Chapter 5: Auctions

Algorithmic Game Theory
Four basic formatsVickrey truthfulness proofBid shadingRevenue equivalenceReserve pricesWinner's curseDouble auctions
Sercan KÜLCÜ · Giresun University · Computer Engineering

Basic Concepts

Warm-up Game: The Vickrey Auction

The rules

Everyone writes a bid in a sealed envelope; the highest bidder gets the item but pays the second-highest bid, not their own.

Think about it
  • If the item is worth 70 TL to you, would you write less than 70, exactly 70, or more?
  • Would your bid change if you knew the rivals' bids?
  • If the winner paid their own bid (first price), would writing 70 still be a good idea?
70 TLYou60 TLRival 145 TLRival 2winssets the priceYou get the item and pay 60 TL.If your value is 70 TL, you gain 70 − 60 = 10.

Basic Concepts

Learning Outcomes

Running the formats

Find the winner and payment in English, Dutch, first-price and second-price auctions.

Proving truthfulness

Show by case analysis that truthful bidding is weakly dominant in the Vickrey auction.

Equilibrium bids

Derive the first-price shading b = (n−1)v/n; compute revenue equivalence (2 bidders → 1/3) and the effect of a reserve price (5/12).

Market phenomena

Explain the winner's curse; find the matching and price in a double auction.

Basic Concepts

Why Auctions Matter

  • Efficient mechanisms for allocating resources: the item tends to go to whoever values it most.
  • Price discovery and market transparency: the price is found without knowing the buyers' values.
  • The most widespread application of mechanism design: an allocation rule + a payment rule.
Art

Ascending-price English auctions at auction houses.

Online advertising

Ad auctions run on every search.

Telecom licenses

Spectrum auctions; multi-item designs worth billions.

Procurement, treasury

Sealed bids; government bond auctions.

Basic Concepts

Key Terms and the Model

Value vi

The bidder's true value for the item; private information.

Bid bi

The number reported to the mechanism; it may differ from the value.

Utility

ui = vi − pi if they win (pi: payment), 0 if they lose.

Independent private values (IPV)

Values are drawn independently and each bidder knows their own. The baseline model of auction theory.

Cem winspayment = 70uu = 85 − 70 = 15v = 85Cem losesno item, no paymentu = 0v = 85

Second-price example (Cem 85, second bid 70): quasi-linear utility.

Examples usually assume vi ~ U[0, 1] (uniform) and risk-neutral bidders.

Basic Concepts

Valuation Models

Private values

Each bidder values the item based on their own information (e.g. a painting for their own wall).

Common value

The item's true value is the same for everyone, but the bidders' estimates differ (e.g. an oil field).

Interdependent values

Both private information and the other bidders' information affect valuations.

private valueeach has their own: v₁ ≠ v₂ …common valueVone V, noisy estimatesinterdependentvᵢ = f(own + others' signals)
The common-value model leads to the winner's curse: the winner tends to be the most optimistic estimator, which produces expected losses.

Four Basic Formats

Four Basic Auction Formats

FormatHow it worksWinnerPays
EnglishThe price rises openly; bidders drop out one by one.Last one standingPrice at which the last rival drops out
DutchThe price starts high and falls; the first to say “mine” takes it.First to acceptThe current price
First-price sealed-bidBids are submitted secretly, once.Highest bidTheir own bid
Second-price (Vickrey)Bids are submitted secretly, once.Highest bidThe second-highest bid
Let's run all four rules with the same four people. The most each is willing to pay (their limit): Ayşe 70, Bora 55, Cem 85, Deniz 60. For now everyone bids their limit.

Four Basic Formats

Step by Step: The English Auction (Ascending Price)

0102030405060708090100Ayşe70Bora55Cem85Deniz60price = 50

The price starts at 40 and rises. At 50, everyone's limit is above 50: all four are still in.

0102030405060708090100Ayşe70Bora55outCem85Deniz60price passed 55

Once the price passes 55, Bora drops out (limit 55). Left: Ayşe, Cem, Deniz.

0102030405060708090100Ayşe70Bora55outCem85Deniz60outprice passed 60

Once the price passes 60, Deniz drops out (limit 60). Left: Ayşe, Cem.

0102030405060708090100Ayşe70outBora55outCem85winnerDeniz60outprice passed 70

Once the price passes 70, the last rival, Ayşe, drops out → Cem wins and pays ≈ 70 (exactly 70 as the increment shrinks): the second-highest value.

Four Basic Formats

Step by Step: The Dutch Auction (Descending Price)

0102030405060708090100Ayşe70Bora55Cem85Deniz60price = 100

At 100 the price is above everyone's limit; nobody accepts, and it drops by 5.

0102030405060708090100Ayşe70Bora55Cem85Deniz60price = 95

At 95: still nobody accepts (the highest limit is 85).

0102030405060708090100Ayşe70Bora55Cem85Deniz60price = 90

At 90: still nobody accepts (the highest limit is 85).

