation and pricing: Google vs old Yahoo!
– Incentives and equilibrium
• Ad exchanges
• Selected survey of research
• Prediction markets
Auctions Applications
eBay
– 216 million/month
Google / Yahoo!
– 11 billion/month (US)
Auctions Applications
0.00
20.00
40.00
60.00
80.00
100.00
120.00
140.00
160.00
180.00
Market Capitalization (billions)
Ebay (founded 1995) Google (founded 1998)Sotheby's (founded 1744)
• eBay • Google
Auctions Applications
• eBay • Google
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Newsweek June 17, 2002
“The United States of EBAY”
• In 2001: 170 million transactions worth $9.3 billion in 18,000 categories “that together cover virtually the entire universe of human artifacts—Ferraris, Plymouths and Yugos; desk, floor, wall and ceiling lamps; 11 different varieties of pockets watches; contemporary Barbies, vintage Barbies, and replica Barbies.”
• “Since everything that transpires on Ebay is recorded, and most of it is public, the site constitutes a gold mine of data on American tastes and preoccupations.”
“The United States of Search”
• 11 billion searches/month
• 50% of web users search every day
• 13% of traffic to commercial sites
• 40% of product searches
• $8.7 billion 2007 US ad revenue (41% of $21.2 billion US online ads; 2% of all US ads)
• Still ~20% annual growth after years of nearly doubling
• Search data: Covers nearly everything that people think about: intensions, desires, diversions, interests, buying habits, ...
Online ad industry revenue
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
http://www.iab.net/media/file/IAB_PwC_2007_full_year.pdf
ResearchResearch
Introduction tosponsored search
• What is it?• Brief and biased history• Allocation and pricing: Google vs Yahoo!• Incentives and equilibrium
search “las vegas travel”, Yahoo!
Sponsored search auctions
Space next to search results is sold at auction
“las vegas travel” auction
Sponsored search auctions
• Search engines auction off space next to search results, e.g. “digital camera”
• Higher bidders get higher placement on screen
• Advertisers pay per click: Only pay when users click through to their site; don’t pay for uncliked view (“impression”)
Sponsored search auctions
• Sponsored search auctions are dynamic and continuous: In principle a new “auction” clears for each new search query
• Prices can change minute to minute;React to external effects, cyclical & non-cyc– “flowers” before Valentines Day
– Fantasy football
– People browse during day, buy in evening
– Vioxx
Vioxx
0
5
10
15
20
25
30
9/14/089/15/089/16/089/17/089/18/089/19/089/20/089/21/089/22/089/23/089/24/089/25/089/26/089/27/089/28/089/29/089/30/0810/1/0810/2/0810/3/0810/4/0810/5/0810/6/0810/7/0810/8/0810/9/0810/10/0810/11/0810/12/0810/13/08
Date
Price ($)
Example price volatility: Vioxx
Sponsored search today
• 2007: ~ $10 billion industry– ‘06~$8.5B ‘05~$7B ‘04~$4B ‘03~$2.5B ‘02~$1B
• $8.7 billion 2007 US ad revenue (41% of US online ads; 2% of all US ads)
• Resurgence in web search, web advertising
• Online advertising spending still trailing consumer movement online
• For many businesses, substitute for eBay
• Like eBay, mini economy of 3rd party products & services: SEO, SEM
Sponsored SearchA Brief & Biased History
• Idealab GoTo.com (no relation to Go.com)
– Crazy (terrible?) idea, meant to combat search spam
– Search engine “destination” that ranks results based on who is willing to pay the most
– With algorithmic SEs out there, who would use it?
• GoTo Yahoo! Search Marketing
– Team w/ algorithmic SE’s, provide “sponsored results”
– Key: For commercial topics (“LV travel”, “digital camera”) actively searched for, people don’t mind (like?) it
– Editorial control, “invisible hand” keep results relevant
• Enter Google
– Innovative, nimble, fast, effective
– Licensed Overture patent (one reason for Y!s ~5% stake in G)
Sponsored SearchA Brief & Biased History
• Overture introduced the first design in 1997: first price, rank by bid
• Google then began running slot auctions in 2000: second price, rank by revenue (bid * CTR)
• In 2002, Overture (at this point acquired by Yahoo!) then switched to second-price. Still uses rank by bid; Moving toward rank by revenue
Thanks: S. Lahaie
Sponsored SearchA Brief & Biased History
• In the beginning:– Exact match, rank by bid, pay per click, human editors
– Mechanism simple, easy to understand, worked, somewhat ad hoc
• Today & tomorrow:– “AI” match, rank by expected revenue (Google), pay per
click/impression/conversion, auto editorial, contextual (AdSense, YPN), local, 2nd price (proxy bid), 3rd party optimizers, budgeting optimization, exploration exploitation, fraud, collusion, more attributes and expressiveness, more automation, personalization/targeting, better understanding (economists, computer scientists)
Sponsored Search ResearchA Brief & Biased History
• Circa 2004
– Weber & Zeng, A model of search intermediaries and paid referrals
– Bhargava & Feng, Preferential placement in Internet search engines
– Feng, Bhargava, & PennockImplementing sponsored search in web search engines: Computational evaluation of alternative mechanisms
– Feng, Optimal allocation mechanisms when bidders’ ranking for objects is common
– Asdemir, Internet advertising pricing models
– Asdemir, A theory of bidding in search phrase auctions: Can bidding wars be collusive?
