Auction Markets
Jon Levin
Winter 2010
Economics 136
Multi-Unit Auctions
Examples









Treasury auctions
Auction-rate securities
IPO auctions
Privatization
Electricity markets
Asset sales
Condominium sales
Wine/Art/Antiques
Auto auctions









Natural resources
Radio spectrum
Emissions permits
Airport landing slots
Bus routes
Procurement contracts
Sponsored search
Internet display ads
eBay marketplace
Sequential auctions

Auction houses often sell identical goods
sequentially (e.g. lots of wine).

What happens at sequential auctions?


Should you bid your value in the first auction?
Are early prices higher or lower than later prices?
Sotheby Wine Auctions
Source: Ashenfelter (1989, Journal of Economic Perspectives)
Declining Prices
A puzzle?

Standard theory: in a symmetric private value
setting, prices need not be equal across sequential
first or second-price auctions, but…

Weber’s Theorem. Equilibrium prices should follow a
random walk: E[pt+1|p1…pt]=pt

Yet the “declining price anomaly appears to be quite
robust – wine, art, cattle, etc – and variants
observed with non-identical items.

This remains something of an open puzzle.
Simultaneous Sales of
Identical items

Consider auction for k identical items.

Possible “one-shot” auction methods


“Uniform price” (clock and sealed bid)

“Discriminatory price” (pay-your-bid and Vickrey).
We will see that one important issue is whether
bidders want just one item, or are potentially interested
in winning several items.
“Uniform price” auctions

Sellers often want to run an auction in which all
winners pay the same “uniform” price.


Perceived as “fair”; achieves “price discovery”
Uniform price formats

Clock auction: seller announces a sequence of prices and
bidders name quantities until a market-clearing price is found
and auction ends.

Sealed bidding: participants bid a price-quantity schedule and
bids are used to determine the uniform market-clearing price.
British CO2 Auctions

Greenhouse Gas Emissions
Trading Scheme Auction,
United Kingdom, 2002.

UK government aimed to
spend 215 million British
pounds to get firms reduce
CO2 emissions.

Clock auction used to determine

What price to pay per unit?

Which firms to reward?
Greenhouse Auction Rules

Auctioneer calls out price


Each round, bidders state tons of CO2 they will abate



Price starts high and decreases each round.
Tons abated can only decrease as prices decrease.
Auctioneer multiplies tons of abatement times price.

If total cost exceeds budget, lowers the price

When total cost first falls short of budget, auction ends and
that allocation is implemented
Auction results

38 bidders (34 winners), 4m metric tons of CO2 reduction.

Price per metric ton: £215m/4m= £53.75
Graphical treatment
P
p1
UK “Demand Curve, defined
so that Q*P(Q)=£215m
p2
Falling prices
trace out a
“supply curve”.
p*
Q
Sealed bid version

Uniform-price sealed bid auction

Auctioneer posts its demand curve

Bidders submit “supply curves” - i.e. how
much they will supply at each price.

Individual supply curves are aggregated
to form an aggregate supply curve.

Price is set so that supply = demand.
Strategic equivalence?


Are clock and sealed auctions strategically equivalent?

Suppose bidders in the clock auction observe only the
prices and that prices decline in a fixed sequence.

Bidders are then effectively being asked to reveal their
supply curves from the top down, with no new
information each round other than that the current
price is relevant.

So yes, the clock auction is strategically equivalent to
a sealed bid auction in which supply curves get written
down in advance.
Equivalence may fail if more information is revealed
each round.
Incentives with Uniform Price




Suppose each bidder wants a single item.
Values are drawn from U[0,1].
n bidders, k items with n>k.
Bidders submit bids: price = k+1 highest bid.
Theorem. For a bidder with single item
demand, it is a dominant strategy to bid one’s
value.

Proof. Similar to the second price case.
Demand reduction

Example: three items for sale





Bidder 1: value 120 and wants 1 item.
Bidder 2: value 110 and wants 1 item.
Bidder 3: value 100 and wants 1 item.
Bidder 4: value 105 and wants 2 items.
Consider what happens with “truthful” bidding



Bids are 120, 110, 105, 105, 100.
Three highest bids are winners
Fourth highest bid is 105 => winners pay 105 each.
Demand reduction

Example: three items for sale





Bidder 1: value 120 and wants 1 item.
Bidder 2: value 110 and wants 1 item.
Bidder 3: value 100 and wants 1 item.
Bidder 4: value 105 and wants 2 items.
“Demand reduction” by bidder 4



If he bids 105, he wins 1 item and pays 105.
If he bids 101, bids are 120, 110, 101, 101, 100.
He still wins 1 and lowers the price to 101!
Example, cont.


Again, three items and values

Bidder 1: 120

Bidder 2: 110

Bidder 3: 100

Bidder 4: 115 and wants two units.
Demand reduction by bidder 4

Bid 115 for both units => wins two, price =110.

Bid 115 for first unit, 100 for second => wins 1, p=100.

Bidder four optimally exercises “market power”.
Demand Reduction Picture
P
b1
Supply
Q=3
Opponent Bids
and demand curve
b2
“Residual supply curve”
Multi-unit bidder wants to
maximize profit by behaving
as a monopsonist against the
residual supply curve.
b3
b4
3
Q
Low Price Equilibrium
P
Supply Expanded supply
“Residualsupply
Residual
supplycurve
curve”in a
in a price
low
“regular”
equilibrium
equilbrium
Residual supply curve if
seller “adds elasticity”
3
Q
Multi-Unit Auctions and
Financial Assets
Today’s Lecture

Uniform price sealed bid auctions



Comparison to alternatives




Virtues: simple, “fair”, reveal “market price”
Concerns: demand reduction, “low price” eqm
Discriminatory price auctions
Vickrey auction
Application to financial markets
Extensions to multiple goods
Uniform-price sealed bid auction

Auctioneer posts its supply curve

Bidders submit “demand curves” - i.e. how
much they want at each price.

Individual demand curves are aggregated
to form an aggregate demand curve.

Price is set so that supply = demand

Demands at the market clearing price are
satisfied.
Incentives with Uniform Price





N bidders, K items with N>K
Each bidder wants one unit, values U[0,1].
Bids submitted (offer to buy one unit at some price)
Market clearing price = any price between the kth highest bid and
the K+1th highest bid (why?)
Assume lowest clearing price: K+1 highest bid.
Theorem. For a bidder with single item demand, it is a dominant
strategy to bid one’s value.
 Proof. Similar to the second price case.

Theorem is not true if bidders want multiple units (next slide!)
Demand reduction

Example: three items for sale





Bidder 1: value 120 and wants 1 item.
Bidder 2: value 110 and wants 1 item.
Bidder 3: value 100 and wants 1 item.
Bidder 4: value 105 and wants 2 items.
Consider what happens with “truthful” bidding



Bids are 120, 110, 105, 105, 100.
Three highest bids are winners
Fourth highest bid is 105 => winners pay 105 each.
Demand reduction

Example: three items for sale





Bidder 1: value 120 and wants 1 item.
Bidder 2: value 110 and wants 1 item.
Bidder 3: value 100 and wants 1 item.
Bidder 4: value 105 and wants 2 items.
“Demand reduction” by bidder 4



If he bids 105, he wins 1 item and pays 105.
If he bids 101, bids are 120, 110, 101, 101, 100.
He still wins 1 and lowers the price to 101!
Example, cont.


Again, three items and values

Bidder 1: 120

Bidder 2: 110

Bidder 3: 100

Bidder 4: 115 and wants two units.
Demand reduction by bidder 4

Bid 115 for both units => wins two, price =110.

Bid 115 for first unit, 100 for second => wins 1, p=100.

Bidder four optimally exercises “market power”.
Demand Reduction

General model




N bidders, K items, where 2k<N.
Each bidder has positive value for at least two items.
Bidder values random, but always satisfy v1j ≥ v2j
Theorem. In the k+1 price auction, it is weakly
dominant to bid true value for first unit. However:


there is no equilibrium in which bidders all bid their full
values for both items, and
there is no equilibrium in undominated strategies that is
efficient for all realized valuations.
“Low price” Equilibria

With fixed supply, uniform price auction can have
equilibria with low prices due to demand reduction.

Example: three units, three bidders.





Bidders value units at 10, want as many as possible.
The price “should be” 10 if there is competition.
What if each bidder offers to buy one unit for 10, and no
additional units at any price above zero.
Bidders split the units, price is zero!
Because a bidder who wants to purchase additional units
has to pay ten, there is no reason to demand more. The
low price bidding is a Nash equilibrium!
Making supply elastic

Suppose the seller offers



Claim: Any Nash equilibrium for bidders involves
selling four units at a price of at least 4.




To sell 3 units at any price
To sell 4 units if (and only if) price exceeds 4
If bidders bid (10,0), each bidder gets 1 unit, p=0.
Each bidder has incentive to bid (10,4) => win two units
and price increases to four. Profit of 2*(10-4)=12>10.
Equilibrium will have one bidder winning two items, and
fifth highest bid somewhere between 4 and 10.
Somewhat surprisingly, seller has managed to
increase supply and yet also increase prices.
California electricity crisis

The California electricity crisis of 2001
 Prices go from around $45 megawatt-hour to as high as $1400.
 Paul Joskow of MIT: “California electricity crisis is what happens
when a vertical supply curve intersects a vertical demand curve.”
 Steep supply/demand: during the crisis a 5% lowering of demand
(or increase in supply) would have lowered prices by 50%!
 Borenstein, Bushnell, Wolak (AER, 2003): vertical demand
because consumers don’t react to price, vertical supply because
generators strategically withhold power.