0102030405060708090100Ayşe70Bora55Cem85winnerDeniz60price = 85

At 85 the price reaches Cem's limit → Cem says “mine” and pays 85. The auction ends before anyone else can speak.

Four Basic Formats

Step by Step: First- and Second-Price Sealed Bids

0102030405060708090100Ayşe✉ sealed bidBora✉ sealed bidCem✉ sealed bidDeniz✉ sealed bid

Sealed bids: everyone bids once, in secret. Nobody sees the others' bids.

0102030405060708090100Ayşe70Bora55Cem85Deniz60

The envelopes are opened: 70, 55, 85, 60. Under both rules the highest bid wins: Cem.

0102030405060708090100Ayşe70Bora55Cem85winnerDeniz60payment = 85

First price: Cem pays his own bid → 85. If his value is 85, his utility is 0.

0102030405060708090100Ayşe70second highestBora55Cem85winnerDeniz60payment = 70

Second price (Vickrey): Cem pays the second-highest bid → 70. His utility is 85 − 70 = 15.

Four Basic Formats

Same Bids, Four Outcomes

FormatWinnerPaymentWhose bid sets it?
EnglishCem≈ 70The last rival to drop out (Ayşe)
Second priceCem70The second-highest bid (Ayşe)
DutchCem85The winner's own
First priceCem85The winner's own
English ≈ second price

With private values, staying in up to your value is dominant; both charge the second-highest value.

Dutch ≡ first price

Strategically equivalent: “at what price will I accept?” is the same as submitting a sealed bid.

Note: in this table everyone bid their limit. In a first-price auction a rational bidder shades the bid, so expected revenues actually come out equal (revenue equivalence).

Sealed-Bid Strategies

The Strategic Property of the Vickrey Auction

  • The highest bidder wins and pays the second-highest bid.
  • With independent private values and a single item, bidding your true value is a weakly dominant strategy (Vickrey, 1961).
  • The idea: your bid determines only whether you win, not what you pay.
020406080100120-200204060rivals' highest bid Butilityb = 90: losstruthful b = 70b = 50: missed gain

v = 70. Dashed red: the truthful bid. No bid beats it for any B.

The result may fail with common values, budget constraints or interdependent values. Vickrey is the single-item special case of VCG.

Sealed-Bid Strategies

Step by Step: Proving Truthfulness in Vickrey

020406080100120v = 70B = 60
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
 
 
 
 
 
 
 

Your value is v = 70. B = max b−i, the others' highest bid (unknown to you). Bidding truthfully with B = 60, you win and pay 60 → 10.

020406080100120v = 70B = 80
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
Truthful b = 7080loses0—
 
 
 
 
 
 

Bidding truthfully with B = 80, you lose → 0. The truthful bid always makes the right call: win when there is a profit, lose when there would be a loss.

020406080100120v = 70b = 90B = 60
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
Truthful b = 7080loses0—
Over b = 9060wins, pays 6010same
 
 
 
 
 

Overbid b = 90, B = 60 (< v): you still win and still pay 60 → 10. No difference.

020406080100120v = 70b = 90B = 80
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
Truthful b = 7080loses0—
Over b = 9060wins, pays 6010same
Over b = 9080wins, pays 80−10worse
 
 
 
 

b = 90, B = 80 (between v and b): you win but pay 80 → −10. Bidding truthfully you would get 0. A loss!

020406080100120v = 70b = 90B = 95
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
Truthful b = 7080loses0—
Over b = 9060wins, pays 6010same
Over b = 9080wins, pays 80−10worse
Over b = 9095loses0same
 
 
 

b = 90, B = 95 (> b): you lose → 0, same as truthful.

020406080100120v = 70b = 50B = 40
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
Truthful b = 7080loses0—
Over b = 9060wins, pays 6010same
Over b = 9080wins, pays 80−10worse
Over b = 9095loses0same
Under b = 5040wins, pays 4030same
 
 

Underbid b = 50, B = 40 (< b): you win and pay 40 → 30. Same as truthful (B, not b, sets your payment).

020406080100120v = 70b = 50B = 60
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
Truthful b = 7080loses0—
Over b = 9060wins, pays 6010same
Over b = 9080wins, pays 80−10worse
Over b = 9095loses0same
Under b = 5040wins, pays 4030same
Under b = 5060loses0worse
 

b = 50, B = 60 (between b and v): you lose → 0; bidding truthfully you would have gained 70 − 60 = 10. A missed chance!

020406080100120v = 70b = 50B = 80
StrategyB (rival)OutcomeUtilityvs. truthful
Truthful b = 7060wins, pays 6010—
Truthful b = 7080loses0—
Over b = 9060wins, pays 6010same
Over b = 9080wins, pays 80−10worse
Over b = 9095loses0same
Under b = 5040wins, pays 4030same
Under b = 5060loses0worse
Under b = 5080loses0same

b = 50, B = 80: you lose, same as truthful. Conclusion: deviating is never better and sometimes worse → b = v is weakly dominant.