– Mehta, Saberi, Vazirani, & VaziranAdWords and generalized on-line matching
• Key papers, survey, and ongoing research workshop series
– Edelman, Ostrovsky, and Schwarz, Internet Advertising and the Generalized Second Price Auction, 2005
– Varian, Position Auctions, 2006
– Lahaie, Pennock, Saberi, Vohra, Sponsored Search, Chapter 28 in Algorithmic Game Theory, Cambridge University Press, 2007
– 1st-3nd Workshops on Sponsored Search Auctions4th Workshop on Ad Auctions -- Chicago Julu 8-9, 2008
Allocation and pricing
• Allocation
– Yahoo!: Rank by decreasing bid
– Google: Rank by decreasing bid * E[CTR] (Rank by decreasing “revenue”)
• Pricing
– Pay “next price”: Min price to keep you in current position
ResearchResearch
Yahoo Allocation: Bid Ranking“las vegas travel” auction search “las vegas travel”, Yahoo!
pays $2.95per click
pays $2.94
pays $1.02
... bidder ipays bidi+1+.01
ResearchResearch
Google Allocation: $ Ranking“las vegas travel” auction
x E[CTR] = E[RPS]
x E[CTR] = E[RPS]
x E[CTR] = E[RPS]
x E[CTR] = E[RPS]
x E[CTR] = E[RPS]
ResearchResearch
Google Allocation: $ Ranking“las vegas travel” auction search “las vegas travel”, Google
x .1 = .301
x .2 = .588
x .1 = .293
x E[CTR] = E[RPS]
x E[CTR] = E[RPS]
TripReservations
Expedia
pays 3.01*.1/.2+.01 = 1.51per click
pays 2.93*.1/.1+.01 = 2.94
pays bidi+1*CTRi+1/CTRi+.01
LVGravityZone
etc...
Aside: Second price auction(Vickrey auction)
• All buyers submit their bids privately
• buyer with the highest bid wins;pays the price of the second highest bid
$150$120
$90
$50
Only pays $120
Incentive Compatibility(Truthfulness)
• Telling the truth is optimal in second-price (Vickrey) auction
• Suppose your value for the item is $100;if you win, your net gain (loss) is $100 - price
• If you bid more than $100:
– you increase your chances of winning at price >$100
– you do not improve your chance of winning for < $100
• If you bid less than $100:
– you reduce your chances of winning at price < $100
– there is no effect on the price you pay if you do win
• Dominant optimal strategy: bid $100
– Key: the price you pay is out of your control
• Vickrey’s Nobel Prize due in large part to this result
Vickrey-Clark-Groves (VCG)
• Generalization of 2nd price auction
• Works for arbitrary number of goods, including allowing combination bids
• Auction procedure:– Collect bids
– Allocate goods to maximize total reported value (goods go to those who claim to value them most)
– Payments: Each bidder pays her externality;Pays: (sum of everyone else’s value without bidder) - (sum of everyone else’s value with bidder)
• Incentive compatible (truthful)
Yahoo! Confidential
Is Google pricing = VCG?
Well, not really …
Put Nobel Prize-winning theories to work.Google’s unique auction model uses Nobel Prize-winning economic theory to eliminate the winner’s curse – that feeling that you’ve paid too much. While the auction model lets advertisers bid on keywords, the AdWords™ Discounter makes sure that they only pay what they need in order to stay ahead of their nearest competitor.
https://google.com/adsense/afs.pdf
Yahoo! Confidential
VCG pricing
• (sum of everyone else’s value w/o bidder) - (sum of everyone else’s value with bidder)
• CTRi = advi * posi (key “separability” assumption)
• pricei = 1/advi*(∑j<ibidj*CTRj + ∑j>ibidj*advj*posj-1 -∑j≠ibidj*CTRj )
= 1/advi*(∑j>ibidj*advj*posj-1 - ∑j>ibidj*CTRj )
• Notes
– For truthful Y! ranking set advi = 1. But Y! ranking technically not VCG because not efficient allocation.
– Last position may require special handling
Yahoo! Confidential
Next-price equilibrium
• Next-price auction: Not truthful: no dominant strategy
• What are Nash equilibrium strategies? There are many!
• Which Nash equilibrium seems “focal” ?
