Auction Markets

Jon Levin
Winter 2010
Economics 136
FCC and Radio Spectrum
FCC regulates use of electromagnetic radio spectrum:
used for broadcast TV, radio, cell phones, WiFi, etc.
Why regulate?
There is a limited amount of spectrum
There are many potential users
There are interference problems if users overlap.
So, how should the FCC decide who gets a license to
use spectrum?
Historically, licenses were allocated administratively (TV
& radio stations) or by lottery.
Spectrum auctions
Coase (1959) suggested that the FCC should auction
spectrum licenses.
If there were no transaction costs, the initial assignment of
ownership wouldnt matter (the Coase Theorem).
But the real world isnt like that decentralized trade may
not lead to efficient allocations (more on this later).
In the early 1990s, the FCC started to think about
auctions as a way to allocate licenses efficiently, and
adopted a new design proposed by Stanford economists.
Many countries now use auctions that result in hundreds of
millions, or billions, of dollars of government revenue.
Sponsored search auctions
Google revenue in 2008: $21,795,550,000.
Hal Varian, Google chief economist:
What most people dont realize is that all that money
comes pennies at a time.
Google revenue comes from selling ads: there is an
auction each time someone enters a search query
Bids in the auction determine the ads that appear on the
RHS of the page, and sometimes the top.
Google design evolved from earlier, and problematic,
design used by Overture (now part of Yahoo!)
We will see that Googles design is closely related to
the theory of matching weve already studied.
British CO
2
Auctions
In 2002, the British government decided to spend
215 million paying firms to reduce CO
2
emissions.
But what price to pay per unit? And which firms to reward?
Solution: run an auction to find the market price
Greenhouse Gas Emissions Trading Scheme Auction
Price starts high and decreases each round.
Each round, bidders state tons of CO
2
they will abate
Cost to UK: tons of abatement times price.
Auction ended when total cost equaled the budget.
Result: 34 firms paid to reduce emission by a total of
4 million metric tons of CO
2.

Auctions everywhere
Auctions are commonly used to sell (and buy) goods that
are idiosyncratic or hard to price.
Auction theory also has close and beautiful ties to
standard price theory (monopoly theory) and matching.
Real estate
Art, antiques, estates
Collectibles (eBay)
Used cars, equipment
Emissions permits
Natural resources: timber,
gas, oil, radio spectrum
Financial assets: treasury
bills, corporate debt.
Bankruptcy auctions
Sale of companies:
privatization, IPOs, take-
overs, etc.
Procurement: highways,
construction, defense.
Auction Theory
Selling a single good
Well start with sale of a single good (later, consider
many goods).
Why not just set a price?
Seller may not know what price to set.
Potential buyers know what theyd pay, but arent telling.
Remember from last time auction serves as a
mechanism for price discovery.
Lets consider different ways to run an auction.
Canonical model
Potential buyers
Two bidders (later N bidders)
Each bidder i has value v
i

Each v
i
drawn from uniform distribution on [0,1].

Same ideas will apply with N bidders and a value
distribution other than U[0,1].

Seller gets to set the auction rules.
Ascending auction
Price starts at zero, and rises slowly.
Buyers indicate their willingness to continue bidding
(e.g. keep their hand up) or can exit.
Auction ends when just one bidder remains.
Final bidder wins, and pays the price at which the
second remaining bidder dropped out.

How should you bid?
Ascending Auction
Optimal strategy: continue bidding until the
price just equals your value.
Bidder with highest value will win.
Winner will pay second highest value.

Example with three bidders
Suppose values are 25, 33, and 75.
First exits at 25, second at 33 and auction ends.
Ascending auction revenue
Suppose we repeatedly take two draws from U[0,1].
On average, the highest draw will be 2/3
On average, the second highest will be 1/3.

So the average (or expected) revenue from an ascending auction
with two bidders who have values drawn from U[0,1] is 1/3.

If we have N bidders with values drawn from U[0,1]
on average, the highest draw will be N/N+1
and the second highest draw will be (N-1)/(N+1).
So the expected revenue will be (N-1)/(N+1).
Aside: Order Statistics
0 1
0 1
0 1
1/2
1/3 2/3
2/4 1/4 3/4
One draw
Two draws
Three draws
Ascending auction profit
Suppose values are drawn from U[0,1]. What profit
does a bidder with value v expect?
His probability of winning is v

(why?)
If he wins, he gets the object worth v
He also pays the highest losing value.
If he wins, he expects on average to pay half his value
(why?), or v/2.
So average (or expected) profit for value v bidder:
v [v-v/2] = v
2
[1/2] = (1/2)v
2
.
Second price auction
Bidders submit sealed bids.
Seller opens the bids.
Bidder who submitted the highest bid wins.
Winners pays the second highest bid.