Trying all pairs b, B ∈ {0, …, 120} for v = 70, no b ever beats the truthful bid.

Sealed-Bid Strategies

Back to the Game: The Vickrey Auction

Theory's answer: write your value

Opened envelopes: you 70, rivals 60 and 45. Bidding truthfully you win and pay 60: a gain of 10. Writing 65 or 100 you would pay the same 60; writing 55, the item goes for 60 and you miss the gain of 10.

Your bid only decides whether you win; the rivals' highest bid sets the price: truthful bidding is weakly dominant. Under first price, the same 70 leaves 70 − 70 = 0 even if you win; there you must bid below your value.

In practice

In lab experiments, overbidding in second-price sealed-bid auctions is common and does not easily disappear with repetition; the same subjects converge to equilibrium much faster in ascending English auctions (Kagel, Harstad and Levin, 1987).

In everyday life: in eBay's automatic (proxy) bidding, the winner pays one increment above the second-highest bid — very close to the second-price rule.

Moving the bid above or below the value never raises the gain for any B; it only lowers it for some B.

Sealed-Bid Strategies

Bid Shading in First-Price Auctions

  • In a first-price auction, a truthful bid is pointless: even if you win, utility is v − v = 0.
  • Lowering the bid raises the profit when you win but lowers the chance of winning:
E[u(b)] = (v − b) · P(win with b)
  • The equilibrium bid balances this trade-off and depends on how others bid → a Bayes–Nash equilibrium.
  • The Dutch auction is strategically the same: “at what price will I accept?” = a shaded bid.
00.20.40.60.800.51bid bbest b = 0.4 = v/2profit v − bP(win) = 2bE[u(b)]

Two bidders, v = 0.8; the rival bids half their value.

Sealed-Bid Strategies

Step by Step: The Equilibrium Bid b = (n − 1)v / n

1. Let the other n − 1 bidders bid bj = β·vj (vj ~ U[0, 1]).
2. P(win) = P(β vj < b for all j) = (b/β)n−1
3. E[u(b)] = (v − b) · (b/β)n−1
4. dE[u(b)]/db = [ −b + (n − 1)(v − b) ] · bn−2/βn−1 = 0
5. −b + (n − 1)(v − b) = 0  ⇒  b = (n − 1) v / n
6. Consistency: β = (n − 1)/n ✓   n = 2 ⇒ b = v/2

Assumption: rivals bid a fixed fraction of their values.

1. Let the other n − 1 bidders bid bj = β·vj (vj ~ U[0, 1]).
2. P(win) = P(β vj < b for all j) = (b/β)n−1
3. E[u(b)] = (v − b) · (b/β)n−1
4. dE[u(b)]/db = [ −b + (n − 1)(v − b) ] · bn−2/βn−1 = 0
5. −b + (n − 1)(v − b) = 0  ⇒  b = (n − 1) v / n
6. Consistency: β = (n − 1)/n ✓   n = 2 ⇒ b = v/2

For each rival P(vj < b/β) = b/β; by independence these multiply.

1. Let the other n − 1 bidders bid bj = β·vj (vj ~ U[0, 1]).
2. P(win) = P(β vj < b for all j) = (b/β)n−1
3. E[u(b)] = (v − b) · (b/β)n−1
4. dE[u(b)]/db = [ −b + (n − 1)(v − b) ] · bn−2/βn−1 = 0
5. −b + (n − 1)(v − b) = 0  ⇒  b = (n − 1) v / n
6. Consistency: β = (n − 1)/n ✓   n = 2 ⇒ b = v/2

Expected utility: profit when winning × probability of winning.

1. Let the other n − 1 bidders bid bj = β·vj (vj ~ U[0, 1]).
2. P(win) = P(β vj < b for all j) = (b/β)n−1
3. E[u(b)] = (v − b) · (b/β)n−1
4. dE[u(b)]/db = [ −b + (n − 1)(v − b) ] · bn−2/βn−1 = 0
5. −b + (n − 1)(v − b) = 0  ⇒  b = (n − 1) v / n
6. Consistency: β = (n − 1)/n ✓   n = 2 ⇒ b = v/2

The derivative of (v − b) is −1 and that of bn−1 is (n − 1)bn−2; the product rule adds them and factors out bn−2. At the maximum the derivative is zero.

1. Let the other n − 1 bidders bid bj = β·vj (vj ~ U[0, 1]).
2. P(win) = P(β vj < b for all j) = (b/β)n−1
3. E[u(b)] = (v − b) · (b/β)n−1
4. dE[u(b)]/db = [ −b + (n − 1)(v − b) ] · bn−2/βn−1 = 0
5. −b + (n − 1)(v − b) = 0  ⇒  b = (n − 1) v / n
6. Consistency: β = (n − 1)/n ✓   n = 2 ⇒ b = v/2

After simplifying, the solution does not depend on β.