• Locally envy-free equilibrium [Edelman, Ostrovsky, Schwarz 2005]
Symmetric equilibrium [Varian 2006]
Fixed point where bidders don’t want to move or – Bidders first choose the optimal position for them: position i
– Within range of bids that land them in position i, bidder chooses point of indifference between staying in current position and swapping up with bidder in position i-1
• Pure strategy (symmetric) Nash equilibrium
• Intuitive: Squeeze bidder above, but not enough to risk “punishment” from bidder above
Yahoo! Confidential
Next-price equilibrium
• Recursive solution:
posi-1*advi*bi = (posi-1-posi)*advi*vi+posi*advi+1*bi+1
bi = (posi-1-posi)*advi*vi+posi*advi+1*bi+1
posi-1*advi
• Nomenclature:Next price = “generalized second price” (GSP)
ResearchResearch
Ad exchanges
• Right Media• Expressiveness
ResearchResearch
Online Advertising Evolution
1. Direct: Publishers sell owned & operated (O&O) inventory
2. Ad networks: Big publishers place ads on affiliate sites, share revenueAOL, Google, Yahoo!, Microsoft
3. Ad exchanges: Match buy orders from advertisers with sell orders from publishers and ad networksKey distinction: exchange does not “own” inventory
Yahoo! Confidential
Exchange Basics
Exchange
Demand
Inventory
Netflix
Vonage
Auto.com…
Advertisers
Ad.com
CPX
Tribal…
Networks
MySpace
Six Apart
Looksmart
Monster…
Publishers
[Source: Ryan Christensen]
Yahoo! Confidential
Right Media Publisher Experience
• Publisher can select / reject specific advertisers
• Green = linked network
• Light Blue = direct advertiser
• Publishers can traffic their own deals by clicking “Add Advertiser”
The publisher can approve creative from each advertiser
[Source: Ryan Christensen]
Yahoo! Confidential
Right Media Advertiser Experience
• Advertisers can set targets for CPM, CPC and CPA campaigns
• Set budgets and frequency caps
• Locate publishers, upload creative and traffic campaigns
[Source: Ryan Christensen]
Yahoo! Confidential
Expressiveness
• “I’ll pay 10% more for Males 18-35”
• “I’ll pay $0.05 per impression, $0.25 per click, and $5.25 per conversion”
• “I’ll pay 50% more for exclusive display, or w/o Acme”
• “My marginal value per click is decreasing/increasing”
• “Never/Always show me next to Acme”“Never/Always show me on adult sites”“Show me when Amazon.com is 1st algo search result”
• “I need at least 10K impressions, or none”
• “Spread out my exposure over the month”
• “I want three exposures per user, at least one in the evening”
Design parameters: Advertiser needs/wants,computational/cognitive complexity, revenue
ResearchResearch
Expressiveness Example
• Competition constraints
3 x .05 = .15
1 x .05 = .05
b xCTR = RPS
ResearchResearch
Expressiveness Example
• Competition constraints
4 x .07 = .28
b xCTR = RPS
monopoly bid
ResearchResearch
Expressiveness: Design
• Multi-attribute bidding
Advertiser1
Advertiser2
Male users (50%)
$1 $2
Female users (50%)
$2 $1
Un-differentiated
$1.50 $1.50
Advertiser1
Advertiser2
Pre-qualified (50%)
$2 $2
Other (50%) $1 $1
Un-differentiated
$1.50 $1.50
Yahoo! Confidential
Expressiveness: Less is More
• Pay per conversion: Advertisers pay for user actions (sales, sign ups, extended browsing, ...)
– Network sends traffic
– Advertisers rate users/types 0-100Pay in proportion
– Network learns, optimizes traffic, repeat
• Fraud: Short-term gain only: If advertisers lie, they stop getting traffic
Yahoo! Confidential
Expressiveness: Less is More
• “I’m a dry cleaner in Somerset, New Jersey with $100/month. Advertise for me.”
• Can advertisers trust network to optimize?
ResearchResearch
Coming Convergence:ML and Mechanism Design
Mechanism(Rules)
e.g. Auction,Exchange, ...
Stats/ML/OptEngine
Stats/ML/OptEngine
Stats/ML/OptEngine
Stats/ML/OptEngine
Stats/ML/OptEngine
ResearchResearch
ML Inner Loop
• Optimal allocation (ad-user match) depends on: bid, E[clicks], E[sales], relevance, ad, advertiser, user, context (page, history), ...
• Expectations must be learned• Learning in dynamic setting requires
exploration/exploitation tradeoff• Mechanism design must factor all this
in! Nontrivial.
ResearchResearch
Selected Survey ofInternet Advertising Research
An Analysis of Alternative Slot Auction Designs for
Sponsored SearchSebastien Lahaie, Harvard University**work partially conducted at Yahoo! Research
ACM Conference on Electronic Commerce, 2006
Source: S. Lahaie
Objective
•Initiate a systematic study of Yahoo! and Google slot auctions designs.
•Look at both “short-run” incomplete information case, and “long-run” complete information case.
Source: S. Lahaie
Outline•Incomplete information (one shot game)
• Incentives
•Efficiency
• Informational requirements
•Revenue
•Complete Information (long-run equilibrium)
•Existence of equilibria
•Characterization of equilibria
•Efficiency of equilibria (“price of anarchy”)
Source: S. Lahaie
The Model• slots, bidders
•The type of bidder i consists of
•a value per click of , realization
•a relevance , realization
• is bidder i’s revenue, realization
•Ad in slot is viewed with probability
So CTRi,k =
•Bidder i’s utility function is quasi-linear:
Source: S. Lahaie
The Model (cont’d)
• is i.i.d on according to
• is continuous and has full support
• is common knowledge
•Probabilities are common knowledge.
•Only bidder i knows realization
•Both seller and bidder i know , but other bidders do not
Source: S. Lahaie
Auction Formats
•Rank-by-bid (RBB): bidders are ranked according to their declared values ( )
•Rank-by-revenue (RBR): bidders are ranked according to their declared revenues ( )
•First-price: a bidder pays his declared value
•Second-price (next-price): For RBB, pays next highest price. For RBR, pays
•All payments are per click
Source: S. Lahaie
•First-price: neither RBB nor RBR is truthful
•Second-price: being truthful is not a dominant strategy, nor is it an ex post Nash equilibrium (by example):
•Use Holmstrom’s lemma to derive truthful payment rules for RBB and RBR:
•RBR with truthful payment rule is VCG
Incentives
1 61 4
Source: S. Lahaie
Efficiency•Lemma: In a RBB auction with either a
first- or second-price payment rule, the symmetric Bayes-Nash equilibrium bid is strictly increasing with value. For RBR it is strictly increasing with product.