How should you bid?
Second price auction
Theorem. The optimal strategy in the second price auction
is to bid your value.
Proof
Suppose you bid b>v.
If the highest opposing bid is less than v, or higher than b, it
makes no difference.
If the highest opposing bid is between v and b; you win if you bid
b, but pay above your value, so better to bid v.
Suppose you bid b<v.
Only matters if the highest opposing bid is between b and v.
Then bidding v is better you win and pay less than your value.
Second price auction
0 1
0 1
v
b v
Bid true value v
Bid b>v
If opponent bid is here,
win and make money
If opponent bid is here,
lose the auction
If opponent bid is here,
win and make money
If opponent
bid is here,
lose auction
If bid is
here, win
but lose
money!
Second price auction
In equilibrium, everyone bids their value.
Bidder with highest value wins.
Pays an amount equal to second highest value.
Exactly the same as the ascending auction!
i.e. same winner, same revenue, same expected
profit for a bidder with value v.

Vickrey auction
Second price auction is an example of a more
general Vickrey auction that can be used to sell
multiple goods.
Rules for Vickrey auction (will return to this later)
Everyone submits their value(s)
Seller allocates to maximize surplus
Set prices so the profit of each winner equals his
contribution to the total surplus the difference between
social surplus if he is or is not counted as a participant.
Equivalently, each winner pays the externality he imposes
by displacing other possible winners.
Think about how the 2
nd
price auction does this!
Sealed tender
Bidders submit sealed bids.
Seller opens the bids and
Bidder who submitted highest bid wins.
Winner pays his own bid.

Now what is the optimal strategy?
Sealed tender, cont.
Best to submit a bid less than your true value.
How much less?
Submitting a higher bid
increases the chance you will win
increases the amount youll pay if you do win
Optimal bid depends on what you think the others
will bid (unlike in the second-price auction!).
We need to consider an equilibrium analysis.
Nash equilibrium
Defn: A set of bidding strategies is a Nash
equilibrium if each bidders strategy choice
maximizes his payoff given the strategies of
the others.
In an auction game, bidders do not know their
opponents values, i.e. there is incomplete
information.
So each bidders equilibrium strategy must
maximize her expected payoff accounting for the
uncertainty about opponent values.
Solving the sealed bid eqm
Suppose j=i uses the strategy: bid b
j
= |v
j
.
Bidder i understands js strategy (the eqm assumption), but
doesnt know js bid exactly because he doesnt know v
j
.
Suppose i bids b
i
. Hell win with probability
Pr(b
i
>|v
j
)= Pr(b
i
/| >v
j
)= b
i
/|
So bidder is bidding problem is
max
b
(b/|)(v
i
-b)
First order condition for optimal bidding
0 = (1/|)(v
i
-b)

(b/|)

Solving for equilibrium
First order condition
0 = (1/|)(v
i
-b)

(b/|)

Re-arranging and cancelling out
b/| = (1/|)(v
i
-b)
b = v
i
-b
b =(1/2)v
At symmetric equilibrium, b
i
=(1/2)v
i
, b
j
=(1/2)v
j
---
both bidders bid half their value.
With N bidders, equilibrium is b=[(N-1)/N]v.
Sealed bid equilibrium
Weve derived the eqm: b(v)=(1/2)v

Bidder with the highest value wins in eqm.

What is the revenue, on average?
Revenue equals bid of the high value bidder
High value, on average, is 2/3
So highest bid, on average, is 1/3

Same as the ascending and second price!

Is it the same for each realization of bidder values?
Example: suppose bidder values are and .

Descending price
Price starts high (at least $1).
Price drops slowly (continuously).
At any point, a bidder can claim the item at the
current price (and pay that price).
Auction ends as soon as some bidder claims item.

How should you bid?
Descending price
Strategically equivalent to the first price auction!
Suppose the bidders had to send in computer programs to
do the bidding it would be a 1
st
price auction!
Does actually being there make it any different?
No, in both auctions, your bid only matters if you are the
winner or tied for winning a slightly lower bid means
paying less if you win, but maybe you lose out.
For any strategies by opponents, bidder i chooses the
stopping price (bid) to maximize Pr(win)(v-b) same
problem as in the first price auction.
So eqm strategies in first price and descending
auction are the same, and so is expected revenue.
All-pay auction
Bidders submit bids
Seller opens bids
Bidder submitting the highest bid wins
All bidders pay their bids.

How should you bid?
All pay auction
Clearly want to bid less than your value
Bidding more means
Greater chance of winning
Pay more for sure

Suppose we find equilibrium bid strategies.
How will the bids compare to the first-price auction?
Will seller raise more or less revenue in equilibrium?
Comparison of Auctions
At least in our example, a number of standard
auctions share the following properties of
bidders play according to equilibrium.
the allocation is efficient
average revenue is the same
average profit of a value v bidder is the same.
Next time well explore this result further.
Bidder Strategy
Are bidders really strategic?
Game theory models of auctions assume that
bidders understand the environment and
behave strategically. A good assumption?