1. Let the other n − 1 bidders bid bj = β·vj (vj ~ U[0, 1]).
2. P(win) = P(β vj < b for all j) = (b/β)n−1
3. E[u(b)] = (v − b) · (b/β)n−1
4. dE[u(b)]/db = [ −b + (n − 1)(v − b) ] · bn−2/βn−1 = 0
5. −b + (n − 1)(v − b) = 0  ⇒  b = (n − 1) v / n
6. Consistency: β = (n − 1)/n ✓   n = 2 ⇒ b = v/2

My best response has the same form → a symmetric Bayes–Nash equilibrium. With two bidders: half the value.

Maximizing E[u(b)] on a grid (10,001 points) for n = 2, 3, 5, 10 and v = 0.3, 0.6, 0.8, the best bid matches (n−1)v/n.

Sealed-Bid Strategies

Code Trace: best_bid(0.8, 2)

00.10.20.30.40.50.60.70.800.2bE[u]

Rivals bid b = v/2 (β = 0.5). Grid: b = 0, 0.1, …, 0.8.

00.10.20.30.40.50.60.70.800.2bE[u]0

b = 0: P(win) = 0, E[u] = (0.8 − 0)·0 = 0 → new best.

00.10.20.30.40.50.60.70.800.2bE[u]00.14

b = 0.1: P(win) = 0.2, E[u] = (0.8 − 0.1)·0.2 = 0.14 → new best.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.24

b = 0.2: P(win) = 0.4, E[u] = (0.8 − 0.2)·0.4 = 0.24 → new best.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.240.3

b = 0.3: P(win) = 0.6, E[u] = (0.8 − 0.3)·0.6 = 0.3 → new best.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.240.30.32

b = 0.4: P(win) = 0.8, E[u] = (0.8 − 0.4)·0.8 = 0.32 → new best.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.240.30.320.3

b = 0.5: P(win) = 1, E[u] = (0.8 − 0.5)·1 = 0.3 < 0.32: best unchanged.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.240.30.320.30.2

b = 0.6: P(win) = 1, E[u] = (0.8 − 0.6)·1 = 0.2 < 0.32: best unchanged.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.240.30.320.30.20.1

b = 0.7: P(win) = 1, E[u] = (0.8 − 0.7)·1 = 0.1 < 0.32: best unchanged.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.240.30.320.30.20.10

b = 0.8: P(win) = 1, E[u] = (0.8 − 0.8)·1 = 0 < 0.32: best unchanged.

00.10.20.30.40.50.60.70.800.2bE[u]00.140.240.30.320.30.20.10

Result: b = 0.4 = v/2, E[u] = 0.32. Bidding low kills the chance of winning; bidding high kills the profit.

def best_bid(v, n, step=0.1):
  beta = (n - 1) / n        # rivals bid beta*v
  best, b_best = -1, 0
  b = 0.0
  while b <= v:
    p = min(1, b / beta) ** (n - 1)  # P(win)
    eu = (v - b) * p
    if eu > best:
      best, b_best = eu, b
    b = round(b + step, 10)
  return b_best, best

Sealed-Bid Strategies

More Competition, Less Shading

000.20.20.40.40.60.60.80.811value vbid btruthful b = vn=2n=3n=5n=10
nb(v)v = 0.8 → bshading
21/2·v0.40.4
32/3·v0.530.27
54/5·v0.640.16
109/10·v0.720.08

With many bidders, winning requires bidding close to your value; as n → ∞, b → v.

Sealed-Bid Strategies

Bid Shading as the Number of Bidders Grows

Equilibrium bid b(v) = (n − 1)v / n  ·  shading = v − b(v) = v/n  ·  expected revenue = (n − 1)/(n + 1)

As the number of bidders grows, competition reduces shading and bids approach values; expected revenue is the same under first and second price (revenue equivalence) and rises with n.

Revenue and Information

Step by Step: Expected Revenue with Two Bidders = 1/3

RuleSeller's revenueExpected revenue
 
 
 
 
 

Two bidders with independent values v1, v2 ~ U[0, 1]. Question: which rule earns the seller more?

RuleSeller's revenueExpected revenue
Second pricemin(v1, v2)E[min] = ∫01 (1 − t)² dt = 1/3
 
 
 
 

Second price: everyone is truthful and the seller gets the smaller value. The smaller of two values exceeds t only if both do: P(min > t) = (1 − t)·(1 − t). Using E[X] = ∫01 P(X > t) dt, the expectation is 1/3.

RuleSeller's revenueExpected revenue
Second pricemin(v1, v2)E[min] = ∫01 (1 − t)² dt = 1/3
First pricemax(v1, v2) / 2E[max] = ∫01 (1 − t²) dt = 2/3 → 1/3
 
 
 

First price: the equilibrium bid is b = v/2 and the higher value wins. P(max < t) = t², so P(max > t) = 1 − t² → E[max] = 2/3, revenue 1/3.

RuleSeller's revenueExpected revenue
Second pricemin(v1, v2)E[min] = ∫01 (1 − t)² dt = 1/3
First pricemax(v1, v2) / 2E[max] = ∫01 (1 − t²) dt = 2/3 → 1/3
English≈ second price1/3
 
 

With private values the English auction gives the same result as second price → 1/3.