•RBB is not efficient (by example).
•Proposition: RBR is efficient (proof).
0.5 6
1 4
Source: S. Lahaie
First-Price Bidding Equilibria• is the expected resulting
clickthrough rate, in a symmetric equilibrium of the RBB auction, to a bidder with value y and relevance 1.
• is defined similarly for bidder with product y and relevance 1.
•Proposition: Symmetric Bayes-Nash equilibrium strategies in a first-price RBB and RBR auction are given by, respectively:
Source: S. Lahaie
Informational Requirements
•RBB: bidder need not know his own relevance, or the distribution over relevance.
•RBR: must know own relevance and joint distribution over value and relevance.
Source: S. Lahaie
Revenue Ranking
•Revenue equivalence principle: auctions that lead to the same allocations in equilibrium have the same expected revenue.
•Neither RBB nor RBR dominates in terms of revenue, for a fixed number of agents, slots, and a fixed .
Source: S. Lahaie
Complete Information Nash
Equilibria
Argument: a bidder always tries to match the next-lowest bid to minimize costs. But it is not an equilibrium for all to bid 0.
Argument: corollary of characterization lemma.
Source: S. Lahaie
Characterization of Equilibria
•RBB: same characterization with replacing
Source: S. Lahaie
Price of AnarchyDefine:
Source: S. Lahaie
Exponential Decay
• Typical model of decaying clickthrough rate:
• [Feng et al. ’05] find that their actual clickthrough data is fit well by such a model with
• In this case
Source: S. Lahaie
Conclusion
• Incomplete information (on-shot game):
•Neither first- nor second-pricing leads to truthfulness.
•RBR is efficient, RBB is not
•RBB has weaker informational requirements
•Neither RBB nor RBR is revenue-dominant
• Complete information (long-run equilibrium):
•First-price leads to no pure strategy Nash equilibria, but second-price has many.
•Value in equilibrium is constant factor away from “standard” value.
Source: S. Lahaie
Future Work
•Better characterization of revenue properties: under what conditions on does either RBB or RBR dominate?
•Revenue results for complete information case (relation to Edelman et al.’s “locally envy-free equilibria”).
Source: S. Lahaie
Research Problem: Online Estimation of Clickrates
•Make virtually no assumptions on clickrates.
•Each different ranking yields (1) information on clickrates and (2) revenue.
•Tension between optimizing current revenue based on current information, and gaining more info on clickrates to optimize future revenue (multi-armed bandit problem...)
•Twist: chosen policy determines rankings, which will affect agent’s equilibrium behavior.
Source: S. Lahaie
ResearchResearch
Equilibrium revenue simulations of hybrid sponsored search mechanisms
Sebastien Lahaie, Harvard University**work conducted at Yahoo! Research
David Pennock, Yahoo! Research
Yahoo! Confidential
Revenue effects
• What gives most revenue?– Key: If rules change, advertiser bids will change
– Use Edelman et al. envy-free equilibrium solution
OvertureHighest bid wins
Google/Yahoo!Highest bid*CTR wins
s=0s=1/2 ?
s=1s=3/4 ?
HybridHighest bid*(CTR)s wins
Yahoo! Confidential
Monte-Carlo simulations
• 10 bidders, 10 positions
• Value and relevance are i.i.d. and have lognormal marginals with mean and variance (1,0.2) and (1,0.5) resp.
• Spearman correlation between value and relevance is varied between -1 and 1.
• Standard errors are within 2% of plotted estimates.
Source: S. Lahaie
Yahoo! Confidential
Source: S. Lahaie
Yahoo! Confidential
Source: S. Lahaie
Yahoo! Confidential
Source: S. Lahaie
Yahoo! Confidential
Preliminary Conclusions
• With perfectly negative correlation(-1), revenue, efficiency, and relevance exhibits threshold behavior
• Squashing up to this threshold can improve revenue without too much sacrifice in efficiency or relevance
• Squashing can significantly improve revenue with positive correlation
Source: S. Lahaie
Pragmatic Robots and Equilibrium Bidding in GSP
Auctions
Michael Schwarz, Yahoo! Research
Ben Edelman, Harvard University
Source: M. Schwarz
Yahoo! Confidential
Testing game theory
• Empirical game theory– Analytic solutions intractable in all but simplest settings
– Laboratory experiments cumbersome, costly
– Agent-based simulation: easy, cheap, allow massive exploration; Key: modeling realistic strategies
• Ideal for agent-based simulation: when real economic decisions are already delegated to software
“If pay-per-click marketing is so strategic, how can it be automated? That’s why we developed Rules-Based Bidding. Rules-Based Bidding allows you to apply the kind of rules you would use if you were managing your bids manually.” Atlas http://www.atlasonepoint.com/products/bidmanager/rulesbased
Thanks: M. Schwarz
Yahoo! Confidential
Bidders’ actual strategiesSource: M. Schwarz
Yahoo! Confidential
Models of GSP
1. Static game of complete information
2. Generalized English Auction (simple dynamic model)
More realistic model
• Each period one random bidder can change his bid
• Before the move a bidder observes all standing bids
Source: M. Schwarz
Yahoo! Confidential
Pragmatic Robot (PR)
• Find current optimal position iImplies range of possible bids: Static best response (BR set)
• Choose envy-free point inside BR set:Bid up to point of indifference between position i and position i-1
• If start in equilibrium PRs stay in equilibrium
Source: M. Schwarz
Yahoo! Confidential
Convergence of PRSimulation
0 100 200 300 400 500 600 700 8000.2
0.4
0.6
0.8
1
1.2
1.4
1.6
simulation rounds - convergence to 0.000001 after 329 iterations
Total Surplus Search Engine RevenueAdvertiser Surplus Computed Equilibrium
Source: M. Schwarz
Yahoo! Confidential
Convergence of PRSource: M. Schwarz
Yahoo! Confidential
Convergence of PR
• The fact that PR converges supports the assertion that the equilibrium of a simple model informs us about the outcome of intractable dynamic game that inspired it
Complex game that we can not solve
Simple model inspired by a complex game
?
