Beruflich Dokumente
Kultur Dokumente
In analyses of sponsored search auctions for search advertisements. We also consider a cor-
online advertising, it is customary to model the responding static game of complete information.
dynamic game of incomplete information by con- We analyze the underlying dynamic game of in-
sidering a static games of complete information. complete information, and we establish an upper
This approach is used in Edelman, Ostrovsky bound on the revenue of any equilibrium of any
and Schwarz (2007) (EOS), Varian (2007), and dynamic game in this environment. We then ex-
the subsequent literature. clude equilibria of the corresponding static game
Modeling complex interactions in uncertain with revenue that exceeds this upper bound. See
environments as games of complete information Section D.
has a long history. For example, the Bertrand We use this equilibrium selection criteria to
model of oligopolistic competition posits that assess optimal design of search engine advertis-
companies know their competitors’ costs based ing platforms. We characterize optimal reserve
on their experience from prior interactions. prices and show that the optimal reserve price is
The use of a static game of complete informa- independent of the number of bidders and inde-
tion often offers important benefits. For one, it is pendent of the rate at which click-through rate
tractable – avoiding complex multi-period infor- declines over positions. See Section E.
mation sets in a dynamic game. Furthermore, a Our analysis of reserve prices also lets us as-
suitably-chosen static game can capture impor- sess their welfare effects. We separate the effects
tant characteristics of the underlying dynamic of reserve price increases into direct effects (caus-
game. When a game is repeated over an ex- ing the lowest value bidder to face a higher pay-
tended period, there is good reason to think par- ment) and indirect effects (inducing other bid-
ticipants will learn many characteristics of their ders to increase their bids, thereby increasing
counterparts – supporting the use of a complete others’ payments). We show that most of in-
information model. cremental revenue from setting reserve price op-
Yet the use of a static game of complete in- timally comes not from the low bidder’s direct
formation is also unsatisfying. There is no clear effect, and not from indirect effects on other low
way to identify which (if any) equilibria of the bidders, but rather from the indirect effects on
static game of complete information are a rele- high bidders. This result may appear counter-
vant approximation of the equilibria of the dy- intuitive because top bidders’ large valuations
namic game of incomplete information. place them, in an important sense, “furthest
In this paper, we consider a dynamic game from” the reserve price. See Section II.
of incomplete information used to sell sponsored
I. Environment and Mechanisms
∗ Edelman: Harvard Business School, Boston,
MA 02163 (e-mail: bedelman@hbs.edu). Schwarz: A. Environment
Yahoo! Labs, 2397 Shattuck Ave, Berkeley, CA
94704 (e-mail: mschwarz@yahoo-inc.com). We
Our model for the sponsored search environ-
thank Susan Athey, Brendan Daley, Scott Komin-
ers, Robin Lee, Preston McAfee, Michael Ostrovsky, ment follows EOS. We consider a market for ad-
Al Roth, and Adam Szeidl. A draft of this paper was vertisements triggered by searches for a single
circulated under the title “Optimal Auction Design keyword. Each period, a slate of ads is shown to
in a Multi-Unit Environment: The Case of Spon- users, and the world lasts for 𝑇 periods. Each
sored Search Auctions.” position has a commonly-known click-through
1
2 DRAFT MAY 2010
rate (CTR), and higher positions have higher Advertisers can change their bids at any time.
CTR. In expectation, an ad in position 𝑖 receives Each period, advertisers observe positions of
𝛼𝑖 clicks per period, and the number of clicks de- their competitors. By seeing other advertis-
pends on position only. In a given period, each ers’ positions, each advertiser updates its beliefs
ad can appear in at most one position. about others’ valuations, influencing bids in fu-
There are 𝑁 advertisers. Advertiser 𝑘 values ture periods. When choosing a bid, an advertiser
a click at 𝑠𝑘 , which is 𝑘’s private information, may consider how its action will influence future
does not change over time, and does not depend play of other advertisers. Historic bid informa-
on the position where the ad appears. Bidder tion is relevant in equilibrium because an adver-
values are i.i.d. drawn from a commonly-known tiser’s best response depends on its expectation
distribution with support [0, 𝑠], pdf 𝑓 (.), and cdf of bidding behavior of other advertisers. This in-
𝐹 (.) satisfying the regularity condition from My- formation structure creates a complex dynamic
erson (1981). In particular, we assume virtual game of incomplete information.