RuleSeller's revenueExpected revenue
Second pricemin(v1, v2)E[min] = ∫01 (1 − t)² dt = 1/3
First pricemax(v1, v2) / 2E[max] = ∫01 (1 − t²) dt = 2/3 → 1/3
English≈ second price1/3
Dutch≡ first price1/3
 

The Dutch auction is strategically equivalent to first price → 1/3. Four rules, the same expected revenue.

RuleSeller's revenueExpected revenue
Second pricemin(v1, v2)E[min] = ∫01 (1 − t)² dt = 1/3
First pricemax(v1, v2) / 2E[max] = ∫01 (1 − t²) dt = 2/3 → 1/3
English≈ second price1/3
Dutch≡ first price1/3
Simulation200,000 drawsSecond price 0.333 · first price 0.333

A Monte Carlo simulation confirms it: revenue differs auction by auction but is equal on average.

Revenue and Information

The Revenue Equivalence Theorem

  • With independent private values, risk-neutral symmetric bidders and the same allocation (highest value wins), standard auctions yield the same expected revenue (Vickrey 1961; Myerson 1981; Riley and Samuelson 1981).
  • Why? Myerson's payment formula p(v) = v·x(v) − ∫0v x(z) dz fixes the payment from x(·) and the lowest type's payment (0): same x(·) → same revenue. For n = 2, U[0, 1]: p(v) = v² − v²/2 = v²/2; first price: bid v/2 × win probability v; second price: v × expected rival value v/2. All agree.
  • Risk aversion, budgets, common values or asymmetry can break it.
nTheory: (n−1)/(n+1)Second price (simulation)First price (simulation)
21/3 = 0.330.3320.333
31/2 = 0.50.5000.500
43/5 = 0.60.6000.600
52/3 = 0.670.6680.667
So formats are chosen for simplicity, transparency and participation, not just revenue.

Revenue and Information

Change the Assumption: What If Bidders Are Risk-Averse?

Two bidders, values U[0, 1]. The winner's profit is x = v − price and their utility u(x) = xρ (0 < ρ ≤ 1). ρ = 1: risk-neutral; ρ < 1: risk-averse (ρ = ½ gives u(x) = √x).

If bidders are risk-averse, which rule earns the seller more?

Two bidders, values U[0, 1]. The winner's profit is x = v − price and their utility u(x) = xρ (0 < ρ ≤ 1). ρ = 1: risk-neutral; ρ < 1: risk-averse (ρ = ½ gives u(x) = √x).

If bidders are risk-averse, which rule earns the seller more?
Second price: no changeThe rival's bid, not yours, sets the price. Truthful bidding is still weakly dominant. Revenue: E[min] = 1/3.
First price: bids riseBidders raise their bids to reduce the risk of losing: b(v) = v / (1 + ρ). ρ = 1: v/2; ρ = ½: 2v/3.

Two bidders, values U[0, 1]. The winner's profit is x = v − price and their utility u(x) = xρ (0 < ρ ≤ 1). ρ = 1: risk-neutral; ρ < 1: risk-averse (ρ = ½ gives u(x) = √x).

If bidders are risk-averse, which rule earns the seller more?
Second price: no changeThe rival's bid, not yours, sets the price. Truthful bidding is still weakly dominant. Revenue: E[min] = 1/3.
First price: bids riseBidders raise their bids to reduce the risk of losing: b(v) = v / (1 + ρ). ρ = 1: v/2; ρ = ½: 2v/3.
Expected revenueρ = 1ρ = ½
Second price1/3 ≈ 0,3331/3 ≈ 0,333
First price (E[max] = 2/3)(2/3)·(1/2) = 1/3(2/3)·(2/3) = 4/9 ≈ 0,444
Risk aversion breaks revenue equivalence: first price earns more.

Revenue and Information

Revenue Maximization: Myerson and the Reserve Price

  • What if the seller wants to maximize expected revenue rather than efficiency? (Myerson, 1981)
  • Virtual value ψ(v) = v − (1 − F(v))/f(v). F(v): the probability that the value is below v; f(v): the density. For U[0, 1], F(v) = v, f(v) = 1 ⇒ ψ(v) = 2v − 1.
  • Expected revenue equals the expected virtual value of the winner. The optimal mechanism: give the item to the largest ψ, and do not sell if ψ < 0.
  • If ψ is increasing and bidders are symmetric, this is a second-price auction with a reserve price: ψ(r) = 0 ⇒ 2r − 1 = 0 ⇒ r = 1/2 (independent of n).
  • With two bidders and reserve r, expected revenue is 1/3 + r² − 4r³/3; at r = 1/2 it rises from 1/3 → 5/12 ≈ 0.417. The cost: sometimes nothing is sold (an efficiency loss).
00.51-101value vψ(v) = 2v − 1r = 1/2ψ < 0: don't sell

Revenue and Information

Code Trace: best_reserve()