Source: M. Schwarz
Yahoo! Confidential
Playing with Ideal Subjects
Largest Gap (commercially available strategy)Moves your keyword listing to the largest bid gap within a specified set of positions
Regime One: 15 robots all play Largest Gap
Regime Two: one robot becomes pragmatic
By becoming Pragmatic pay off is up 16%Other assumptions: values are log normal, mean valuation 1, std dev 0.7 of the underlying normal, bidders move sequentially in random order
Source: M. Schwarz
Yahoo! Confidential
ROI
• Setting ROI target is a popular strategy
• For any ROI goal the advertiser who switches to pragmatic gets higher payoff
Source: M. Schwarz
Yahoo! Confidential
If others play ROI targeter
• Bidders 1,...,K-1 bid according to the ROI targeting strategy
• What is K’s best response?
bidder
bidder payoffs if bidder K plays
ROI targeting
PR
1
…
K-1
K 0.0387 0.0457
Source: M. Schwarz
Yahoo! Confidential
Reinforcement Learnervs Pragmatic Robot
• Pragmatic learner outperforms reinforcement learner (that we tried)
• Remark: reinforcement learning does not converge in a problem with big BR set
Source: M. Schwarz
Yahoo! Confidential
Conclusion
• A strategy inspired by theory seems useful in practice: PR beats commercially available strategies and other reasonable baselines
• Since PR converges and performs well, the equilibrium concept is sound in spite the fact that some theoretical assumptions are violated and there are plenty of players who are “irrational”
• When bidding agents are used for real economic decisions (e.g., search engine optimization), we have an ideal playground for empirical game theory simulations
Thanks: M. Schwarz
ResearchResearch
First Workshop on Sponsored Search Auctionsat ACM Electronic Commerce, 2005
Organizers:
Kursad Asdemir, University of Alberta Hemant Bharghava, University of California Davis Jane Feng, University of Florida Gary Flake, Microsoft David Pennock, Yahoo! Research
ResearchResearch
Papers
• Mechanism Design• Pay-Per-Percentage of Impressions: An Advertising
Method that is Highly Robust to Fraud, J.Goodman• Stochastic and Contingent-Payment Auctions,
C.Meek,D.M.Chickering, D.B.Wilson• Optimize-and-Dispatch Architecture for Expressive
Ad Auctions, D.Parkes, T.Sandholm• Sponsored Search Auction Design via Machine
Learning, M.-F. Balcan, A.Blum, J.D.Hartline, Y.Mansour• Knapsack Auctions, G.Aggarwal, J.D. Hartline• Designing Share Structure in Auctions of Divisible
Goods, J.Chen, D.Liu, A.B.Whinston
•
•
ResearchResearch
Papers
• Bidding Strategies• Strategic Bidder Behavior in Sponsored Search Auctions,
Benjamin Edelman, Michael Ostrovsky• A Formal Analysis of Search Auctions Including
Predictions on Click Fraud and Bidding Tactics, B.Kitts, P.Laxminarayan, B.LeBlanc, R.Meech
• User experience• Examining Searcher Perceptions of and Interactions with
Sponsored Results, B.J.Jansen, M. Resnick• Online Advertisers' Bidding Strategies for Search,
Experience, and Credence Goods: An Empirical Investigation, A.Animesh, V. Ramachandran,
• S.Vaswanathan•
ResearchResearch
Stochastic Auctions C.Meek,D.M.Chickering, D.B.Wilson
• Ad ranking allocation rule is stochastic
• Why?• Reduces incentive for “bid jamming”• Naturally incorporates explore/exploit mix• Incentive for low value bidders to join/stay?
• Derive truthful pricing rule
• Investigate contingent-payment auctions:Pay per click, pay per action, etc.