Remedies?
 Create elasticity in electricity demand (how?).
 Restrict slope of submitted supply curves.
 Encourage build-out of additional capacity (how?).
 Forward contracts (unravel the market!).
Increasing Returns…

Discussion implicitly focused on bidders with
decreasing marginal values who submit downwardsloping demand curves.

What if there are scale economies?





Two units for sale
Bidder 1 offers 10 for one unit.
Bidder 2 offers 5 for first unit and 11 for second.
There is no uniform price that clears the market!
Not much is known about performance of uniform
price auctions where there are scale economies.
Expanding the example…

Two units for sale
 Bidder 1 offers 10 for one unit.
 Bidder 2 offers 5 for first unit and 11 for second.

Consider possible prices
 At p < 8, demand = 3
 At p = 8, demand = 1 or 3
 At p in (8,10), demand = 1
 At p = 10, demand = 0 or 1
 At p > 10, demand = 0

Problem: demand “crosses” supply at price = 8, but at that price,
Bidder 1 demands 1, and Bidder 2 demands 0 or 2 but is
unwilling to buy 1! So can’t have demand=supply!!!
Discriminatory Price Auctions




Alternative is a “pay-your-bid” format
Bidders submit bids (offers to buy different
quantities at different prices)
Seller finds price where supply=demand
All bids above clearing price are satisfied, but
winners pay their bid rather than the clearing
price.
Example

Example: three items for sale
 Bidder 1: value 120 and wants 1 item.
 Bidder 2: value 110 and wants 1 item.
 Bidder 3: value 100 and wants 1 item.
 Bidder 4: value 105 and wants 2 items.

Suppose “truthful bids”
 Bidder 1: (1, 120) “buy one at any price < 120”
 Bidder 2: (1, 110)
 Bidder 3: (1, 100)
 Bidder 4: (2, 105)

Outcome: 1, 2, 4 win and pay 120, 110, 105.
Is this an equilibrium? Why or why not?

Example, cont.

Example
 Bidders 1, 2, 3 want 1 item and values 120, 110, 100.
 Bidder 4: wants two items and value 105.

Possible equilibrium bids?
 Bidder 1: (1, 105)
 Bidder 2: (1, 105)
 Bidder 3: (1, 100)
 Bidder 4: (2, 105)

So 1, 2, 4 win and all pay 105. Is this an eqm? Why or why not?
Example, cont.

Example



Bidders 1, 2, 3 want 1 item and values 120, 110, 100.
Bidder 4: wants two items and value 105.
Possible equilibrium bids?




Bidder 1: (1, 100 + penny)
Bidder 2: (1, 100 + penny)
Bidder 3: (1, 100)
Bidder 4: (2, 100 + penny)

So 1, 2, 4 win and all pay 100 + penny. Is this an eqm?

The actual equilibrium involves mixed strategies with bids
distributed between 100 and 105!
Uniform vs. Discriminatory

Both auctions can be inefficient.


Does one lead to higher prices?


Both auctions create an incentive to “reduce demand” if
bidders want multiple units.
Not clear in general.
Does one lead to higher or lower participation or
reveal more useful information?

Sometimes argued that uniform price auction is good for
small bidders b/c its easy to participate and get the “market
price”, but discriminatory price can sometimes be painful
for large high-value bidder.
US Treasury Experience



US Treasuy historically ran discriminatory auctions
(since the 1970s) to sell bonds.
Beginning in 1992, switched to uniform price.
Rationale




Aim for more liquid market (transparency)
Encourage broader participation
Encourage competition (slightly vague)
Features of the market: large, highly liquid



There is some “price impact” from the new issuance.
There is also a “when-issued” market that runs prior to the
auction, so participants can guage likely price.
Many participants (75-85 bidders), but relatively small
number of primary dealers win a lot of the bonds.
Treasury experience, cont.

Evidence from US transition




Switch to uniform price led to somewhat lower spread
between auction price and WI price (but not very large or
stat. significant).
Somewhat more volatility between auction price and WI
price b/c more dispersed bids w/ uniform price auction.
Awards to primary dealers similar under the two types of
auctions, but share of awards to the very top dealers
decreased with uniform price.
Debate in other countries: possible that design may
matter more if market is thinner or less transparent.
TARP Warrant Auctions

As part of TARP, Treasury received warrants from
banks that were “bailed out”




Warrants give holder a right to buy the stock at some
“strike price” at any point over the next 10 years.
Like an option except that when a warrant is exercised, the
firm issues new shares rather than buying back shares (so
there is dilution)
Treasury negotiated with banks to sell them back the
warrants but some negotiations failed, leaving treasury to
dispose of the warrants.
Question: how to design an auction to sell the
warrants?
Warrant auctions, cont.

Questions one must address





Uniform price or discriminatory?
Sell all warrants at once, or over time?
Sealed bid or ascending auction?
Potential for a “winners curse”
Treasury (via auction agent) decided on a standard
treasury format, ran three in fall.


Evidence from JPM auction (largest at $1 bn) suggests a
large price impact (auction price 30% below subsequent
after-market price.
Now treasury must consider whether this was a big
number, and whether to change the design. What data
would you want to look at?
An Efficient Auction?


Back to our example with three items for sale

Bidder 1: value 120 and wants 1 item.

Bidder 2: value 110 and wants 1 item.

Bidder 3: value 100 and wants 1 item.

Bidder 4: value 105 and wants 2 items.
Is there a pricing rule that would make it a dominant
strategy for each bidder to bid truthfully – and would
lead to an efficient allocation of the items?
“Vickrey” Prices

Set price for each bidder equal to the value per unit
the bidder “displaces.

Equivalently: use submitted values to compute total value
with and without the bidder present.

Set price for the bidder so that his profit equals the value
he creates.

Vickrey pricing makes bidding truthfully a dominant
strategy.

But Vickrey prices are not uniform prices!
Vickrey pricing

Example: three items for sale





Bidder 1: value 120 and wants 1 item.
Bidder 2: value 110 and wants 1 item.
Bidder 3: value 100 and wants 1 item.
Bidder 4: value 105 and wants 2 items.
Vickrey pricing if truthful bids


Bidders 1 and 2 win, pay 105 each (displace 4).
Bidder 4 wins one unit, pays 100 (displaces 3).
Vickrey pricing, again

Example: three items for sale





Bidder 1: value 120 and wants 1 item.
Bidder 2: value 110 and wants 1 item.
Bidder 3: value 100 and wants 1 item.
Bidder 4: value 115 and wants 2 items.
Vickrey pricing if truthful bids


Bidder 1 wins and pays 110 (displaces bidder 2).
Bidder 4 wins two units, pays 100 for first unit (displaces
bidder 3) and 110 for second (displaces bidder 2).
General Lessons

Uniform price auctions have fairness, transparency virtues

But encourage demand reduction when bidders want multiple
units

Also create the possibility of low price “collusive seeming”
equilibria (making supply elastic can help with this problem).

Discriminatory price auctions also encourage demand
reduction, but sometimes viewed as a way to raise revenue
from high value bidders – although revenue implications
generally unclear.

Vickrey pricing can eliminate demand reduction and restore
efficiency, but the uniform price property is lost.
Multi-Item Auctions
and Matching
Multiple Kinds of Goods

Can we design successful auctions that allow
for multiple kinds of goods?





Spectrum licenses covering different cities.
Electricity delivered from/to different places.
Multiple kinds of financial assets.
Emissions reductions in different years.
Different kinds of sponsored search placements.
General issues


What would be desirable properties?

Auction finds “market clearing” prices (uniform price)

Auction is strategyproof (Vickrey), or nearly so.

Auction is relatively simple, robust to collusion, etc.
The challenge

Bidder preferences may be complex and complex
preferences are hard to state in sealed bid auction.

Complementarities (like increasing returns) may imply that
market clearing prices don’t exist.

Auction complexity and strategy can be serious issues.
Connection to matching

There is a connection to matching theory…




Treat bidders as one side of the market
Treat items for sale as the other side
We are interested in a matching (for now, one-toone, but potentially many-to-one).
But now we have to determine prices as well as
the assignment.
Deferred acceptance?

Each bidder submits a preference list


Example: first choice is to pay zero for item 1, second
choice is to pay $1 for item 1, third choice is to pay $0 for
item 2, fourth choice is to pay $2 for item 1, etc..
Seller runs deferred acceptance algorithm

Bidders “propose” to the items.

Items prefer to sell for more money, accept highest offer.

Algorithm will eventually terminate.
Auctions & Matching

Matching algorithm
(Gale & Shapley, 1962)
Offers made by computer.
1. Doctors apply to most
preferred remaining
acceptable program.
2. Hospitals hold best doctor,
reject others.
3. Rejected doctor strikes the
hospital from his/her list.
4. Process continues until no
new offers or rejections.
5. Implement last held
allocation.

Ascending auction
(Kelso & Crawford, 1982)
“Bids” made by computer.
1.
Bidders offer most
preferred remaining
acceptable purchase.
2.
Items hold best bid, reject
others.
3.
Rejected bidder strikes
offer from his/her list.
4.
Process continues until no
new offers or rejections.
5.
Implement last held
allocation.
Deferred acceptance “auction”



What we know from matching theory

Suppose bidders are each interested in a single item.

Algorithm will converge to a “stable” allocation.

Bidder-offering auction is strategy-proof for the bidders.
Connection to auctions/markets

Stability: at the deferred acceptance final price for each item,
there is exactly one bidder, so supply = demand!

A stable allocation is a competitive equilibrium

Strategy-proof: the competitive equilibrium prices are the same
as one would get from a Vickrey auction!
Extension to multi-good demand?
Sponsored Search
Auctions
Sponsored Search Auctions

Google revenue in 2008: $21,795,550,000.