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/3

Try reserve prices on the grid r = 0, 0.1, …, 1; revenue(r) averages the payment over a 100 × 100 grid of values.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,333

r = 0: revenue(r) = 0.333 → new best. No reserve = second price: 1/3.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,342

r = 0.1: revenue(r) = 0.342 → new best.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,363

r = 0.2: revenue(r) = 0.363 → new best.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,387

r = 0.3: revenue(r) = 0.387 → new best.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,408

r = 0.4: revenue(r) = 0.408 → new best.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,4080,417

r = 0.5: revenue(r) = 0.417 → new best.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,4080,4170,405

r = 0.6: revenue(r) = 0.405 < 0.417.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,4080,4170,4050,366

r = 0.7: revenue(r) = 0.366 < 0.417.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,4080,4170,4050,3660,291

r = 0.8: revenue(r) = 0.291 < 0.417.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,4080,4170,4050,3660,2910,171

r = 0.9: revenue(r) = 0.171 < 0.417.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,4080,4170,4050,3660,2910,1710

r = 1: revenue(r) = 0 < 0.417.

00.10.20.30.40.50.60.70.80.9100.2reserve rrevenue1/30,3330,3420,3630,3870,4080,4170,4050,3660,2910,1710

The best reserve is r = 0.5: revenue ≈ 0.417 = 5/12 > 1/3. A high r often leaves the item unsold.

def revenue(r, m=100):   # 2 bidders, U[0.1]
  v = [(i + .5) / m for i in range(m)]
  total = sum(max(min(a, c), r) for a in v
              for c in v if max(a, c) >= r)
  return total / m**2     # mean payment
def best_reserve(step=0.1):
  best, r_best = -1, None
  for k in range(round(1/step) + 1):
    R = revenue(k * step)
    if R > best: best, r_best = R, k * step
  return r_best, best

Revenue and Information

The Cost of a Reserve: Revenue Up, Efficiency Down

Two bidders, values U[0, 1], a second-price auction with reserve price r. If both values are below r, the item is not sold.

rExpected revenue
1/3 + r² − 4r³/3
Probability of no sale
r²
Expected welfare loss
2r³/3
01/3 ≈ 0,33300
0.253/8 = 0,3751/16 = 6.25%1/96 ≈ 0,010
0.55/12 ≈ 0,4171/4 = 25%1/12 ≈ 0,083
0.751/3 ≈ 0,3339/16 = 56.25%9/32 ≈ 0,281
  • No sale: both values below r: r · r = r².
  • Welfare loss: the item should go to the highest-value bidder; if that value is below r, it is lost. The highest value has density 2t: loss ∫0r t · 2t dt = 2r³/3.
  • Welfare without a reserve: E[max] = 2/3.
r = 0.5 maximizes revenue, but the item goes unsold with probability 25% and expected welfare falls from 2/3 to 7/12 (a 12.5% loss).

Revenue and Information

All-Pay and Multi-Unit Auctions

All-pay auction

The highest bid wins but everyone pays their bid: lobbying, R&D races, election campaigns.

Two bidders, U[0, 1]: in equilibrium b(v) = v²/2. Value 0.8 → bid 0.32 (0.4 under first price). Types with little chance of winning bid almost nothing.

Expected revenue is again 1/3 (revenue equivalence): a type-v bidder's expected payment is v²/2, as under first and second price.

Multi-unit auctions

With identical units (government bonds), winners pay either a uniform price per unit or their own bids; for items valuable together (spectrum licenses), package bids are used.

Further reading: Bulow and Klemperer (1996): attracting one extra bidder earns as much as the optimal reserve price, and often more.

The all-pay model describes contests with a single prize winner: when values are equal and commonly known, expected total spending equals the value of the prize.

Revenue and Information

The Winner's Curse

  • In a common-value auction, the highest bid comes from the bidder who overestimates the value the most.
  • If the winner ignores their estimation error, they may pay more than the item's true value.
  • Even if estimates are right on average, the largest estimate is systematically too high.
  • To reduce this risk, bidders bid below their signals and take the quality of their information into account.
  • First documented in Gulf of Mexico oil-lease auctions (Capen, Clapp, Campbell, 1971).
1234567891005101520number of firms nE[largest error]13.316.4

Estimation errors U[−20, 20]: the winner's expected overestimate is 20(n−1)/(n+1). More rivals, bigger curse (simulation gives the same values).

Revenue and Information

Step by Step: An Oil Lease Auction

7580859095100105110115120125true value V = 10082F191F296F3107F4124F5
CaseWinner / paymentProfit
 
 
 
 
 

The field's true value is V = 100 (nobody knows it). Five firms estimate 82, 91, 96, 107, 124. The mean is 100: the estimates are right on average.

7580859095100105110115120125true value V = 10082F191F296F3107F4124F5
CaseWinner / paymentProfit
First price, unshaded bidsF5 → 124−24
 
 
 
 

If everyone bids their estimate (first price), the most optimistic, F5, wins and pays 124 → profit −24.