• Investigate bid jamming, exploration strategies
ResearchResearch
Expressive Ad AuctionsD.Parkes, T.Sandholm
• Propose expressive bidding semantics for ad auctions (examples next)• Good: Incr. economic efficiency, incr. revenue• Bad: Requires combinatorial optimization;
Ads need to be displayed within milliseconds
• To address computational complexity, propose “optimize and dispatch” architecture: Offline scheduler “tunes” an online (real-time) dispatcher
ResearchResearch
Expressive bidding I
• Multi-attribute bidding
Advertiser1
Advertiser2
Male users (50%)
$1 $2
Female users (50%)
$2 $1
Un-differentiated
$1.50 $1.50
Advertiser1
Advertiser2
Pre-qualified (50%)
$2 $2
Other (50%) $1 $1
Un-differentiated
$1.50 $1.50
ResearchResearch
Expressive bidding II
• Competition constraints
3 x .05 = .15
1 x .05 = .05
b xCTR = RPS
ResearchResearch
Expressive bidding II
• Competition constraints
4 x .07 = .28
b xCTR = RPS
monopoly bid
ResearchResearch
Expressive bidding III
• Guaranteed future delivery• Decreasing/increasing marginal value• All or nothing bids• Pay per: impression, click, action, ...• Type/id of distribution site (content match)• Complex search query properties• Algo results properties (“piggyback bid”)• Ad infinitum• Keys: What advertisers want; what
advertisers value differently; controlling cognitive burden; computational complexity
Second Workshop on Sponsored Search Auctions
Kursad Asdemir, University of Alberta
Jason Hartline, Microsoft Research
Brendan Kitts, Microsoft
Chris Meek, Microsoft Research
Organizing Committee
Source: K. Asdemir
Objectives
Diversity Participants
Industry: Search engines and search engine marketers Academia: Engineering, business, economics schools
Approaches Mechanism Design Empirical Data mining / machine learning
New Ideas
Source: K. Asdemir
History & Overview
First Workshop on S.S.A. Vancouver, BC 2005 ~25 participants 10 papers + Open discussion 4 papers from Microsoft Research
Second Workshop on S.S.A. ~40-50 participants 10 papers + Panel 3 papers from Yahoo! Research
Source: K. Asdemir
Participants
Industry Yahoo!, Microsoft, Google Iprospect (Isobar), Efficient Frontier, HP Labs, Bell
Labs, CommerceNet
Academia Several schools
Source: K. Asdemir
Papers Mechanism design
Edelman, Ostrovsky, and Schwarz Iyengar and Kumar Liu, Chen, and Whinston Borgs et al.
Bidding behavior Zhou and Lukose Szymanski and Lee Asdemir Borgs et al.
Data mining Regelson and Fain Sebastian, Bartz, and Murthy
Source: K. Asdemir
Panel: Models of Sponsored Search:What are the Right Questions? Proposed by
Lance Fortnow and Rakesh Vohra
Panel members Kamal Jain, Microsoft Research Rakesh Vohra, Northwestern University Michael Schwarz, Yahoo! Inc David Pennock, Yahoo! Inc
Source: K. Asdemir
Panel Discussions Mechanisms
Competition between mechanisms Ambiguity vs Transparency: “Pricing” versus “auctions” Involving searchers
Budget Hard or a soft constraint Flighting (How to spend the budget over time?)
Pay-per-what? CPM, CPC, CPS Risk sharing Fraud resistance
Transcript available!
Source: K. Asdemir
ResearchResearch
Web resources
• 1st Workshop website & papers:http://research.yahoo.com/workshops/ssa2005/
• 1st Workshop notes (by Rohit Khare):http://wiki.commerce.net/wiki/RK_SSA_WS_Notes
• 2nd Workshop website & papers:http://www.bus.ualberta.ca/kasdemir/ssa2/
• 2nd Workshop panel transcript:(thanks Hartline & friends!)http://research.microsoft.com/~hartline/papers/panel-SSA-06.pdf
• 3rd Workshop websitehttp://opim-sun.wharton.upenn.edu/ssa3/index.html
• 4th Workshop websitehttp://research.yahoo.com/workshops/adauctions2008/
More Challenges
• Unifying search, display, content, offline
• Economics of attention
• Directly rewarding users, control, privacy3-party game theoretic equilibrium
• Predicting click through rates
• Detecting spam/fraud
• Pay per “action” / conversion
• Number/location/size of of ads
• Improved targeting / expressiveness
• $15B Question: Monetizing social networks, user-generated content
ResearchResearch
Prediction Markets
David Pennock, Yahoo! Research
ResearchResearch
Bet = Credible Opinion
• Which is more believable?More Informative?
• Betting intermediaries• Las Vegas, Wall Street, Betfair, Intrade,...• Prices: stable consensus of a large
number of quantitative, credible opinions• Excellent empirical track record
Obama will win the 2008 US Presidential election
“I bet $100 Obama will win at 1 to 2 odds”
ResearchResearch
A Prediction Market
• Take a random variable, e.g.