Consider two examples:
Skewed bidding in Forest Service auctions
Excessive bidding in internet auctions
Scoring rule auctions
In many auctions, bidders dont just bid a price, but
many prices that are combined into a score.
The score determines who wins, but the unit prices
determine the eventual payments.
Contractors bid an hourly rate for labor and a cost for
materials. Owner picks the bid that appears the most
cheapest but actual payment depends on the work done.
Firms bidding for timber in the national forests bid a unit
price for each species. The prices are multiplied by the
estimated quantities to determine the winner.
What are the bidders incentives?
Timber auction example
Forest Service estimates there are 100 mbf of Douglas Fir and
100 mbf of Western Spruce.
Bidders make per-unit bids for Fir and Spruce.
A bidder offers prices of $50 and $60.
Total bid is 100*50 + 100*60 = $11,000.
High bid wins, but payment depends on actual quantities.
Suppose the ($50,$60) bid wins.
Estimated quantities are 100, 100; expected payment $11,000.
But if the actual quantities are 100 of fir and 50 of spruce, the
actual payment is 100*50 + 50*60 = $8,000.
Bidding incentives
Suppose FS estimates (100,100) of fir, spruce.
Bidder can submit a total bid of $10,000 by:
Bidding $50 for Fir and $50 for Spruce.
Bidding $100 for Fir and $0 for Spruce
Bidding $0 for Fir and $100 for Spruce.
Suppose bidder estimates (200,0) for fir, spruce.
The three alternative bids lead to expected payments of
$10,000, $20,000, and $0!
Incentive to skew ones bid onto the species you think the
seller has over-estimated!
Are bids skewed?
Athey and Levin (2001) show that there is a lot of
skewed bidding in Forest Service auctions.
FS revenue is about $10,000 less per sale than if
total bids were the same but balanced.
But, in sales when FS made a mistake, total bids
were higher, so there is no correlation between FS
mistakes and actual revenue.
Explanation?
Bidders appear to game the auction by making bids for
which they expect to pay less.
But since theyre all doing it, they have to bid more to win
they compete away the potential profit from gaming!
Overbidding at eBay
Lee and Malmendier (2010) document
excessive bidding in eBay auctions.
Look at auctions for a game: Cashflow 101.
Available from two eBay retailers for $129.95.
Auction prices on eBay can exceed this:
42% of auctions in their data!
73% if one accounts for shipping costs.
What should we make of this?
Source: Lee and Malmendier (2010)
Overbidding at eBay, cont.
Does everyone overbid?
Appears to be a small number of bidders: only
17% of bidders ever bid above the retail price.
Maybe auctions let you fish for fools?
How general is the phenomenon?
Appears to be relatively common in other
categories, not a specific Cashflow 101 effect.
So again, what do we make of it?
Swoopo Auction
Swoopo sells common products by auction.
Auction rules
Placing a bid costs $0.50.
Price starts at $0, and clock starts to run down.
If there are bids before clock runs out, new round.
In each new round, price increases by $0.10.
If clock expires, auction ends. Standing high
bidder is the winner and pays the current price.
Excess Bidding
Data from 650 auctions of $50 bills
Average revenue: $104!

More generally, Augenblick examines data
from over 100,000 Swoopo auctions.

Excessive bidding?
Ending Times: Theory vs Data
Revenue Equivalence
Background
We saw that (in theory) some common auction
designs all lead to efficient outcomes and yield the
same revenue, at least on average.
Next:
How general is this result?
Why is this the case?
What are the implications?
and then a bit of empirical evidence.
Canonical model
Potential buyers
Two bidders.
Bidder i has value v
i

Each v
i
drawn from uniform distribution on [0,1].

Our next result also applies with N bidders so long
as their values come from the same distribution
(doesnt have to be a uniform distribution).

Seller gets to set the auction rules.
Revenue equivalence theorem
Thm. Consider the model on the last slide and any
auction game with the feature that in equilibrium,
the bidder with the highest value wins, and
if a bidder has the lowest possible value, he pays nothing.
The average revenue and bidder profits in this auction
game are the same as in the 2
nd
price auction.

Examples: the four main auctions from before
Ascending auction
Second price sealed bid
Descending auction
First price sealed bid
Envelope Theorem
Consider a parameterized maximization problem
Maximization means that when b=b*(v)
The envelope theorem says that
) ), ( ( ) , ( max ) (
*
v v b u v b u v U
B b
= =
e
( ) v v b u v U
v
), ( ) ( '
*
=
0 ) , ( = v b u
b
By the chain rule for differentiation
) ), ( ( ) ( ' ) ), ( ( ) ( '
* * *
v v b u v b v v b u v U
v b
+ =
Proof of the RET
Consider auction game each bidder has to submit a
bid b. (sealed bid, stopping price..)
Bidder with value v solves
Let U(v) be the bidders expected profit, i.e. the value
of the bidding problem. The Envelope Theorem says
So the bidders expected profit is
( ) v d v Win U v U
v
~ ~
| Pr ) 0 ( ) (
0
}
+ =
( ) ] | [ | Pr max b Payment E b Win v
B b