Hal Varian, Google chief economist:


“What most people don’t realize is that all that money
comes pennies at a time.”
Today we’ll discuss internet keyword auctions.

References: Varian 2008, Edelman et al. 2007.
Keyword Auctions



Advertiser submit bids for keywords

Offer a dollar payment per click.

Alternatives: price per impression, or per conversion.
Separate auction for every query

Positions awarded in order of bid (more on this later).

Advertisers pay bid of the advertiser in the position below.

“Generalized second price” auction format.
Some important features

Multiple positions, but advertisers submit only a single bid.

Search is highly targeted, and transaction oriented.
Brief History of Sponsored
Search Auctions




Pre-1994: advertising sold on a per-impression basis,
traditional direct sales to advertisers.
1994: Overture (then GoTo) allows advertisers to bid for
keywords, offering some amount per click. Advertisers
pay their bids.
Late 1990s: Yahoo! and MSN adopt Overture, but
mechanism proves unstable - advertisers constantly
change bids to avoid paying more than necessary.
2002: Google modifies keyword auction to have
advertisers pay minimum amount necessary to maintain
their position (i.e. GSP)- followed by Yahoo! and MSN.
Example

Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Efficient allocation




Advertiser 1 gets top slot: value created 200*10 = 2000
Advertiser 2 gets 2nd slot: value created 100*4 = 400
Total value creation: $2400
Example: competitive eqm

Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Competitive equilibrium



Example: p2 = 2 and p1 = 4




Set prices for slots (p1, p2) so that demand = suppy
Advertiser 3 demands nothing
Advertiser 2 demands slot 2: 100*(4-2)>200*(4-4)=0
Advertiser 1 demands slot 1: 200*(10-4)>100*(10-2)
Efficient outcome with revenue: $800+$200= $1000
Example: competitive eqm

Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Many possible mkt-clearing prices (p1, p2)





Advertiser 3 must demand nothing, so p1, p2  2
Advertiser 2 must demand slot 2, so
 p2  4, and 2p1- p2  4 (so that 200*(4- p1) < 100*(4-p2) )
Advertiser 1 must demand slot 1, so
 2p1- p2  10 (so that 200*(10- p1) < 100*(10-p2) )
Allocation efficient with revenue: 200p1+100p2
Competitive Eqm
p1
Possible competitive
equilibrium prices!
8
6
Can we use an
auction to “find”
these prices?
4
2
2
4
p2
Example: Pay-your-Bid

Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Overture auction (pay your bid)






Advertiser 3 will offer up to $2 per click
Advertiser 2 has to bid $2.01 to get second slot
Advertiser 1 wants to bid $2.02 to get top slot.
But then advertiser 2 wants to top this, and so on.
Pay your bid auction is unstable!
Overture bid patterns

Edelman & Ostrovsky (2006): “sawtooth” pattern
caused by auto-bidding programs.
Overture bid patterns, cont.
Vickrey Auction


Bidders submit bids (price per click)
Auctioneer treats bids as values



Finds allocation that maximizes value created
So high bid gets top slot, and so forth
Vickrey payment for bidder j



Note that if bidder j gets a slot, he is “displacing” other
bidders by moving them down a slot.
Compute the lost value from this displacement (e.g. if j
pushes k down a slot, k loses clicks that are worth some
amount to k)
Bidder j’s payment equals total “displacement” cost, or
equivalently the externality j imposes on other bidders.
Vickrey Auction, cont.

Second price auction is a Vickrey auction




Winner “displaces” second highest bidder
Winner pays the displaced value: 2nd high bid
Also a Google auction with one click for sale!
General properties of Vickrey auction


Dominant strategy to bid truthfully (bid = value)!
Outcome is efficient (maximizes total value)!
Vickrey auction, aside

Recall our example with three items for sale





Bidder 1: value 120 and wants 1 item.
Bidder 2: value 110 and wants 1 item.
Bidder 3: value 100 and wants 1 item.
Bidder 4: value 105 and wants 2 items.
Vickrey pricing if truthful bids


Bidders 1 and 2 win, pay 105 each (displace 4).
Bidder 4 wins one unit, pays 100 (displaces 3).
Example: Vickrey auction


Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Vickrey auction
 Advertisers are submit bids, assigned efficiently given submitted
bids, and have to pay the value their ad displaces.
 Dominant strategy to bid one’s true value.

Vickrey outcome
 Advertiser 1 gets top, then 2, and 3 gets nothing.
 Adv. 2 pays $200 (displaces 3) for 100 clicks, or $2 /click.
 Adv. 1 pays $600 (displaces 3 & 2) for 200 clicks, $3 /click.

Efficient allocation with revenue of $800.
Deriving the Vickrey prices


Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Vickrey payment for Bidder 2
 Bidder 2 displaces 3 from slot 2
 Value lost from displacing 3: $2 * 100 = $200
 So Bidder 2 must pay $200 (for 100 clicks), or $2 per click.

Vickrey payment for Bidder 1
 Displaces 3 from slot 2: must pay $200
 Displaces 2 from slot 1 to 2: must pay $4*(200-100)=$400
 So Bidder 1 must pay $600 (for 200 clicks), or $3 per click.

Vickrey “prices” are therefore p2 = 2 and p1 = 3, revenue $800.
Vickrey prices
p1
Vickrey prices are
the lowest competitive
equilibrium prices!
8
6
4
2
Vickrey prices
2
4
p2
Google “GSP” Auction

Generalized Second Price Auction




Bidders submit bids (per click)
Top bid gets slot 1, second bid gets slot 2, etc.
Each bidder pays the bid of the bidder below him.
Do the bidders want to bid truthfully?
Truthful bidding?

Not a dominant strategy to bid “truthfully”!

Two positions, with 200 and 100 clicks.

Consider bidder with value 10

Suppose competing bids are 4 and 8.


Bidding 10 wins top slot, pay 8: profit 200 • 2 = 400.

Bidding 5 wins next slot, pay 4: profit 100 • 6 = 600.
If competing bids are 6 and 8, better to bid 10…
Example: GSP auction


Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Generalized Second Price Auction
 In this case, it is an equilibrium to bid truthfully
 Bidder 2 gets slot 2 and pays $2 per click (or $200)
 Bidder 1 gets slot 1 and pays $4 per click (or $800)

Efficient allocation, revenue is $1000 (> Vickrey!)

Why an equilibrium?
 Bidder 3 would have to bid/pay $2 to get slot 2 – not worth it.
 Bidder 2 would have to bid/pay $10 to get slot 1 – not worth it.
 Bidder 1 could bid/pay $2 and get slot 2, but also not worth it.
GSP equilibrium prices
p1
GSP prices are
also competitive
equilibrium prices!
8
6
GSP eqm
4
2
Not the only GSP
equilibrium, however
Vickrey prices
2
4
p2
Example: GSP auction

Two positions: receive 200 and 100 clicks per day
Advertisers 1,2,3 have per-click values $10, $4, $2.

Another GSP equilibrium (with Vickrey prices!)





Bidder 3 bids $2 and gets nothing
Bidder 2 bids $3 and pays $2 per click for slot 2
Bidder 1 bids $10 and pays $3 per click for slot 1
And another GSP equilibrium (w/ higher prices)



Bidder 3 bids $3 and gets nothing
Bidder 2 bids $5 and pays $3 per click for slot 2
Bidder 1 bids $6 and pays $5 per click for slot 1
GSP equilibrium prices
p1
8
6
GSP eqm
GSP equilibrium
prices are also
competitive
equilibrium prices!
GSP eqm
4
2
Vickrey prices
2
4
p2
General Model

K positions k=1,..,K
N bidders i = 1,…,N

Bidder i values position k at uik = vn • xk




xk is quantity of clicks, x1>x2>…>xK
vn is value of a click, v1>v2>…>vK
Efficient allocation is assortative

Bidder k should get slot k to max total value.
GSP Auction Rules

Each agent i submits bid bi

Interpret as “maximum amount i offers to pay per click”

Positions assigned in order of bids

Agent i’s price per click is set equal to the bid of
agent in the next slot down.

Let bk denote kth highest value and vk value.

Payoff of kth highest bidder:
vk • xk – bk+1 • xk = (vk - bk+1) • xk
GSP equilibrium analysis

Full information Nash equilibrium


NE means no bidder wants to change positions
Nash eqm is a bid profile b1,…, bK such that
(vk - bk+1) • xk  (vk - bm+1) • xm for m>k
(vk - bk+1) • xk  (vk - bm) • xm

for m<k
Lots of Nash equilibria, including some that
are inefficient…
Locally Envy-Free

Definition: An equilibrium is locally envy-free
if no player can improve his payoff by
exchanging bids with the player ranked one
position above him.

Motivation: “squeezing” – if an equilibrium is not
LEF, there might be an incentive to squeeze.

Add the constraint for all k
(vk - bk+1) • xk  (vk - bk) • xk-1
Stable Assignments

Close connection between


Think of bidders being “matched” to positions.
Matching postion i to bidder k with price pk gives


GSP equilibria / Competitive eqm / Stable assignments !
bidder payoff: (vi -pk)xk and position payoff pkxk
Stability: no bidder/position want to block.


All stable assignments are efficient (assortative).
Relevant blocks are bidders looking to move up or down
one position. (think about this).
Stable assignments


At a stable assignment, matching is efficient.
Each position k commands a price pk.