• Turn it into a financial instrument payoff = realized value of variable
$1 if $0 if
I am entitled to:
Bird Flu Outbreak US 2008?(Y/N)
Bird FluUS ’08
Bird FluUS ’08
ResearchResearch
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
http://intrade.com
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
ResearchResearch
Prediction Markets:Examples & Research
ResearchResearch
The Wisdom of CrowdsBacked in dollars• What you can say/learn
% chance that• Obama wins• GOP wins Texas• YHOO stock > 30• Duke wins tourney• Oil prices fall• Heat index rises• Hurricane hits Florida• Rains at place/time
• Where
• IEM, Intrade.com• Intrade.com• Stock options market• Las Vegas, Betfair• Futures market• Weather derivatives• Insurance company• Weatherbill.com
ResearchResearch
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Prediction MarketsWith Money Without
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
ResearchResearch
The Widsom of CrowdsBacked in “Points”• HSX.com• Newsfutures.com• InklingMarkets.com• Foresight Exchange• CasualObserver.net• FTPredict.com• Yahoo!/O’Reilly Tech Buzz• ProTrade.com• StorageMarkets.com• TheSimExchange.com• TheWSX.com• Alexadex, Celebdaq, Cenimar, BetBubble, Betocracy, CrowdIQ,
MediaMammon,Owise, PublicGyan, RIMDEX, Smarkets, Trendio, TwoCrowds
• http://www.chrisfmasse.com/3/3/markets/#Play-Money_Prediction_Markets
http://tradesports.com
http://betfair.com
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Screen capture 2007/05/18
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Screen capture 2008/05/07
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
ResearchResearch
Example: IEM 1992
[Source: Berg, DARPA Workshop, 2002]
ResearchResearch
Example: IEM
[Source: Berg, DARPA Workshop, 2002]
ResearchResearch
Example: IEM
[Source: Berg, DARPA Workshop, 2002]
Does it work? Yes, evidence from real markets, laboratory
experiments, and theory Racetrack odds beat track experts [Figlewski 1979] Orange Juice futures improve weather forecast [Roll 1984] I.E.M. beat political polls 451/596 [Forsythe 1992, 1999][Oliven
1995][Rietz 1998][Berg 2001][Pennock 2002]
HP market beat sales forecast 6/8 [Plott 2000]
Sports betting markets provide accurate forecasts of game outcomes [Gandar 1998][Thaler 1988][Debnath EC’03][Schmidt 2002]
Laboratory experiments confirm information aggregation[Plott 1982;1988;1997][Forsythe 1990][Chen, EC’01]
Theory: “rational expectations” [Grossman 1981][Lucas 1972]
Market games work [Servan-Schreiber 2004][Pennock 2001]
[Thanks: Yiling Chen]
ResearchResearch
Prediction Markets:Does Money Matter?
ResearchResearch
The Wisdom of CrowdsWith Money Without
IEM: 237 Candidates HSX: 489 Movies
1 2 5 10 20 50 100estimate
1
2
5
10
20
50
100
actual
ResearchResearch
The Wisdom of CrowdsWith Money Without
ResearchResearch
Real markets vs. market gamesHSX FX, F1P6
probabilisticforecasts
forecast source avg log scoreF1P6 linear scoring -1.84F1P6 F1-style scoring -1.82betting odds -1.86F1P6 flat scoring -2.03F1P6 winner scoring -2.32
ResearchResearch
Does money matter? Play vs real, head to headExperiment• 2003 NFL Season• ProbabilitySports.com
Online football forecasting competition
• Contestants assess probabilities for each game
• Quadratic scoring rule• ~2,000 “experts”, plus:• NewsFutures (play $)• Tradesports (real $)
• Used “last trade” prices
Results:• Play money and real
money performed similarly• 6th and 8th respectively
• Markets beat most of the ~2,000 contestants• Average of experts
came 39th (caveat)
Electronic Markets, Emile Servan-Schreiber, Justin Wolfers, David Pennock and Brian Galebach
ResearchResearch
0
25
50
75
100
TradeSports Prices
0 20 40 60 80 100NewsFutures Prices
Fitted Value: Linear regression
45 degree line
n=416 over 208 NFL games.Correlation between TradeSports and NewsFutures prices = 0.97
Prices: TradeSports and NewsFutures
Prediction Performance of MarketsRelative to Individual Experts
020406080
100120140160180200220240260280300
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Week into the NFL season
Rank
NewsFutures
Tradesports
0
10
20
30
40
50
60
70
80
90
100
Observed Frequency of Victory
0 10 20 30 40 50 60 70 80 90 100Trading Price Prior to Game
TradeSports: Correlation=0.96NewsFutures: Correlation=0.94
Data are grouped so that prices are rounded to the nearest ten percentage points; n=416 teams in 208 games
Market Forecast Winning Probability and Actual Winning ProbabilityPrediction Accuracy
ResearchResearch
Does money matter? Play vs real, head to head
Probability-Football Avg
TradeSports(real-money)
NewsFutures(play-money)
DifferenceTS - NF
Mean Absolute Error
= lose_price
[lower is better]
0.443
(0.012)
0.439
(0.011)
0.436
(0.012)
0.003
(0.016)
Root Mean Squared Error
= ?Average( lose_price2 )
[lower is better]
0.476
(0.025)
0.468
(0.023)
0.467
(0.024)
0.001
(0.033)
Average Quadratic Score
= 100 - 400*( lose_price2 )
[higher is better]
9.323
(4.75)
12.410
(4.37)
12.427
(4.57)
-0.017
(6.32)
Average Logarithmic Score
= Log(win_price)
[higher (less negative) is better]
-0.649
(0.027)
-0.631
(0.024)
-0.631
(0.025)
0.000
(0.035)
Statistically:TS ~ NFNF >> AvgTS > Avg