e
( ) ) ( | Pr ) ( '
*
v b Win v U =
Math behind the RET, cont.
So, independent of the exact auction rules (how bids
determine the winner and payments), a bidders
expected profit depends only on his equilibrium
probability of winning as a function of his value.
So if some auction game always results in the same
allocation (same winner) as the 2nd auction, it must have
the same expected bidder profits.
It must also have the same expected revenue. Why?
Revenue equals the total surplus minus the sum of the bidder
profits, and both of these are the same across auctions!
( ) v d v Win U v U
v
~ ~
| Pr ) 0 ( ) (
0
}
+ =
Revenue Equivalence Recap
Bidder values drawn
from same distribution.
Standard auction with
equilibrium in which
high value bidder wins.
Expect payoff if value v
( ) *] | [ * | Pr ) ( b Pay E b Win v v U =
Envelope argument
( ) ) ( | Pr ) ( ' v b Win v U =
Using calculus
RET: if two auctions
would result in the
same winner, they have
the same bidder profits,
and hence seller
revenue.
}
}
+ =
+ =
v
v
dx x Win U
dx x U U v U
0
0
) | Pr( ) 0 (
) ( ' ) 0 ( ) (
Hidden assumptions
Bidders know their own values
What if they may want to resell the good?
What if another bidder has relevant information?
Bidder values are independent.
Bidder i and j will have same belief about ks value.
What if bidders conduct surveys of the same oil field?
Bidders care about their auction profits
Value of the object minus the price they pay.
If opponent wins, dont care about what it pays.
No risk aversion if auction outcome is uncertain.

RET as a benchmark
The revenue equivalence theorem is a central result in
auction theory, but it depends on strong assumptions.
Many important results in economics are similar in their
dependence on implausible assumptions
Modigliani-Miller capital structure theorem
Ricardian equivalence theorem
First & second welfare theorems
Coase theorem
What is the significance of these kinds of results?
Evidence on RET
Does the RET Really Hold?
Not in laboratory experiments
Bidders behave as predicted in ascending
auctions, but tend to bid too much in first and
second price auctions, and descending auctions!
There is some evidence, however, that this goes
away with more experienced bidders.

What about in actual markets?
USFS Timber Auctions

The US Forest Service sells timber from the
national forests sometimes by open auction
and sometimes by sealed bid.

Athey, Levin and Seira (2008) use data from
these auctions to test predictions of the RET.
Data from Timber Auctions
Empirical Test
Tracts of timber sold by open and sealed bid auction
are not identical
California: small sales are usually sealed and large sales
usually open have to control for auction size.
Idaho/Montana: format was partially randomized but varied
by location and date of sale have to control for date/area.
Solution: estimate regression model
where outcome can be participation, revenue, etc.
i i i i
X sealed y c | o + + =
Estimated effect of sealed bidding
More loggers at sealed bid auctions!
Logger more likely to win
a sealed bid auction
Revenue advantage for sealed bidding?
Explaining the results
Basic RET assumed symmetric bidders, but here we
have small loggers and large mills competing.
What happens with weak and strong bidders?
Open auction: outcome is still efficient high value wins.
Sealed auction: strong bidders may shade bids a lot, and
so weak bidders win more often.
So theory predicts weak bidders will do better
(attend more, win more) with sealed bidding.
Combining theory and data
Athey-Levin paper asks
Suppose mills draw their values from one distribution and
loggers from another: what distributions are consistent with
the data?
Can do this for the open and/or the sealed bid auctions.
How do we infer value distributions from bidding data?
One approach: pick a distribution of values, simulate a lot
of auctions, see if the simulation data is similar to the
actual data; iterate until you find a good fit.
In practice, can be a bit more nuanced.
Findings
California
The data matches the theory very closely.
Idaho/Montana
The data doesnt match the theory as wel: the prices
in the open auctions are too low relative to what the
theory predicts.
Why? Maybe bidders arent competing hard in the
open auctions they bid as if they have low values.
Why might open auctions be vulnerable to collusion?
Applying the RET
Applying the RET
The basic idea of the RET is that so long as
the equilibrium of the auction leads to the
high value bidder winning
The bidders will make the same expected profits
as in the 2
nd
price auction
The seller will make the same expected revenue
as in the 2
nd
price auction
Now lets consider some applications
Solving the first-price auction
Two bidders and values U[0,1].
In a 2
nd
price auction, a bidder with value v will win with
probability v and expects to pay v/2 when he wins.
Suppose the first price auction has an equilibrium where
bidders use an increasing bid strategy b(v), so the high
value bidder wins.
A bidder with value v must expect to pay the same as in a
2
nd
price auction, i.e. to pay v/2 if he wins.
In the first-price auction, if the winner with value v pays his
bid b(v) if he wins, so therefore b(v)=v/2.
All-pay auction
Everyone puts their bid in an envelope and mails the
money to the seller.
The seller keeps the bids, and gives the object to
the highest bidder.
What are the equilibrium bids?