Prices that support a stable allocation satisfy:
(vk - pk) • xk  (vk – pk-1) • xk-1
(vk - pk) • xk  (vk – pk+1) • xk+1

These are the conditions for a competitive equilibria


Essentially they say that bidder k demands position k
So therefore at these prices, supply = demand!
Equivalence Result

Theorem:


Outcome of any locally envy-free equilibrium of
the GSP is a stable assignment (i.e. competitive
equilibrium allocation)
Provided that |N|>|K|, any stable assignment (i.e.
competitive equilibrium allocation) is an outcome
of a locally envy-free equilibrium.
GSP equilibrium prices
p1
8
6
GSP eqm
The set of competitive
equilibrium prices
corresponds to the set
of GSP equilibrium prices!
GSP eqm
4
2
Vickrey prices
2
4
p2
Revenue and Prices

Theorem



There exists a bidder-optimal competitive
equilibrium (equivalently, GSP equilibrium) and a
seller-optimal one.
The bidder optimal competitive equilibrium is
payoff-equivalent to the Vickrey outcome.
Corollary: any locally envy free GSP equilibrium
generates at least as much revenue as Vickrey.
Internet Advertising Markets
Today’s Lecture

Sponsored Search Market




Recap of last time
Examples of GSP/Comp Eqm/Vickrey
Auction design & platform competition
Display Advertising Market



Structure and organization of the market
Form of contracts, auctions vs prices
Heterogeneity, targeting and conflation
Sponsored Search Recap

Search engines sell positions on results page



Advertisers bid for keywords on per-click basis
“Generalized second price” auction for each query
Simple model of auction setting shows that:




Usually range of market clearing (per click) prices
Vickrey auction leads to lowest mkt-clear. prices.
Many possible GSP equilibrium outcomes, but the
payments coincide with mkt-clearing prices.
Equivalence btwn mkt-clear prices, GSP outcomes and
stable matchings of advertisers to positions.
Example

Two positions with 200, 100 clicks

Three bidders with values $2, $1, $1

Efficient assignment is assortative

Supporting (stable) prices

Bidder 2 pays $100 for slot 2, (or p2 = $1 /click).


High enough to deter bidder 3 => at least $100, but
not so much that bidder 2 wants to drop out.
Bidder 1 pays $200-300 for slot 1, p1 [1,3/2].

High enough to deter bidder 2 => at least $200, but
not so much that bidder 1 wants to drop down.
Example, continued

Two positions with 200, 100 clicks
Three bidders with values $2, $1, $1

GSP equilibrium





Efficient matching, payments of $1, and $p1.
Bids required to support these payments
 Bidder 3 bids $1 per click
 Bidder 2 bids $p1  [1,3/2].
 Bidder 1 bids at least p1
Easy to check that no bidder benefits from deviating…
For each set of competitive prices, there is a set of
GSP eqm bids, and vice-versa…. Revenue is $100
+ $(200-300), or $300 to $400 total.
Compare to Vickrey auction

Two positions with 200, 100 clicks

Three bidders with values $2, $1, $1

Vickrey auction


Efficient matching.

Bidder 2 pays $100 (displacing 3)

Bidder 1 pays $200 (displacing 2,3)

VCG payments are $100, $200. Total $300.
GSP prices are at least as high as Vickrey!
Another example

Three positions with 300, 200, 100 clicks

Four bidders with values $3, $2, $1, $1

Efficient assignment is assortative

Supporting (stable) prices


Bidder 3 pays $100 for slot 3, p3 = 1.

Bidder 2 pays $200-300 for slot 2, p2 [1,3/2].

Bidder 1 pays $400-600 for slot 3, p3 [4/3,2].
Relationship between VCG and GSP eqm

VCG payments are $100, $200, $400, revenue $700.

GSP payments range from $700 up to $1000.
Stable prices, generally

Stable prices satisfy
(vk - pk) • xk  (vk – pk-1) • xk-1
(vk - pk) • xk  (vk – pk+1) • xk+1

Re-arranging we get
pk-1xk-1  pk xk + vk(xk-1-xk)
pk-1xk-1  pk xk + vk-1(xk-1-xk)

This gives us a relationship between price of one
slot and the price of the slot above… can solve for
stable prices from the bottom up…
Features of Equilibrium

Allocation is efficient (assortative)

Increasing price of marginal clicks


Varian points out this is testable.

Implies bidders are click-constrained!

Pricing should be linear if bidders satiated…
Bids “reveal” bounds on bidder values.

Apparently not so easy to invert in practice.

Actual bidding is surprisingly unstable…
Ascending auction
What if there is incomplete information about
values – does this change things, or is there a
naural process through which market equilibrates?
Suppose price rises from zero, advertisers can
drop out at any time, fixing their bid.


Theorem (Edelman et al.).




There is a unique perfect equilibrium in which advertisers
drop out in order of their values.
The equilibrium is efficient and the prices are Vickrey.
The equilibrium is an ex post equilibrium – no one wants to
go back and change their bidding after the auction ends.
Choices in Auction Design



How many slots to sell?

Would search engine want to restrict slots available?

Could this ever increase revenue? Efficiency?
Setting a reserve price?

What is the optimal reserve price?

Is it better to use a reserve price, or restrict slots?
Clickability and “squashing”

What if ads have different “clickability”

Should you incorporate this in the auction? How?
Example: Slot Restriction

Two positions with 200, 100 clicks

Three bidders with values $2, $1, $1

Should the seller sell only one position?

Focus on lowest equilibrium (Vickrey) prices.

Selling two positions: revenue of $300.

Selling one position: revenue of $200.

Bidder 2 and 3 will pay up to $1 per click.

Market clearing price is $1 per click for Bidder 1.
Example: Slot Restriction

Two positions with 200, 100 clicks

Three bidders with values $3, $3, $1

Should the seller sell only one position?


Selling two positions: revenue of $500.

Bidder 2 will pay $1 /click for slot 2

Bidder 1 will pay $2 /click for slot 1
Selling one position: revenue of $600.

Bidder 2 will pay up to $3 per click for slot 1

Market clearing price is $3 per click for Bidder 1.
Example: Reserve Prices

Two positions with 200, 100 clicks

Three bidders with values $2, $1, $1

Can the seller benefit from a reserve price?

No reserve price: revenue of $300.

Reserve price of $2: revenue of $400

Sell only one position, but for $2 per click!
Example: Reserve Price

Two positions with 200, 100 clicks

Three bidders with values $3, $3, $1

Can the seller benefit from a reserve price?


No reserve price: revenue of $500 (or $600 if
sell just one position).

Reserve price of $3 (and sell both positions):
revenue of $900!
In general, is it better to use a reserve price,
or to adjust the number of slots for sale?
Bidder-Specific Click Rates

Some ads may be more relevant than others.




eg if query is “Pottery Barn,” what ad will get clicked?
Extended model where click rates differ.

Suppose Pr(click) = aixk

Values: uik= vi  (ai  xk)= (vi ai )xk

Bids are still made on a per-click basis
Value rank: rank bids by expected revenue, by bi  ai

Eqm allocation will maximize total value.

Bidder-optimal eqm will be payoff-equivalent to Vickrey
Bid rank: rank bids directly by bi .

May not be efficient, but may raise revenue.
“Squashing” Example

Two positions with 200, 100 “base” clicks

Three bidders


per-click values $2, $1, $1

click-thru rates: 2,1,1
Rank bids by bid*CTR

Bidder 2 pays $1 per-click for position 2

Bidder 1 pays $0.50 per-click for position 1 (why?)

Total revenue: $1*100 + $0.50*400 = $300.
“Squashing” Example

Two positions with 200, 100 “base” clicks

Three bidders



per-click values $2, $1, $1

click-thru rates: 2,1,1
Rank bids by bid (i.e. treat B1 “as if” CTR=1)

Bidder 2 pays $1 per-click for position 2

Bidder 1 pays $1 per-click for position 1. (why?)

Revenue: $1*100 + $1*400 = $500!
Can squashing would lead to inefficient matching of
bidders and positions? When?
Issues not modeled

Is each query a separate competition?




Model doesn’t allow for much uncertainty



Advertisers really have portfolio of bids & broad match…
They also have budget constraints, decreasing returns.
They also have a choice between competing platforms.
Click rates, effectiveness of advertising are known.
Seems to be a lot of experimentation in practice. Why?
Many aspects of search not captured



How do people decide whether/what to click?
Is there an interaction with “organic” search?
“Broad match” and the use of algorithms… very important.
Platform competition

Current search market (approx.)




Yahoo! and Msft strategic partnership



Google: 70% market share, RPS maybe $0.08
Yahoo!: 20% market share, RPS maye $0.05
Microsoft: 10% market share, RPS maybe $0.04
Agreement from summer 2009, just approved.
Searches on Y! will show Msft results and ads.
Questions:


What explains the difference between the platforms,
particularly the difference in monetization?
Are the platforms competing in a meaningful sense?
Platform competition, cont.

Scale economies: more searches means…





Cheaper for advertisers to bid on a per-search basis, if
there are fixed costs to campaigns.
Easier for advertisers to reach critical mass of consumers,
if they want to raise awareness.
More possibilities for platforms to experiment, estimate
click rates and improve broad match algorithms.
How could we try to distinguish these hypotheses?
Is scale a “barrier to entry” or “barrier to
competition”?
Platform competition

Competition for users/searchers



Competition for advertisers



Try to make algorithmic results “better”
Syndication deals to “buy” searches
Is advertising on platform A a substitute for advertising on
platform B? Not immediately clear.
If advertisers have diminishing returns or budget
constraints, yes – there is a “market for clicks”.
How should will changes in the design of platform A
to impact the competition on platform B?