ResearchResearch
A Problem w/ Virtual CurrencyPrinting Money
Alice1000
Betty1000
Carol1000
ResearchResearch
A Problem w/ Virtual CurrencyPrinting Money
Alice5000
Betty1000
Carol1000
ResearchResearch
YootlesA Social Currency
Alice0
Betty0
Carol0
ResearchResearch
YootlesA Social Currency
I owe you 5
Alice-5
Betty0
Carol5
ResearchResearch
YootlesA Social Currency
credit: 5 credit: 10
I owe you 5
Alice-5
Betty0
Carol5
ResearchResearch
YootlesA Social Currency
credit: 5 credit: 10
I owe you 5 I owe you 5
Alice-5
Betty0
Carol5
ResearchResearch
YootlesA Social Currency
credit: 5 credit: 10
I owe you 5 I owe you 5
Alice3995
Betty0
Carol5
ResearchResearch
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
YootlesA Social Currency• For tracking gratitude among friends• A yootle says “thanks, I owe you one”
ResearchResearch
Combinatorial Betting
ResearchResearchCombinatorics ExampleMarch Madness
ResearchResearch
Combinatorics ExampleMarch Madness• Typical today
Non-combinatorial• Team wins Rnd 1• Team wins Tourney• A few other “props”• Everything explicit
(By def, small #)• Every bet indep:
Ignores logical & probabilistic relationships
• Combinatorial• Any property• Team wins Rnd k
Duke > {UNC,NCST}ACC wins 5 games
• 2264 possible props(implicitly defined)
• 1 Bet effects related bets “correctly”;e.g., to enforce logical constraints
Expressiveness:Getting Information
• Things you can say today:– (43% chance that) Hillary wins
– GOP wins Texas
– YHOO stock > 30 Dec 2007
– Duke wins NCAA tourney
• Things you can’t say (very well) today:– Oil down, DOW up, & Hillary wins
– Hillary wins election, given that she wins OH & FL
– YHOO btw 25.8 & 32.5 Dec 2007
– #1 seeds in NCAA tourney win more than #2 seeds
Expressiveness:Processing Information
• Independent markets today:– Horse race win, place, & show pools
– Stock options at different strike prices
– Every game/proposition in NCAA tourney
– Almost everything: Stocks, wagers, intrade, ...
• Information flow (inference) left up to traders
• Better: Let traders focus on predicting whatever they want, however they want: Mechanism takes care of logical/probabilistic inference
• Another advantage: Smarter budgeting
ResearchResearch
Automated Market Makers
• A market maker (a.k.a. bookmaker) is a firm or person who is almost always willing to accept both buy and sell orders at some prices
• Why an institutional market maker? Liquidity! • Without market makers, the more expressive the betting
mechanism is the less liquid the market is (few exact matches)• Illiquidity discourages trading: Chicken and egg• Subsidizes information gathering and aggregation: Circumvents
no-trade theorems
• Market makers, unlike auctioneers, bear risk. Thus, we desire mechanisms that can bound the loss of market makers
• Market scoring rules [Hanson 2002, 2003, 2006]
• Dynamic pari-mutuel market [Pennock 2004]
[Thanks: Yiling Chen]
Overview: Complexity Results
Permutations Boolean
General Pair Subset General 2-clause Restrict Tourney
Call Market
NP-hard NP-hard Poly co-NP-complete
? ?
Market Maker
(LMSR)
#P-hard #P-hard #P-hard #P-hard #P-hard Poly
ResearchResearch
New Prediction Game
ResearchResearch
Mech Design for Prediction
Financial Markets Prediction Markets
Primary Social welfare (trade)Hedging risk
Information aggregation
Secondary Information aggregation Social welfare (trade)Hedging risk
ResearchResearch
Mech Design for Prediction
• Standard Properties• Efficiency• Inidiv. rationality• Budget balance• Revenue• Truthful (IC)• Comp. complexity
• Equilibrium• General, Nash, ...
• PM Properties• #1: Info aggregation• Expressiveness• Liquidity• Bounded budget• Truthful (IC)• Indiv. rationality• Comp. complexity
• Equilibrium• Rational
expectations
Competes with:experts, scoringrules, opinionpools, ML/stats,polls, Delphi
ResearchResearch
Discussion
• Are incentives for virtual currency strong enough?• Yes (to a degree)• Conjecture: Enough to get what people already know;
not enough to motivate independent research• Reduced incentive for information discovery possibly
balanced by better interpersonal weighting
• Statistical validations show HSX, FX, NF are reliable sources for forecasts• HSX predictions >= expert predictions• Combining sources can help
ResearchResearch
Catalysts
• Markets have long history of predictive accuracy: why catching on now as tool?
• No press is bad press: Policy Analysis Market (“terror futures”)
• Surowiecki's “Wisdom of Crowds”• Companies:
• Google, Microsoft, Yahoo!; CrowdIQ, HSX, InklingMarkets, NewsFutures
• Press: BusinessWeek, CBS News, Economist, NYTimes, Time, WSJ, ...http://us.newsfutures.com/home/articles.html
CFTC Role
• MayDay 2008: CFTC asks for help
• Q: What to do with prediction markets?
• Right now, the biggest prediction markets are overseas, academic (1), or just for fun
• CFTC may clarify, drive innovation
• Or not
ResearchResearch
Conclusion
• Prediction Markets:hammer = market, nail = prediction• Great empirical successes• Momentum in academia and industry• Fascinating (algorithmic) mechanism design
questions, including combinatorial betting
• Points-paid peers produce prettygood predictions