Suppose theres an equilibrium in which a higher
value leads to a higher bid then high value bidder
will win, and we can use the RET to find out what
the bids must be!
Solving the all-pay auction
Two bidders and values U[0,1].
In a 2
nd
price auction, a bidder with value v will win with
probability v and expects to pay v/2 when he wins.
So his overall expected payment is v
2
/2.
Suppose the all pay auction has an equilibrium where
bidders use an increasing bid strategy b(v), so the high
value bidder wins.
A bidder with value v must expect to pay the same as in a
2
nd
price auction, i.e. v*v/2=v
2
/2
In the all-pay auction equilibrium, each bidder pays its bid
with probability 1, so the bid strategy must be b(v) = v
2
/2!
Bidding with Reserve prices
Again consider two bidders with values U[0,1].
Consider a second price auction where the seller
sets a reserve price r below which she will not sell.
A bidder with value below r wont bid,
A bidder with value v>r bids her value.
Suppose a bidder has value v>r.
Expects to win with probability v.
Conditional on winning, will pay either r or opponent
value.
Expects to pay (r/v)*r + [(v-r)/v]*(v+r)/2 if he does win.
Re-writing expected payment: r+[(v-r)/v]*(v-r)/2.
First price auction with reserve
Look for an equilibrium where high value bidder wins
if his value is above the reserve price.
By RET, in the first price auction equilibrium,
A bidder with v<r doesnt bid.
A bidder with v>r bids b(v)=r+[(v-r)/v]*(v-r)/2.
Reserve prices and revenue
Can a reserve price increase revenue?
Suppose there is just one bidder with value v
distributed uniformly between 0 and 1.
Setting a reserve price r means
No sale with probability r
Sale at price of r with probability 1-r
Want to maximize r(1-r) => r=1/2!
Same as the monopoly price for a seller facing a
linear demand curve Q(p)=1-p. Why?
Reserve prices, cont.
Suppose there are two bidders with values U[0,1].
Consider choosing r or slightly more r+A
Doesnt matter if both values above r+A or below r.
Increases revenue by A if one bidder has value above r+A
and one has value below, i.e. with prob. 2r(1-r)
Decreases revenue by r if one bidder has value between r
and r+A, and the other has value below, i.e with prob. 2Ar.
At the optimal r, small change cant matter, so we
must have A*2r(1-r)= r*2Ar.
Optimal reserve: r=1/2 same as with one bidder!
Reserve price and RET
The optimal reserve price is independent of the
number of bidders.
And by RET, its the same in the first and second
price auction (and ascending/descending).
Is it the same in the all-pay auction?
No but can still use RET logic.
If r* is the optimal reserve in the second-price auction, and
b(v) is the all-pay eqm, the optimal all-pay reserve is
R*=b(r*).
Buy-it-Now Auctions
Two bidders with values U[100,200]
Seller runs ascending auction where bidders can
buy-it-now at price p.
Claim: buy it now will be used at least some of the
time if p<50.
Suppose opponent never buys it now.
If a buyer with value v doesnt buy it now, he expects to win
with prob. (v-100)/200 and pay (100+v)/2, so expected
profit (v-100)
2
/400.
Buyer can buy it now for profit v-p, so best to buy-it-now if
v-p > (v-100)
2
/400 if p=150, buy-it-now if v>158.
Buy-it-Now Auctions
Buy it now equilibrium: each bidder tries to
buy it now if v>v*, otherwise waits and bids.
Here v* is such that v* bidder is just indifferent
between buying it now and waiting, given the
other bidder tries to buy it now if his v>v*.
Claim: introducing buy it now reduces profit.
With values drawn U[100,200], MR>0 for all v.
So optimal auction awards object to high bidder.
May not happen with buy it now: because both
bidders may try to buy it now, leading to coin flip!
RET and Bargaining
Consider bargaining situation as follows
Seller with value s drawn U[0,1]
Buyer with value v, also drawn from [0,1].