valuation 𝜓(𝑠) = 𝑠 − 1−𝐹 (𝑠)
𝑓 (𝑠)
is decreasing in 𝑠. Our key observation is that we can establish
The per-period payoff of advertiser 𝑘 in posi- an upper bound on revenue in any equilibrium
tion 𝑖 is 𝛼𝑖 (𝑠𝑘 − 𝑝𝑘 ), where 𝑝𝑘 denotes the pay- of any dynamic game in this environment with-
ment per-click made by advertiser 𝑘 in that pe- out characterizing the equilibrium of a dynamic
riod. The total payoff of advertiser 𝑘 is the sum game. Importantly, the environment is fairly
∑𝑇
of 𝑘’s period payoffs, i.e. 𝑡=1 𝛼𝑘(𝑡) (𝑠𝑘 − 𝑝𝑘 (𝑡))
simple – letting us characterize an optimal mech-
where 𝑘(𝑡) denotes the position of advertiser 𝑘 anism for this environment, thereby bounding
in period 𝑡 and 𝑝𝑘 (𝑡) denotes the payment made revenue in any dynamic mechanism in this en-
by advertiser 𝑘 in period 𝑡. vironment without solving for an equilibrium of
a dynamic game. An optimal mechanism is pre-
B. The Dynamic Game of Incomplete sented in Section D.
Information
C. Approximation with a One-Shot Game
We begin with a stylized model that captures of Complete Information
the dynamic aspects of the mechanism used in
practice. Search engines sell advertisements us- The dynamic mechanism described above is
ing real-time generalized second price (GSP) sufficiently complex that it is difficult or per-
auctions. As each search occurs, an auction is haps impossible to solve analytically for an equi-
conducted to determine which ads should be dis- librium strategy. The literature on sponsored
played to the corresponding user. Advertisers search auctions sidesteps the complexity of the
are ranked based on bids: all else equal, the dynamic game of incomplete information by
higher the bid, the higher the position assigned modeling it as a one shot game of complete in-
to the corresponding advertiser, and hence the formation. The complete information game cor-
more clicks the advertiser receives. If there are responding to the incomplete information game
more positions than advertisers, a position is al- retains payoffs and the actions exactly as in the
located to all advertisers whose bid exceeds the previous section. However, there are two impor-
reserve price. The per-click payment of the ad- tant differences: the complete information game
vertiser in the bottom position equals the re- lasts only one period, and the per-click values of
serve price. For each other advertiser, the per- all bidders are common knowledge. The justi-
click payment equals the bid of an advertiser fication for considering a complete information
who is ranked one position lower. (For exam- model is that in a dynamic game of incomplete
ple, the advertiser in the third position pays an information, players learn each others’ values,
amount equal to the fourth-largest bid.) The making complete information plausible in the
mechanisms used by Google, Yahoo!, and Mi- long-run. This justification follows longstand-
crosoft adCenter build on this GSP approach ing practice in the industrial organization liter-
– though they add variations such as adjusting ature, modeling price competition among firms
prices based on ad quality. See e.g. Abrams and as a one-shot game with common knowledge of
Schwarz (2008). each firm’s production costs.
VOL. 100 NO. 2 OPTIMAL AUCTION DESIGN IN SPONSORED SEARCH 3
The one-shot game of complete information librium of the complete information game, one
has multiple equilibria. Some of these equilibria can view a complete information game as a
can be ruled out by the envy-free restriction in- valid approximation of the incomplete informa-
troduced in EOS and Varian (2007). The envy- tion game. This suggests an equilibrium selec-
free condition requires that no bidder wishes to tion criteria: An equilibrium of complete infor-
exchange positions with a bidder ranked above mation game can be ruled out if there does not
him. Note that the payment of the advertiser exist an equilibrium of the corresponding game
in position 𝑖 − 1 equals the bid of advertiser 𝑖, of incomplete information that converges to the
so envy freeness means that for each advertiser, same outcome. As a result, expected revenue
𝛼𝑖 (𝑠𝑘 −𝑝𝑘 ) ≥ 𝛼𝑖−1 (𝑠𝑘 −𝑏𝑘 ) where 𝑝𝑘 is advertiser in a “plausible” equilibrium of complete infor-
𝑘’s payment per click in his current position, and mation game integrated over all possible real-
𝑏𝑘 is 𝑘’s current bid. EOS and Varian (2007) izations of bidder values cannot exceed the ex-
show that the equilibrium with VCG-equivalent pected revenues in some equilibrium of the in-
payoffs yields the lowest revenues of all envy-free complete information game. We call this the
equilibria. Both EOS and Varian identify the Non-Contradiction Criteria (NCC).