Example: Yahoo! implements are reserve price system.
Competition between auction platforms is analagous to
price competition but not as well understood.
Internet display advertising


Real estate on non-search web pages

Wide variety of ads: text, graphical, video, etc.

Wide variety of advertisers: “brand”, “performance,” etc.
Differences with search advertising

Intent: search query makes it easier to discern users intent; Less
clear if user is reading their email or browsing the web.

Search traffic is controlled by small set of firms that get to impose
standards, specify form of contract and design markets.

Display advertising opportunities are controlled by many, many
publishers, so market is more fragmented. Result has been
intermediaries (such as Google, but others as well) trying to
create/design market for advertising to be traded.
Contract design



Different types of contracts
 Pay-per-impression (CPM): advertisers pay to have ad shown to
a fixed number of “eyeballs”.
 Pay-per-click (CPC): Advertisers pay for clicks.
 Pay-per-action (CPA): Advertisers pay for a “conversion” or sale,
or for some action (e.g. filling out a form post-click).
Contract design involves trade-offs in
 Incentives for publishers/advertisers
 Costs of certifying monitoring behavior and reporting
 Risk-sharing between advertisers and publishers.
Different advertisers/publishers use different contracts.
 Performance advertisers (CPC), brand advertisers (CPM)
 Large publishers (CPM), small blogs and publishers (CPC).
Market design

Different models for advertising markets/sales
 Large publishers typically have sales forces that negotiate sales
of guaranteed impressions --- contracting can take place well in
advance of delivery (e.g. buy now for August).
 “Remnant inventory” is often sold through ad networks and
exchanges. These are often spot auction sales. Advertisers
submit bids, and market makers use algorithms to predict clicks
and determine allocation.
 Example: Google uses a version of its search auction to place
search ads on non-query web pages (AdSense).

Market design questions:



When should the market clear: Advance or real time?
How should the market clear: posted prices or auctions?
Will there be a “dominant” platform for ad sales?
Targeting and Conflation

Targeted advertising




Traditional advertising involved limited targeting
(e.g. everyone watching Law & Order)
Internet advertising allows advertisers to target
users based on demographics, geography, time,
recent search behavior, etc.
Many believe that internet advertising will
become progressively more targeted, with
better and better “matching” of ads to users.
But, there are also costs to targeting!
Conflation


Evolution toward targeting means that each ad
becomes its own “product”. This is not, historically,
how markets have evolved.
Example: market for wheat (Debreu)



Initially sold “by sample” each transaction different
Standard contracts and grades of wheat allowed for thick
markets, futures contracts, lower transaction costs.
Conflation: treat things that aren’t exactly the same
as the same in order to create better functioning
markets.
Market Thickness


Targeting can create thin markets
Facebook example, prices in spring 2009



Show ad to 1,000 Harvard undergraduates: $0.50
Show ad to 1,000 Havard econ majors: $0.05
What’s going on?



This is a very thin market for ads shown to Harvard econ
majors, resulting in a low price!
Targeting creates a large number of markets, some are
likely to be thin markets and this makes it hard to get all the
prices right.
Conflation: treat econ majors as undergraduates, or at
least apply undergraduate bids to econ major impressions.
Cherry-picking and “safety”


Targeting can also make markets “unsafe” by
allowing savvy advertisers to cherry-pick
Example: Yahoo! Happy Contract



Similar problems can arise in search auctions


McDonalds asks Yahoo! to show its ads only on sunny
days and when the stock market is up.
This leaves Burger King and Wendy’s with the inventory on
rainy days when the market is down.
Bid for auto insurance, but only in Palo Alto.
Solutions: eliminate targeting? ensure advertisers a
“representative” set of impressions?
Conflation: search advertising


Search advertising appears very targeted: can bid
for any of millions of keywords.
Yet there is also a lot of conflation




One bid applies to all position on the page
One bid (eg auto insurance) can be applied to many other
keywords (eg auto insurance companies).
One bid applies to many users, and maybe to AdSense.
What is being “conflated” are clicks in different
contexts. Conflation can mean same bid applies or
is scaled mechanically according to an exchange
ratge (e.g. a discount for AdSense clicks).
Targeting and Conflation


When many heterogeneous goods are being
sold, there are important set of trade-offs in
defining the products for sale…
Targeting, or finer product definition, means




Improved (more efficient) matching, but
Potential for thin markets
Potential for cherry-picking
Conflation is a key element of market design
Simultaneous Ascending Auctions
(and Radio Spectrum Licenses)
Today’s Lecture


Background on radio spectrum auctions
The simultaneous ascending auction





Example, and “magic of the market”
Theory of these auctions
Evidence on how they work in practice
What can go wrong, and why.
Next time: complex auctions, bidder strategy
and innovative spectrum auction designs.
FCC Spectrum Auctions

Auctions to allocate radio spectrum



Suggested by Coase (1960), and adopted by the FCC in
1994, followed by UK, Germany, Netherlands, Belgium,
Mexico, India, etc.
Large auctions in which telecommunications companies
may pay billions of dollars for spectrum licences.
Structure of typical auction


Government (e.g. FCC) specifies a set of licenses to be
sold. Each license conveys the right to use a portion of the
spectrum in a certain geographic area.
Licenses are sold in an auction, often using a simultaneous
ascending auction designed by Milgrom-Wilson-McAfee.
Structure of the Problem

Potentially many different goods for sale


Potentially bidders with different objectives


E.g. license for San Francisco very different than
license for Death Valley.
E.g. Verizon may want spectrum to add 3G; rural
telco might want spectrum for another purpose.
Substantial uncertainty about price/value

Illiquid secondary market, not many licenses or
auctions, uncertainty about technology evolution.
Example




Two licenses: New York and San Francisco
Three bidders: A, B, C.
NY
SF
A
40
35
B
60
50
C
80
60
Let’s run an SMR auction…
Efficient allocation: C wins #1, B wins #2.
SMR/SAA Rules

Simple case: each bidder allowed at most one license.

Auction Rules

Seller sets initial price, and is the “high bidder” on each license.

First round: each bidder can submit a bid for one of the licenses.
If a license gets multiple bids, a coin flip determines the new high
bidder. If it gets no bids, the previous high bidder remains.

In subsequent rounds, bidders who are standing high bidders
don’t bid, other bidders must submit a bit or else exit the auction.

Each new bid must be some increment (say 5% or 10%) above
the standing high bid on a license.

The auction ends when there are no new bids. Each standing
high bidder pays its current bid and receives its license.
SMR/SAA Rules

General case: bidders can bid for many licenses.

Auction Rules





Seller sets initial price, and is the “high bidder” on each license.
Bidders can submit bids on any set of licenses including raising
their own high bids, but must respect an “activity rule”. Bids must
be at least an increment above the current price.
If a license gets one new bid, that bidder becomes the new high
bidder. If no new bids, standing high bidder remains. If multiple
new bids, a coin flip determines the new high bidder.
After each round of bidding, information about bids revealed.
Auction ends when there are no new bids. Each standing high
bidder receives its license at its last bid price.
Why a simultaneous auction?

Relative to sequential auctions

Bidders have more chance to coordinate their
bids on different licenses – may be important if
bidders want to assemble packages.

Information is aggregated all at once – may be
important if lots of uncertainty about value.

Should leads to greater degree of arbitrage and
similar prices for similar items.
Why a multiple round auction?



Relative to sealed bidding, information revelation…

Allows bidders to identify target licenses “on the fly”

Mitigates inefficiency due to the winner’s curse

Helps bidders to assess “roaming” opportunities.
The SAA design has some other virtues..

It’s transparent, and easy to check up on the gov’t.

Activity rule prevents super-slow bidding.
Skeptics might argue…

Design is vulnerable to demand reduction/collusion.

Design does not facilitate new entry or “package” bidders.
Why an SMR auction?


Suppose we make the following assumptions:

Each bidder places a dollar value on each of the licenses,
and wants to maximize its “profit” – equal to difference
between value and price paid.

The government’s objective is efficiency, defined as
allocating the licenses to maximize the sum of the dollar
values of the winners.

The government does not know the license values of the
bidders. (What if they did?).
We’re ignoring many potential complications, to
which we’ll return.
The magic of the market

Bidding is “straightforward” if in each round bidders submit bids
on the license that offers them the most profit at current prices.

Theorem. If bidders want one license (or multiple licenses but
“substitutes” values”) and bid straightforwardly, the SMR auction
leads to an efficient allocation and competitive eqm prices.

“Magic of the market”


Auction outcome is “as if” the seller knew all the values and used
a computer to find the efficient allocation & mkt-clearing prices.

Biders have “no regret” – at the final prices, each winner gets
exactly the license that gives it the most profit, and no loser
would like to be a winner.
Put another way, the auction is a price discovery mechanism to
find the efficient (market clearing) prices!
Example




Three bidders A, B, C.
Two licenses: NY and SF.
NY
SF
A
40
35
B
60
50
C
80
60
Efficient allocation: C wins NY, B wins SF.
Let’s see how the SMR auction works (in theory)…
SMR Auction

License Values
A
B
C
NY
40
60
80
SF
35
50
60




Suppose prices start at $0, and the
minimum raise is $1.
Initially everyone bids on NY.
When prices reach $5, $0: A bids
on SF; B and C continue on NY.
When prices reach $11,$1, bidder
B starts switching back and forth
between NY and SF. (“arbitrage”)
When prices reach $45, $35, bidder
A drops out, so B wins SF at $35
and C wins NY at $45.
Alternative: “Vickrey” auction



Sellers asks bidders to write down their
values and mail them in.
Seller allocates licenses to maximize total
value.
Prices are set so that (two equivalent ways)


Each bidder makes a profit equal to total value
with them in the auction minus the total value
without them in the auction.
Each bidder pays the value he “displaces” by
entering the auction and receiving his allocation.
Vickrey auction

License Values
#1
A 40
B 60
C 80
#2
35
50
60



Efficient allocation gives $130

C gets 1, worth $80

B gets 2: worth $50
Without C, value would be $95

B would get 1, worth $60

A would get 2, worth $35

So C adds $130 - $90 = $35 in value.