In a bargaining protocol, the seller and
buyer report their bids/offers and the
mechanism determines whether or not they
should trade and who pays/receives what.
Examples
Buyer, seller submit offers b, s (or can opt out)
If B,S opt in, trade at price 1/2 (posted price)
If b>s, trade at price b (buyer offer )
If b>s, trade at price s (seller offer )
If b>s, trade at price (b+s)/2 (split the difference)
If both bid, trade for sure at offer closer to 1/2 (arbitration)
Dynamic bargaining
Seller and buyer alternate making offers
Seller makes offers, buyer says when to accept
Actually fits into this framework if we think of bid b and offer
s as bargaining strategies rather than numbers.
Vickrey pivot mechanism
Seller and buyer make bid and ask offers.
The parties trade if and only if b>s.
If trade takes place,
The seller receives a price b
The buyer pays a price s
This mechanism is strategy-proof (optimal to offer
true value/cost) and leads to efficient trade.
But the payments dont balance whenever there is
trade, the mechanism requires a subsidy of v-s.
Vickrey mechanism
0 1
0 1
v
b v
Buyer: bid true value v
Buyer: bid b>v
If seller offer is here,
trade at price s<v
If seller offer is here,
dont trade
If seller offer is here,
trade at price s<v
If seller offer
is here, dont
trade
If seller
offer is
here,
trade at
price > v!
Cost of Vickrey Mechanism
If v>s, trade occurs
Gains from trade v-s
Subsidy required is also v-s
Expected outcomes
Expected gains from trade: E[max{v-s,0}] = 1/6
Expected buyer profit: 1/6
Expected seller profit: 1/6
Expected subsidy: 1/6
Efficiency in Bargaining?
Analogue of the RET for bargaining says that any bargaining
protocol that leads to efficient trade must give buyer and seller
the same profit theyd get under the Vickrey mechanism.
Theorem. There is no bargaining protocol in that always leads to
efficient trade and does not require an external subsidy!

Proof.
Under Vickrey mechanism, expected surplus is E[max{v-s,0}],
buyer profit is E[max{v-s,0}] and seller profit is E[max{v-s,0}].
These numbers are the same in any efficient mechanism.
So buyer and seller profit minus surplus created is E[max{v-s,0}],
which can only happen if there is an outside subsidy!
Auctions vs Bargaining
Recall Coases argument for FCC auctions.
If licenses are allocated according to some
inefficient process, parties may not trade to
realize the efficient allocation.
Myerson-Satterthwaite Theorem provides a
rigorous justification: bargaining with private
information is inherently inefficient!
But in many cases, it is possible to run an efficient
auction (in the symmetric U[0,1] case, any
standard auction will do!).
Optimal Auctions
Optimal auctions
A version of the RET says that the expected
revenue from any auction will be the expected
marginal revenue from the winning bidder.

The optimal auction would award the object to the
bidder with highest marginal revenue, conditional on
his marginal revenue being positive.
Reserve price example
Two bidders with values U[0,100]
Define P(q)=100-100q
Define MR(q)=100-200q
For a given value v, the corresponding quantity is the
probability of having a value more than v, so MR(v)=100-2(1-
v/100).
The auction maximizes the sellers revenue if
The bidder with highest value wins, and the object is awarded if
and only v>50.
A first-price auction or a Vickrey auction with reserve price of 50
is revenue-maximizing!
Optimal Auctions
Myersons Theorem. Across all auctions, the
following achieves the maximum expected revenue:
Bidders report their values to seller.
Bidder with highest MR wins, provided his MR>0.
Winner pays an amount equal to the least value he could
have reported and still won.

Notes
Payment rule ensures that auction is strategy-proof.
If bidders are symmetric, optimal auction is a standard
auction with a reserve price.
If asymmetric, optimal auction biases the allocation toward
bidders with high MRs, but perhaps lower values.
Asymmetric Bidders
Two bidders with values U[20,60] and U[0,80]
Bidder 1: P(q)=60-40q and MR(q)=60-80q
Bidder 2: P(q)=80-80q and MR(q)=80-160q
Marginal revenue as a function of value
If value is v, then q is prob. of higher value.
Bidder 1: MR(v) = 2v-60
Bidder 2: MR(v) = 2v-80
Revenue-maximizing auction (picture)
Optimal auction
Bidder 1 could have lower value, but
higher marginal revenue.
MR
MR
1
(v)=2v-60 MR
2
(v)=2v-80
80 40
0
-40
-80
80
40
v
1
v
2
*
If bidder 1 has value v
1
, he
wins if v
2
< v
2
*!
The optimal auction is biased toward bidder one!
Relation to Monopoly Theory
Two Stanford GSB professors, Jeremy Bulow and John Roberts,
wrote a famous paper that explained optimal auctions as identical to
monopoly pricing.
Each bidder is analagous a separate market in which the good can
be sold.
Quantity is analogous to probability of winning.
The monopolist can price discriminate, setting different prices in
different markets.
Allocating the probability of winning to individual bidders is
analogous to allocating quantities across separated markets.
The best strategy for a monopolist is to allocate quantity to the
market where marginal revenue is highest, and the same is true for
the auctioneer!
Marginal Revenue
There is a useful connection to monopoly theory.
Define a bidders demand at price p to be the
probability his value is above p
Define his inverse demand P(q) to be the price at
which there is probabilty q he will buy.
Example, a bidder has value U[20,60]
The bidders demand is P(q)=60-40q.
The revenue curve is R(q)=qP(q)=60q-40q
2
Define the bidders marginal revenue as dR(q)/dq.
If values are U[20,60], MR(q)=60-80q.
Marginal Revenue Recap
From RET analysis, bidder profits and seller
revenue are determined by the auction
allocation who wins as a function of the
bidder values.
Optimal auction design: choose the revenue-
maximizing allocation rule.
Optimal Auction Design
v
1