lowest-revenue envy-free (LREF) equilibrium as Let us define NCC precisely. Denote by 𝑅(𝑇 )
the most plausible. Varian offers an informal the expected revenue in the highest revenue per-
argument for the LREF equilibrium, while the fect Bayesian equilibrium of the incomplete in-
main result of EOS shows that in a generalized formation game that lasts for T periods. Let 𝑅
English auction, the LREF equilibrium is the denote the (largest possible) expected per period
unique outcome. Cary et al. (2007) show that equilibrium revenues in an incomplete informa-
if bidders play myopic strategies in a repeated tion game 𝑅 = 𝑇1 lim𝑇 →∞ 𝑅(𝑇 ). Consider equi-
GSP auction, the system converges to the LREF librium per-period revenues in an equilibrium 𝜇
equilibrium. A number of subsequent papers use of the complete information game. Denote by
the LREF equilibrium as the solution concept 𝑟𝜇 (s) the expected revenues in Nash equilibrium
for sponsored search auctions. 𝜇 of a complete information GSP auction where
It remains uncertain whether the LREF equi- values of bidders are given by vector s.
librium is the “right” equilibrium to select in
GSP auctions. Athey and Nikopolev (2010) use DEFINITION 1: A NE 𝜇 of a complete in-
noise in estimates of bidders’ click-through-rate formation game fails NCC if it generates
to pin down the equilibrium. Thompson and greater expected revenues than any equilib-
Leyton-Brown (2009) computationally explore rium of the corresponding incomplete infor-
the set of all equilibria in complete informa- mation
tion GSP. This approach would be unnecessary if ∫ ∫ game. That is, 𝜇 fails NCC if
𝑠1
... 𝑠
𝑟 𝜇 (s)𝑓 (𝑠1 )...𝑓 (𝑠𝑁 )𝑑𝑠1 ...𝑑𝑠𝑁 >𝑅.
a convincing equilibrium selection criteria were 𝑁
that NCC is not a restriction on the set of equi- son framework to the case of multiple heteroge-
libria in Γ; NCC says nothing about reasonable neous objects when bidders’ private information
equilibria in Γ. Rather, NCC restricts equilibria is one-dimensional.1
of Γ when Γ is meant to capture behavior in 𝐺. Obtaining an optimal mechanism is a key
We use NCC to analyze sponsored search ad- intermediate result in our analysis. Simi-
vertising. But the underlying idea is general. larly, Roughgarden and Sundararajan (2007)
Many models of complete information games are and Athey and Elison (2009) also obtain an
simplified representations of complex dynamic optimal mechanism as an intermediate result,
games of incomplete information. In these cases, but for entirely different purposes. Iyengar and
the set of plausible equilibria of a static game Kumar (2006) consider computational methods
should be considered in the context of the cor- to estimate optimal auctions when clickthrough
responding incomplete information game. Our rates are arbitrary. In contrast, the structure of
insight is that even if the incomplete informa- clickthrough rates assumed in our model allows
tion game is very complex, it may be possible us to obtain an analytic result.
to bound the set of equilibria of the incomplete The following proposition characterizes the
information game. optimal mechanism2 :
COROLLARY 3: The optimal reserve price in are more slots than advertisers, that 𝑛 bidders
the generalized English auction is independent of had values greater than r𝑜𝑙𝑑 , and that 𝑗 bidders
the number of bidders, number of slots being auc- drop out as a result of the increase in reserve.
tioned, and the rate of decline in click-through In the following proposition, we compare the
rate from position to position. revenues resulting from a change in reserve price.