7580859095100105110115120125true value V = 10082F191F296F3107F4124F5
CaseWinner / paymentProfit
First price, unshaded bidsF5 → 124−24
Second price, unshaded bidsF5 → 107−7
 
 
 

Even under second price F5 pays 107 → profit −7. Winning means answering yes to “am I the one who overestimated most?”

7580859095100105110115120125true value V = 10082F191F296F3107F4124F5
CaseWinner / paymentProfit
First price, unshaded bidsF5 → 124−24
Second price, unshaded bidsF5 → 107−7
Expected largest error (n = 5)20·4/6≈ 13.3
 
 

With errors U[−20, 20], the largest of five estimates is off by 20·4/6 ≈ 13.3 on average. Bidding your estimate under first price gives expected profit ≈ −13.3; under second price you pay the second-largest estimate, whose error averages 20·(n − 3)/(n + 1) = 20·2/6 ≈ 6.7, so ≈ −6.7.

7580859095100105110115120125true value V = 10082F191F296F3107F4124F5
CaseWinner / paymentProfit
First price, unshaded bidsF5 → 124−24
Second price, unshaded bidsF5 → 107−7
Expected largest error (n = 5)20·4/6≈ 13.3
Simulation, first price, unshaded bidsmean−13.3
Simulation, first price, shaded by 13.3mean0

First-price simulation (100,000 auctions): bidding your estimate averages a profit of −13.3; shading by 13.3 brings it to ≈ 0. With 10 firms the required shading is 16.4: more rivals, shade more.

Two-Sided Markets

Double Auctions

  • Many buyers and many sellers submit orders at once: buyers post bids, sellers post asks.
  • Stock exchanges, electricity markets and carbon permit markets work this way.
  • Matching: sort bids in decreasing and asks in increasing order; pairs match as long as bid ≥ ask.
  • Price: a single market price is chosen from the interval between the last matched pair and the first unmatched pair (e.g. the midpoint).
  • Total surplus (buyer + seller surplus) = Σ (bid − ask) over matched pairs → the largest possible value.

Two-Sided Markets

Step by Step: Matching in a Double Auction

020406080100number of tradesprice12345bidask70604085903555758050

Order book: buyers are willing to pay 70, 40, 90, 55, 80; sellers to accept 60, 85, 35, 75, 50.

020406080100number of tradesprice12345demand (buyers)supply (sellers)bids ↓asks ↑90358050706055754085

Bids are sorted decreasing (the demand curve), asks increasing (the supply curve).

020406080100number of tradesprice12345demand (buyers)supply (sellers)bids ↓asks ↑90358050706055754085

Pair 1: 90 ≥ 35 → match; this trade creates surplus 90 − 35 = 55.

020406080100number of tradesprice12345demand (buyers)supply (sellers)bids ↓asks ↑90358050706055754085

Pair 2: 80 ≥ 50 → match; this trade creates surplus 80 − 50 = 30.

020406080100number of tradesprice12345demand (buyers)supply (sellers)bids ↓asks ↑90358050706055754085

Pair 3: 70 ≥ 60 → match; this trade creates surplus 70 − 60 = 10.

020406080100number of tradesprice12345✗demand (buyers)supply (sellers)bids ↓asks ↑90358050706055754085

Pair 4: 55 < 75 → the buyer won't pay what the seller asks; stop. 3 trades take place.

020406080100number of tradesprice12345demand (buyers)supply (sellers)price = 65bids ↓asks ↑90358050706055754085

Price interval: [max(60, 55), min(70, 75)] = [60, 70]. The midpoint is price = 65; every matched trader trades at this price.

020406080100number of tradesprice12345demand (buyers)supply (sellers)price = 65bids ↓asks ↑90358050706055754085

Buyer surplus (90−65)+(80−65)+(70−65) = 45, seller surplus (65−35)+(65−50)+(65−60) = 50; total 95 = the largest possible total surplus.

Two-Sided Markets

Code Trace: double_auction(bids, asks)

b9080705540s6085357550

Sort bids from high to low: the most eager buyer first.

b9080705540s3550607585

Asks from low to high: the cheapest seller first.

b9080705540s3550607585▲ k = 0

k: number of matched pairs; n: length of the shorter list.

b9080705540s3550607585▲ k = 0

Condition: b[0] = 90 ≥ s[0] = 35? Yes.

b9080705540s3550607585▲ k = 1

Pair matched → k = 1.

b9080705540s3550607585▲ k = 1

Condition: b[1] = 80 ≥ s[1] = 50? Yes.

b9080705540s3550607585▲ k = 2

Pair matched → k = 2.

b9080705540s3550607585▲ k = 2

Condition: b[2] = 70 ≥ s[2] = 60? Yes.

b9080705540s3550607585▲ k = 3

Pair matched → k = 3.

b9080705540s3550607585▲ k = 3

Condition: b[3] = 55 ≥ s[3] = 75? No. The loop ends.

b9080705540s3550607585▲ k = 3

k ≠ 0: at least one trade, so compute the price.