C values 1 at $80, so must pay $45.
Without B, value would be $115

C gets 1, A gets 2, so $80 + $35.

So B adds $130 - $115 = $15.

B values 2 at $50, so must pay $35.
Vickrey prices: $45, $35 - as in SMR!
Vickrey auction


Under the Vickrey rules, bidders do best to
reveal their true values: it’s strategy-proof!

If you bid more or less than your value, you don’t
change what you pay unless you change what
you win.

If you bid more or less and it changes what you
win, you make less profit.
Logic is similar to the other Vickrey cases we
considered (T-bills & sponsored search).
Vickrey strategy

Suppose B bids $60 and $0.

License Values
#1
A 40
B 60
C 80
#2
35
50
60



Then B wins 1, C wins 2.
Total value is $120

B’s value for 1 is $60

C’s value for 2 is $60
Without B, value would be $115

C gets 1, A gets 2, so $80 + $35.

So B adds $130 - $115 = $15.
So B must pay $55 for #1 and makes a
profit of only $5, less than the $15 it gets
from truthful reporting.
SMR auctions in practice

SMR auctions (and “clock auction” variants) are very
common for spectrum and other goods.

What is the evidence on their performance?

We’ll start with a very successful UK auction,

Then we’ll look at some European auctions that weren’t so
successful and think about what can go wrong.

Next time, we ‘ll discuss some of the US auctions, which
are often more complicated, and sometimes surprising.

Finally, we’ll talk about some innovative new auction
designs that are starting to be used, and their properties.
UK Auction of 3G Spectrum

In 1998, British Radiocommunications Authority (RA)
was designated to sell spectrum licenses for third
generation (3G) wireless services.

The RA decided to follow the FCC and use an auction:
banks estimated an auction might raise £500 million
($750 million).

The RA employed a small group of auction designers.

Paul Klemperer (former Stanford GSB student, and
my advisor at Oxford) was principal auction theorist.

Economists Jeremy Bulow of Stanford GSB and Ken
Binmore of LSE were also involved.
Deciding what to Sell


An important question is what to sell

How many licenses? Exactly what spectrum?

In this case, european countries had agreed to all
designate certain spectrum for 3G use; the main question
was whether to sell 4 or 5 licenses.
Licenses grant owner the right to use the spectrum,
and an obligation to build out a network of cell
towers to provide coverage.

More service is desirable, but concern as to whether it was
reasonable to expect buildout of five separate networks.
Identifying Bidders

There were four incumbent phone companies in the
UK, operating “2G” services.



Vodafone, Orange, British Telecom, One-to-One.
Complication: Vodafone was trying to take over
Mannesman, which owned Orange.

Vodafone had agreed to divest Orange, but post-auction.

Gov’t decided to let both Vodafone and Orange bid.
Now, the big problem: what auction design to use,
and would the design attract several other bidders?
Ensuring competition


SMR auction is efficient and prices are competitive

given the set of bidders who participate

assuming bidding is straightforward
In the UK, entry was the serious concern

With four licenses, and four strong incumbents, new
bidders might not bother to show up.

Without a new bidder, prices might be very low.

The bidding team considered an “Anglo-Dutch” design –
SMR until five remaining bidders, then a final sealed bid
round … could this have helped?
What happened


The RA finally decided to sell five licenses.

License A reserved for a new entrant

Licenses A, B a bit bigger than C,D,E.

Bidders can only bid on one license at a time

In first round, everyone bids whatever they want on one of
the five licenses – then single increment bids.
There were thirteen entrants in total

Incumbents plus nine entrants including major players like
Telefonica, Hutchinson Whampoa, etc.
What happened
What happened

Auction ends after 150 rounds….
Success in the UK

Four incumbents and one entrant won
licenses.

Auction netted £22 billion, about $39 billion
dollars – “the biggest auction ever”.

Some other auctions held around the same
time, however, were not as successful….
Entry problems

Netherlands auction of 3G spectrum in 1999.

Following UK lead, decided to sell five licenses
using an SMR auction.

However, there were some differences


The Netherlands has five incumbent (2G) operators.

There was no prohibition on bidding partnerships.
What happened?
Netherlands auction

Prior to auction, major outside telecom firms (Deutsche
Telekom, DoKoMo, Hutchinson Whampoa) all reach
partnership agreements with an incumbent.

This left just one entrant, a startup called Versatel.

What happened in the auction

On day 1, Telfort (owned by BT) sends Versatel a letter saying
that it “can’t win” and should drop out immediately!

Versatel shortly drops out: total revenue of 3bn euros – at UK
prices, auction would have raised 10bn euros.
Bidding problems


German GSM auction in 2000.

Ten nationwide licenses.

Bidders allowed to win multiple licenses.

First bid at 10m DM, then 10% price increments.

Bidders: two very strong bidders, Mannesman
and T-Mobile, and some small guys.
What might you be worried about here?
German GSM Auction

What happened in the auction

Round 1: Mannesman bids 36.6m for each of 5
bands, and reduces eligibility.

Round 2: T-Mobile (Deutsche Telekom) bids 40m
for the other five bands, reduces eligibilty.

No bids in round 3!
Complications

Even in the UK setting where things look nice, one
might be worried about a number of issues

Consumers might care who are the winners: maybe “total
value” isn’t the right objective.

Firms may care who are the other winners, or may be just
learning their values in the auction.

The SMR is designed for efficiency, but maybe isn’t the
design that maximizes revenue.

Maybe competitive prices are too high – because they
leave firms without enough money to build their networks.

The SMR design didn’t deal with the problem of how many
licenses – the Germans tried to, but with limited success!
Conclusion

Magic of markets: auctions can be a powerful tool
for price discovery and efficient allocation.

But plenty of things can go wrong. Successful
auctions need to:


Induce bidders to participate

Induce bidders to bid competitively
Considerations around the auction must be
accounted for (what to sell, what are the objectives,
etc.): “The auction is always bigger than you think!”
Spectrum Auctions:
Strategy and Design
Today’s Lecture

Simultaneous Ascending Auctions

Bidder strategy in “complex” auctions




The exposure problem
Budget constraints and forecasting
The AWS auction
New auction designs
Refresher: SAA Rules


Auction consists of multiple rounds.

Round begins with standing high bid on each license
(initially the seller), and a minimum bid increment.

Each bidder can submit bids on any number of items,
subject to an eligibility and activity rule.

If no bids on a license, standing high bidder remains. If
multiple bids, one bid selected at random to be high bid.

Information about bids is revealed to bidders.
Auction ends when no new bids are submitted.
Theory of SMR auctions


Suppose bidders view licenses as substitutes and bid
straightforwardly, i.e. each round bid for most desired
licenses at current prices. Then,

Arbitrage: the final prices for identical items will differ by at
most one bid increment.

Competitive equilibrium: the final prices will approximately
competitive equilibrium prices.

Efficiency: If the bid increments are small, the final license
allocation will be efficient.
Results due to Gul and Stacchetti (2000), Milgrom (2000).
SAA auctions in practice

SAA auctions (and “clock auction” variants) are very
common for spectrum and other goods.

What is the evidence on their performance?


In the UK spectrum auction, the SMR auction appeared to
work very well – simple setting, not much strategy.

In some other European auctions, we observed problems.

Now we’ll look at some evidence from US auctions, where
things are often more complicated, and surprising.
Focus on elements that create role for strategy

Exposure problem, activity rules, budgets & complexity.
The exposure problem


New entry may require a package of licenses

Danger for entrant: might end up with very expensive spectrum
but not enough for viable entry.

Why not re-sell? Problems include opponent budgets, other
package bidders, bargaining and agency problems.

Fear of being “exposed” to losses can lead to conservatism, and
auction outcome may not be efficient.
Exposure problem is caused by uncertainty.

Bidders may have to make committing bids early in the auction,
when they are uncertain about how much it will cost to complete
their target package.
Exposure problem

Two licenses A and B




Entrant has value 100 for the pair, else zero.
Individual bidders for A, B with values U[0,125].
Suppose A is sold first, followed by B.
Solve for entrant’s optimal strategy



In entrant loses A, won’t bid for B and gets profit = 0.
If entrant wins A, will bid to 100 for B (why?) and expects a
profit (4/5)* [100 – 50] – pA = 40 - pA
Therefore in the first auction, entrant will bid up to 40.
Efficiency vs equilibrium
pB
125
Entrant wins
A only
100
LOSS
Entrant
wins both
licenses
Efficient for Entrant to
win if individual bidder
values are inside pink line,
ie if 100 > pA + pB.
PROFIT
0
0
100
125
pA
Strategy for exposure problem

Strategy can potentially resolve uncertainty.

Example


Entrant has value 100 for the pair A & B.

Individual bidder for A has value U[0,100]

Individual bidder for B has value U[0,60]
Possible ways the auction could go

A sells first, B sells first, or prices rise together?

Entrant may be able to have some control over this.
Auction timing


Suppose license B sells first

If entrant wins B, expects to pay 50 for A.

Value of winning B: 100 – 50 – price of B

Optimal strategy: stop bidding at pB=50.

Expected profit 20.8.
Suppose license A sells first

If entrant wins A expects to pay 30 for B.

Value of winning A: 100 – 30 – price of A.