v
2

Bidder 1
wins
Bidder 2
wins
Efficient allocation: e.g. vickrey auction
What happens if
bidder 1 wins a
bit more often?
Marginal Revenue Recap
Q
P
P(q)
MR(q)
Probability bidder would
pay a price p defines
demand curve for a bidder
1 0
v
1-F(v)
MR(v)
MR(v)=marginal revenue
from allocating a bit more
quantity to the bidder (by letting
him win when his value is v-dv
rather than just above v)
Optimal Auction Design
v
1

v
2

Bidder 1
wins
Bidder 2
wins
What happens if
bidder 1 wins a
bit more often?
For each value v of
bidder 1 that gets a little
extra chance of winning,
seller gains MR
1
(v)
For each value v of
bidder 2 that gets a little
less chance of winning,
seller loses MR
2
(v)
Expected revenue
Starting from the case where the item is
never sold, seller gains MR1(v) when
allocates probabilty to type v of bidder 1, and
MR2(v) when allocates probability to type v of
bidder 2.
So expected revenue from the auction, is
average marginal revenue of the winner!
Optimal Auction Design
v
1

v
2

Bidder 1
wins
Bidder 2
wins
MR
1
(v
1
) =MR
2
(v
2
)
Optimal auction awards
item to bidder with the
higher marginal revenue.
Optimal Auction Design
v
1

v
2

Bidder 1
wins
Bidder 2
wins
MR
1
(v
1
) =MR
2
(v
2
)>0
MR
1
(v
1
)=MR
2
(v
2
)=0
MR
1
(v
1
)<0
MR
2
(v
2
)>0
Reserve price used to
ensure no bidder wins
with a negative
marginal revenue!
No sale
Summary
RET is powerful tool for thinking about auctions (or bargaining
situations) where bidders have private information.
With symmetric bidders many auctions have the property that
In equilibrium, the high value bidder wins.
Exp. revenue and bidder profit are identical to a Vickrey auction.
Auctioneer may want to distort efficiency to increase revenue
Reserve price even with symmetric bidders
Biasing the auction to favor high MR bidders.
Concerns we havent yet discussed
Getting bidders to show up
Getting bidders to bid competitively
Common Value
Auctions
Wallet auction

Whats for sale: the money in my wallet.
Second-price auction

How much will you bid?
Wallet auction

Whats for sale: the money in my wallet.
Second-price auction

What if one person gets to look in the wallet.
Does this change your bid? Why?
The winners curse
Winning the wallet means that everyone else
in the class was more pessimistic about its
contents.
Winning is bad news
If you had an initial estimate of $100, knowing that
everyone else had a lower estimate should cause
you to revise your estimate downward.
Equilibrium bidding should account for this.
Common value auctions
Imperfect estimate model
Two bidders with common value v
Bidder 1 has estimate s
1
= v+e
1

Bidder 2 has estimate s
2
= v+e
2
Errors e
1
,e
2
are independent
Second price auction

How should bidders account for the winners
curse in equilibrium?
Equilibrium bidding
Claim: In the equilibrium of the second price auction,
bidders use the strategy b(s
i
)=E[v|s
i
,s
j
=s
i
]
Proof.
Suppose b(s) is an equilibrium bid strategy. Then i prefers
to bid as if his value is s
i
as opposed to s
i
-ds or s
i
+ds.
Changing the bid a little will matter if and only if bidder j
bids exactly b(s
i
), that is, if j has the same value.
In order that i not want to slightly raise or lower his bid,
winning and finding out that j has exactly the same value
must result in zero profit.
Therefore b(s
i
)=E[v|s
i
,s
j
=s
i
].
Winners or Losers Curse?
Does accounting for the information of others mean
you bid higher or lower than your baseline estimate?
Case 1: 10 bidders, 1 item, second price auction.
Eqm: b(s) = E[v|s is tied for highest estimate] < E[v|s].
There is a winners curse.
Case 2: 10 bidders, 8 items, ninth price auction.
Eqm: b(s) = E[v|s is tied for eigth highest estimate] > E[v|s]
There is a losers curse!
Revenue Equivalence?
Milgrom-Weber: If bidders have correlated estimates of a
common value item, an open auction leads to higher revenue
than sealed bid auction.
Intuition.
In either case, the high estimate bidder will win.
In first price sealed bid auction, the payment will be independent
of the estimate of the second highest bidder.
In an ascending auction, the payment will depend on the second
highest estimate.
The bidder estimates are correlated.
So open auction sets the price for the winner high when the other
bidder has good new about the value, and sets it low when the
other bidder has bad news.
This reduces the winners profit and increases revenue!
Linkage principle
Broader principle: suppose the seller can give bidders access
to better information. Then revenue is increased on average
by making the information publicly available.
Intuition:
The public information can either increase or decrease
everyones bids (depending on if its good or bad news).
The public information will tend to be good news exactly when
the high bidder has a high value.
So releasing information squeezes the high bidder in the right
cases.
Information aggregation
How many miles is it to drive from New York
to Chicago?
Information aggregation
Suppose we have many bidders, and each
has an independent estimate s
i
= v+e
i.

Average of the bidders estimates s
i
is likely
to be a very good estimate of v. Why?

Avg(s
i
) = Avg(v+e
i
) = v+Avg(e
i
) v
.


The wisdom of crowds; Galton example.
Information aggregation
Suppose we have many bidders, and each
has an independent estimate s
i
= v+e
i.


If the bidders compete in a (first or second
price or ascending) auction, is the resulting
price a good estimate of v?
Answer of Stanford profs Wilson, Milgrom
In certain cases YES!, auction prices aggregate
information.
Information Aggregation
In certain cases, auction prices aggregate information, so that
the auction price is a good estimate of true value. Why?
Example:
Bidder estimates are s
i
= v+e
i.

Second price auction, b
i
= E[v| s
i
is tied for high s]
Order bids so that b
1
>b
2
>

>b
N

Price = b
2
= E[v| s
2
is tied for high s]
With many bidders then s
2
is tied for high s is basically a correct
hypothesis, so b
2
is an accurate estimate!
There is one hidden assumption, which bidders with very
extreme estimates are in fact quite confident about the value.
Information Aggregation
Signals
if true value
is high
Signals
if true value
is low
Very high
signal is very
informative!
This signal is also very informative
if you assume its highest.
Common values in practice
What kinds of auction have a common value
flavor to them?
Treasury bill auctions or other financial assets everyone
may have a guess about the trading price after the
auction, but no one knows for sure.
Timber auctions? what kind of timber is actually out there
on the tract thats being sold.
Oil lease auctions: oil is under the Gulf of Mexico, bidders
do independent seismic studies each has valuable
information.
OCS Auctions
The US government auctions the right to drill for oil
on the outer continental shelf.
Value of oil is similar to the different bidders, but no
one knows how much oil there is, or if theres none.
Prior to the auction, the bidders do seismic studies.
Two kinds of sale
Wildcat sale - new territory being sold
Drainage sale - territory adjacent to existing tract.
These are like the wallet auctions we ran in class!
Wildcat vs Drainage
Drainage sales
Explaining the Results
Comparing wildcat and drainage sales
Wildcat sales yield low profits => competition.
Drainage sales are profitable, but only for
insiders => insiders have an advantage.
Next slides (which we wont cover today)
work out equilibrium bidding for an
insider/outsider model.
This model fits the drainage data quite well.
The symmetric model fits the wildcat data!
Drainage sales
Asymmetric information model
Neighbor knows true value v.
Outsider believes v is U[0,1].
In ascending auction, outsider drops out at zero!
What about first price sealed bid auction?

Captures the idea that the bidder with the
neighboring tract has superior information.
Bidding in drainage model
There is no equilibrium in which the outsider uses a
pure strategy.
Suppose the neighbor expects the outsider to bid b.
Neighbor will bid b if b>v, or else less than b.
So the outsider wins and pays b if and only if v<b.
Therefore b>0 cant be optimal, and there cant be an
equilibrium in which the outsider bids b>0 for sure.
There also cant be an equilibrium with b=0. The neighbor
would also bid zero (or epsilon), and the outsider could
profitably bid just a bit more and win for sure.
Mixed strategy equilibrium
In equilibrium, the outsider randomizes between different bids
each bid gives him an expected payoff of zero.
Outsider bids less than x with probability P(x)=2x.
Neighbor bids b(v)=v/2 if value is v.
Features of the equilibrium
Bids of both bidders are uniformly distributed on [0,1/2].
The neighbor wins exactly half the time, but tends to win when
v is high, and makes positive profit.
The outsider also wins half the time, but makes zero profit on
average.
Verifying the equilibrium
Outsider profit (equals zero for all bids x)
Profit from bidding x: t(x)=Pr(b(v)<x)(E[v|b(v)<x]-x)
Define g(x)=b
-1
(x): t(x)=Pr(v<g(x))(E[v|v<g(x)] x)
So, t(x)=g(x)*[g(x)/2-x]=0
Therefore g(x)=2x and b(v)=v/2.
Verifying the equilibrium
Neighbors problem
Given v, choose bid b to maximize P(b)(b-v)
Neighbors first order condition: b=v-P(b)/P(b)
But we know b(v)=v/2, so therefore P(b)=2b.

Testing equal distributions
Summary
Many auctions have a common value flavor.
In common value auctions
Auctions that aggregate participant information
into the price can yield more revenue.
If there are many bidders, the resulting price can
be a useful indicator of the items value.
The distribution of information is very important.