Optimal to stop on A at pA=70, exp. profit 24.5.
Controlling the Pace


Best case for the entrant

Prices rise on both licenses, but faster on A.

If and when prices reach pA=60, pB=20, entrant exits.

At pA=60, pB=20, value of winning A (or B) is zero.

Expected profit is 25.3.
Idea: entrant should raise prices in a way that
provides the most information before
becoming committed to a purchase.
Strategic individual bidder

Entrant with value 20 for licenses A & B together.

Two individual bidders


License A bidder with value 10

License B bidder with value cB ~ U[0,30]
License B bidder prefers to see price on A rise first.


Entrant will exit when pA= 5, so possible to buy B for zero.
License B bidder can use strategy to exacerbate the
exposure problem for the entrant!
Activity Rules

Activity rules necessary to keep auction moving




Each license assigned some number of points
Bidder start with eligibility points, must use them each
round or else have their eligibility reduced.
Problem for a “package bidder” – creates exposure risk.
Activity rules also complicate arbitrage



Suppose NY worth 200, LA worth 100, SF worth 100.
If you’re high on NY and get bid off, can switch to SF/LA,
but what if you’re high on SF/LA and are bid off LA only?
Serious issue if some licenses much bigger than others!
Activity rules and timing


Strategy to deal with activity rules

Bidders can “park” points to save them for later.

Bid on large licenses to maintain “flexibility”
Auction timing: this suggests that…



Bidding will tend to start on large licenses;
Large license licenses will tend to “clear” first.
Similar licenses may not sell for the same price.
Bidding activity (FCC auction 35)
Fraction of bids on
large licenses
Fraction of bids on
small licenses
Timing of final bids (auction 35)
Variation in
clearing round
Round of final bid plotted against
license size in bid units
Large licenses
clear first
The AWS Auction

Use FCC’s auction of Advanced wireless
service in 2006 to illustrate bid strategy and
features of large auctions.

This was a large, complicated auction with a
very surprising outcome that has
subsequently influenced auction design.
US sale of AWS spectrum (2006)


Background for the auction

90 MHz of nationwide spectrum, 1122 licenses

Regional licenses (10,10,20 MHz), 6 to cover US

Smaller licenses (10,20,20 MHz), 176 to cover

Total of 168 bidders, including major incumbents, smaller firms.

Two potential national entrants: SpectrumCo and WirelessDBS.
Entrants face a difficult problem

Theory doesn’t provide much guidance on how to bid in a way
that avoids the exposure problem…
SpectrumCo problem

Goals for the auction



Acquire 20 MHz of spectrum covering 85% of US
population, without spending more than budget.
If this is impossible, don’t buy anything?
Strategic thinking


Beware the exposure problem!
Try to figure out how much it will cost to buy target
amount of spectrum… but how?
Hard to forecast prices!
Role of bidder budgets


Many bidders appear to be limited by budgets, rather
than values --- a neglected but important pattern.

With many substitutable items for sale, a straightforward bidder
will eventually bid its budget and continue doing so each round.

Even if some bidders don’t behave this way, aggregate demand
elasticity will be anchored around -1 as prices rise.
Empirical proposition:

Auction “exposure,” defined as sum of all bids in a round should
rise faster than auction revenue and level off at final revenue.
“Exposure” forecasts prices
Sum of all bids (exposure)
FCC Auction 35
Sum of high bids (revenue)
Forecasting in the AWS auction
Sum of all bids (exposure)
FCC AWS Auction
Sum of high bids (revenue)
Not everyone is a budget bidder
Exposure of individual bidders in the AWS auction
T-Mobile
Spectrumco
Dolans
Applying the budget hypothesis


Why is an accurate forecast of final prices useful?

Avoid exposure problem: allows an entrant to identify if a desired
aggregation is achievable at reasonable price.

Acquire licenses cheaply: allows a bidder to anticipate price
anomalies if licenses clear in sequence.
Would budget forecasting have worked in past auctions?

Requires exposure to peak sufficiently early.

Requires exposure not to overshoot final revenue.
Exposure peaks early in auction
Peak/final exposure FCC sales
Overshooting in
“small” auctions
No overshoot in large auctions
How the AWS auction worked


Recall basic structure of licenses:

“Large” regional licenses (three bands, 40 MHz)

“Small” EA/CMA licenses (three bands, 50 MHz)
Competitive landscape: 168 bidders, major
incumbents, and two potential national entrants

SpectrumCo: cable TV consortium

Wireless DBS: satellite TV consortium
Controlling the pace

Bidding started on large regional licenses.


But due to uniform starting point, prices rose uniformly on
coasts/interior, creating serious exposure problem…
In round 9, Spectrumco makes maximum possible jump
bid on all Northeast and West regional licenses, doubling
their prices from $750m to $1.5 billion.

Shake-out: Wireless DBS takes waivers, then exits.

FCC eliminates jump bidding in subsequent auctions.

But prices continue to rise on the REAG licenses…
Rising prices in AWS auction
Budget forecasting

At round 13, the situation is



High bids on REAGs (40 MHz): $5.0 bn
High bids on EA/CMAs (50 MHz): $0.7 bn
Auction exposure had peaked at $14.2 bn.

Cable consortium gives up REAG licenses and switches
to smaller licenses, other major bidders do not.

Why did no other large bidder switch?


Large licenses “easiest” way to buy large quantity and no
reason a priori to expect they’d be much more expensive.
Incumbents did not face exposure problem: less need to
forecast prices and “discover” budget theory.
Rising prices in AWS auction
Spectrumco
band switch
Rising prices in AWS auction
Timing of final bids in AWS
Large REAG
licenses
Another exposure problem


At round 19, the situation is

High bids on REAG licenses: $7.6 bn

High bids on EA/CMA licenses: $2.3 bn
Budget algebra

Implied maximum budget for small licenses: $6.6 bn.

Estimated price of 20 MHz national: $2.6 bn.
Rising prices in AWS auction
Rising prices in AWS auction
Price per MHz-pop of REAG licenses
Price per MHz-pop
of EA/CMA licenses
Rising prices in AWS auction
Price per MHz-pop of REAG licenses
Price per MHz-pop
of EA/CMA licenses
Rising prices in AWS auction
Price per MHz-pop of REAG licenses
Price per MHz-pop
of EA/CMA licenses
Similar spectrum, different prices

US auction of AWS spectrum (2006)
Band
MHz
License
type
Price
($/MHz-pop)
Price (US
10Mhz)
A
20
CMA
$0.40
$1.1 bn
B
20
EA
$0.43
$1.2 bn
C
10
EA
$0.52
$1.5 bn
D
10
REAG
$0.62
$1.8 bn
E
10
REAG
$0.61
$1.7 bn
F
20
REAG
$0.73
$2.1 bn
SpectrumCo’s Licenses (20 MHz)
Failure of price arbitrage
Table 1: Prices Paid by the Five Largest Buyers in the AWS Auction
Bidder
Total Amount Paid
MHz-Pops
Price per MHz-Pop
$ 2,377,609,000
5,267,189,470
$ 0.45
Cingular
1,334,610,000
2,436,458,880
0.55
T-Mobile
4,182,312,000
6,638,718,070
0.63
Verizon
2,808,599,000
3,840,952,220
0.73
MetroPCS
1,391,410,000
3,840,952,220
0.96
$ 9,716,931,000
14,361,573,190
$ 0.68
SpectrumCo
Four incumbents
New auction designs

AWS auction spurred new designs




700 MHz auction in United States
British WiMax auction
These auctions allow for “package bidding”
British design also involves new twists


“Vickrey” or “near-Vickrey” payment rule
Conflation in license definitions
US 700 MHz Auction

Three main bands



Bidders allowed to submit a “package bid” for the
entire national C band.



A and B bands were 10MHz
C band was 20 MHz, with special rules
Lobbied for by Google, although they didn’t buy.
Package bidding favors entrants – can create a “threshold
problem” for smaller bidders!
In the auction, C band went at a huge discount.
British wimax auction

Very innovative new design to sell spectrum in the UK.

120 MHz of spectrum divided into 42 5-MHz blocks.

First stage clock auction

Seller calls out prices

Bidders call out number of blocks they want

Stage ends when Demand <= Supply.

Sealed bids: if overshoot in first stage, bidders can “add” bids.

Then, seller takes all bids and computes efficient allocation.

Bidders pay Vickrey prices, … but if Vickrey prices are too low
(outside the core), prices are increased until the resulting
allocation is in the core!
British WiMax, cont.


Auction also includes third stage

Winners are guaranteed set amount of spectrum

Third stage determines who gets which blocks.

Sealed bid package Vickrey auction (with prices adjusted
up if result is outside the core).
Conflation in auction design

Different items are treated as identical (conflated)

Additional round used to “de-conflate” the items.

Conflation is very common in setting up markets.
Conclusions

Simultaneous multi-round auctions are commonly used
for selling radio spectrum, and many other goods.

Design has many advantages in terms of revealing
information, giving bidders flexibility, but auctions can
also be complex creating a role for strategy.

Innovative new designs such a British auction are trying
to simplify the bidder problem…

Experience of Spectrumco shows how economic theory
can be practical as well as fun!
Package Auctions
London Bus Routes

City of London auctions off service contracts for bus
routes in Greater London.

How it works

City decides bus network, frequency of buses, types of
buses, exact routing etc. – provision is outsourced.

City uses sealed bid auctions allowing bids for packages of
routes as well as individual routes.

Bids state a fee to be paid to the city to operate the buses
for five years – collected fares go to the city.
Why package bids?


Operator cost structure

To operate routes, must have storage and maintenance
facility – fixed cost, can accommodate several buses.

So costs per route may decline if operator has several
routes – at least until capacity is reached.
Bidding advantage in the auction

Consider operator with zero cost on routes A,B.

Suppose opponent bids on A,B are U[0,10]

Bidding 5 for A, 5 for B => expected profit 2*5*(1/2)=5

Bidding 7.5 for A/B package => exp. profit 7.5*(3/4)=5.6!
London auction data
Source: Cantillon and Pesendorfer, “Auctioning London Bus Routes,” 2006
Package bid “discounts”

Note: marginal discount assigns full discount to
smallest route in the package.
Interpreting the data

Cantillon-Pesendorfer try to infer if package bidding
reflects cost synergies or is used strategically.

They estimate operator costs in the London market
using the bid data…

Findings:

Mininmal or no cost synergies!

Strategically use of package bids: bids are marked up over
cost by 16.4% for individual routes, but only 11.4% for
packages!
Package auctions


Package bidding allows bidders to express
complex demands in multi-good settings.

A set of bus routes in London

A set of spectrum licenses.

A set of airport take-off and landing slots

A complete search/display advertising campaign
Today: discuss underlying theory, different
auction methods, and some evidence.
The problems begin…




Two items for sale, A and B
Two bidders with values:
A
B
AB
1
0
0
12
2
10
10
10
No item prices clear the market.
Such prices must result in bidder 1 efficiently
buying both: pA10, pB10, and pA+pB12!
SMR auction
1
2



A
0
10
B
0
10
AB
12
10
Suppose everyone know values (complete information)

Bidder 1 should not bid!

Bidder 2 will win one item at a very low price.
With uncertainty about opponent values

Bidder 1 may start bidding but will have to pay 10 for each to win both.

More likely outcome – bidders split items, perhaps at low prices.
The outcome is likely to be inefficient and maybe yield low revenue.

Although recall Spectrumco overcoming exposure problem.
Package bids in the SMR

Allow bidders to make “package bids” - e.g. $X for
the package of items A and B.

SMR with package bidding

Each round, there is a provisional winner for each license –
some provisional winning bids could be package bids.

Bidders can submit new bids, individual or package, for any
licenses they like.

Seller takes new bids and existing bids (maybe including
losing bids) and identifies highest revenue allocation.

Auction ends when no new bids are submitted.
Auction of 700 MHz spectrum


US auction of 700 MHz spectrum in 2008.

B block: 10 MHz divided into 176 EA licenses

C block: 20 MHz divided into 6 REAG licenses

Bidders can make a national bid on the C block.
What happened?

Little bidding on C block – Google bid the reserve price,
and Verizon bid a bit higher.

Lots of bidding on the B block – prices about 4x that of the
C block.

But many other special features of the auction make it hard
to identify the exact effect of package bidding.
Threshold problem
1
2
3

A
0
10
0
B
0
0
10
AB
12
10
10
SMR with package bidding




Suppose bidder 1 starts by bidding 10 for AB.
Bidders 2 and 3 have to make bids that sum to 11.
Incentive to wait for other bidder to increase its bid.
But then auction could end with inefficient package winner!
Who gets the advantage?


It is hard to balance the playing field when
some bidders have package preferences

Individual bidding: creates difficult exposure
problem for complements bidder.

Package bidding: creates difficult threshold
problem for individual bidders.
Why package bidding changes things

Effectively gives package bidder a chance to
move second – e.g. bid 10, and let the auctioneer
divide up the bid once the other bids are in!
Vickrey auction

Vickrey auction with package bidding



Bidders submit their values for all possible
packages (could be a lot -- 2N)
Seller finds highest value allocation, sets prices
so that each bidder makes as profit the difference
between value with them and without them.
Desirable properties


The outcome is efficient if bidders are truthful.
Truthtelling is a dominant strategy
Vickrey auction: problems

Consider our “threshold problem” example
1
2
3



A
0
10
0
B
0
0
10
AB
12
10
10
Bidders 2 and 3 win items.
Each pays a Vickrey price of 2 => revenue 4.
BUT, package bidder would pay 10!
Vickrey auction: problems

Now suppose bidder 2,3 have lower values
1
2
3



A
0
3
0
B
0
0
3
AB
12
3
3
With honest bidding, package bidder wins.
If bidders 2,3 report 10, each wins and pays 2
Small amount of collusion has a big effect!
Vickrey auction: problems

Bidders may also want to split their bidding
1
2



A
0
0
B
0
0
AB
12
11
Honest bidding means 1 wins and pays 11.
Bidder 2 can enter as 2A, 2B, each bidding 10
for a single item – wins both and pays 4!
If bidders 2A, 2B bid 11 each, they each win and
pay 2, so higher bids can mean lower revenue!
Budget constraints

We saw that budget constraints seem to be
important in spectrum auctions.

They pose a big problem for Vickrey auctions


Suppose items A, B are for sale.

Bidder values A at 200, B at 100, budget of 150.

Can’t bid true values and be sure to stay under budget.

“Straightforward” bid of 150 for A, 100 for B and 150 for the
pair implies zero value for B if awarded A.
Bidding with a budget & vickrey rules is complicated!
Core outcomes

In markets with complements, market
clearing prices may not exist.

“Core” allocations are a useful generalization.


An allocation is in the core if there is no set or
“coalition” of players that could make a deal on
their own from which all of them would benefit.

“All gains from trade are exploited”
The “bad” Vickrey examples are cases where
the Vickrey outcome is not in the core.
Vickrey and Core outcomes

Same example as before
1
2
3



A
0
10
0
B
0
0
10
AB
12
10
10
Bidders 2 and 3 win items, pay 2 each.
But package bidder would pay 10!
That is: the auction outcome is not in the core!
Core outcomes and auctions

There is always at least one core allocation



Example: assign items efficiently, have buyers
pay their full value to the seller.
Problem: unclear how to get bidders to reveal
values!
Day and Milgrom (2008) propose to use
“core-selecting” auctions in which:



Bidders are asked to submit bids
Bids are treated as values.
Seller finds core allocation that is bidder optimal.
Auctions vs exchanges

B
AB
1
0
0
12
2
10
10
10
With one seller, there is a core allocation.


A
Bidder 1 wins the object, pays between 10 and 12.
But if goods A and B belong to different sellers, the core
is empty, because…



Bidder 2 must get 0
Coalition of either seller and bidder 2 must get 10
So, each seller must get 10, but only 12 is available.
Pay as bid auctions

London bus routes are a “pay-as-bid” auction




Bidders submit bids
Bids treated as values to find efficient allocation
Bidders are asked to pay bids (seller optimal).
Bernheim and Whinston (1986, QJE): the full
information Nash equilibria of the pay-as-bid
package auction correspond to the set of
bidder-optimal core allocations.
Pay as Bid Auction



A
B
AB
1
0
0
12
2
10
10
10
Bidder 1 bids 10 for package A/B
Bidder 2 bids bA and bB that sum to 10.
Allocation is efficient, revenue is 10.
Pay as bid Auction
1
2
3


B
0
0
10
Bidder 1 bids 12 for the package
Bidders 2, 3 submit bids that


A
0
10
0
Are each less than 10, and sum to 12.
Allocation is efficient, revenue is 12.
AB
12
10
10
Experimental evidence


Does package bidding really help?

Published reports of experiment suggest remarkable
efficiency properties.

Porter et al (2003, PNAS): efficiency of 100% in 23 of 25
trials, 99% in the other two!
Hard to interpret these results

As number of items grows, the number of possible values
to use in the experiment grows as 2n - huge!!

Are these experiments focused on “easy” cases?
Easy and hard settings

Straightforward bidding in SMR auction: in
each round, bid for most desirable set of
items at current prices.

Kagel-Lien-Milgrom (2009)


Setting is “easy” if straightforward bidding in SMR
with package bidding leads to efficient outcomes,
otherwise is “hard”.
Use computer simulations to classify different
settings as easy or hard, then run human
experiments to see if difficulty of the setting
matters.
KLM experiment design
KLM Results
KLM results, cont.
Summary of theory


If the bidders view goods as substitutes, then

Competitive equilibrium (CE) prices exist.

There is a CE with “minimal prices”.

This CE coincides with Vickrey auction outcome.

This CE is a core allocation.
In the general package preference case

CE prices may not exist

Vickrey auction may not yield core allocation.

Non-Vickrey designs do not encourage truthful bidding

Auctions suffer from exposure/threshold problems.
British wimax auction

Very innovative new design to sell spectrum in the UK.

120 MHz of spectrum divided into 42 5-MHz blocks.

First stage clock auction

Seller calls out prices

Bidders call out number of blocks they want

Stage ends when Demand <= Supply.

Sealed bids: if overshoot in first stage, bidders can “add” bids.

Then, seller takes all bids and computes efficient allocation.

Bidders pay Vickrey prices, … but if Vickrey prices are too low (outside
the core), prices are increased until the resulting allocation is in the core!
British WiMax, cont.


Auction also includes third stage

Winners are guaranteed set amount of spectrum

Third stage determines who gets which blocks.

Sealed bid package Vickrey auction (with prices adjusted
up if result is outside the core).
Conflation in auction design

Different items are treated as identical (conflated)

Additional round used to “de-conflate” the items.

Conflation is very common in setting up markets.
Conclusion

Package auctions are finding increasing use for
hard resource allocation problems.

Vickrey auctions problematic because of low
revenue, non-core outcomes.

Alternative designs (pay-as-bid, Vickrey, SMR, etc.)
trade off incentive and distributional properties.

Package exchanges are fundamentally hard due to
empty cores, but some interesting new ideas are
being studied here as well.