b9080705540s3550607585▲ k = 3

lo = max(s[k−1], b[k]) = max(60, 55) = 60: the last seller must not lose, and the excluded buyer must not want to buy.

b9080705540s3550607585▲ k = 3

hi = min(b[k−1], s[k]) = min(70, 75) = 70.

b9080705540s3550607585▲ k = 3

Result: 3 trades at price (60 + 70)/2 = 65.

def double_auction(bids, asks):
  b = sorted(bids, reverse=True)  # buyers
  s = sorted(asks)                # sellers
  k, n = 0, min(len(b), len(s))
  while k < n and b[k] >= s[k]:
    k += 1                   # pair matched
  if k == 0:
    return 0, None
  lo = max([s[k-1]] + b[k:k+1])
  hi = min([b[k-1]] + s[k:k+1])
  return k, (lo + hi) / 2

Two-Sided Markets

Incentives in Two-Sided Markets

Impossibility

With independent values and costs whose ranges overlap, bilateral trade cannot achieve efficiency + IC + IR + budget balance together (Myerson–Satterthwaite, 1983).

The midpoint rule

Efficient and budget-balanced, but not IC: the marginal buyer can shade their bid to pull the price down.

McAfee's rule (1992)

Let p0 be the average of the first excluded buyer and seller. If p0 falls between the last matched buyer and seller, everyone trades at p0; otherwise the last (least profitable) pair is dropped. Truthfulness is dominant; the cost is sometimes one trade.

In our example

p0 = (55 + 75)/2 = 65 ∈ [60, 70] → all three trades happen at 65; no trade is lost.

Two-Sided Markets

Code Trace: mcafee(bids, asks)

bids basks spair 19030pair 28050pair 37065pair 44085

A new example (buyers 90, 80, 70, 40; sellers 30, 50, 65, 85): bids sorted decreasing, asks increasing; sentinels (−∞ and ∞) are appended.

bids basks spair 19030≥ ✓pair 28050pair 37065pair 44085

b[0] = 90 ≥ s[0] = 30? Yes: an efficient pair.

bids basks spair 19030pair 28050pair 37065pair 44085

k = 1: 1 efficient trade found.

bids basks spair 19030pair 28050≥ ✓pair 37065pair 44085

b[1] = 80 ≥ s[1] = 50? Yes: an efficient pair.

bids basks spair 19030pair 28050pair 37065pair 44085

k = 2: 2 efficient trades found.

bids basks spair 19030pair 28050pair 37065≥ ✓pair 44085

b[2] = 70 ≥ s[2] = 65? Yes: an efficient pair.

bids basks spair 19030pair 28050pair 37065pair 44085

k = 3: 3 efficient trades found.

bids basks spair 19030pair 28050pair 37065pair 44085< ✗

b[3] = 40 ≥ s[3] = 85? No: the loop ends.

bids basks spair 19030pair 28050pair 37065pair 44085p₀ = 62.5

Candidate price from the first excluded pair: p₀ = (40 + 85)/2 = 62.5.

bids basks spair 19030pair 28050pair 37065≥ ✓pair 44085p₀ = 62.5 ∉ [65, 70]

Is p₀ ∈ [s[2], b[2]] = [65, 70]? No (62.5 < 65): the 3rd seller would lose money.

bids basks spair 19030pair 28050pair 37065pair 440852 trades: buyers 70, sellers 65

Drop the weakest pair (70, 65): 2 trades; buyers pay 70, sellers receive 65. The leftover 2·5 = 10 stays with the market maker.

def mcafee(bids, asks):
  b = sorted(bids, reverse=True) + [float('-inf')]
  s = sorted(asks) + [float('inf')]
  k = 0
  while b[k] >= s[k]:
    k += 1                 # efficient trade
  if k == 0: return 0, None, None
  p0 = (b[k] + s[k]) / 2   # excluded pair
  if s[k-1] <= p0 <= b[k-1]:
    return k, p0, p0       # one price
  return k - 1, b[k-1], s[k-1]  # two prices

Two-Sided Markets

Auctions and Market Design

  • Market design decides who gets matched with whom, and under which rules.
  • The core goals are efficiency, incentive compatibility, fairness, stability and practicality; rules should be strategyproof.
  • Auctions are market-design tools that allocate scarce resources by prices and bids.
  • Electricity markets, emission permits, procurement and spectrum licenses use prices; in markets without money, priority and matching rules take the place of prices.
  • There can be trade-offs among revenue, efficiency, simplicity and strategyproofness.

Two-Sided Markets

From Theory to Practice: A Brief History

An ancient institution

Herodotus describes auctions in Babylon in the 5th century BC. In AD 193 the Praetorian Guard auctioned the Roman throne to Didius Julianus.

Theory

Vickrey (1961) compared the four formats and found revenue equivalence; Myerson (1981) characterized the revenue-maximizing auction. Nobel: Vickrey 1996, Myerson 2007.

Spectrum auctions

In 1994 the US FCC ran its first spectrum auction using the simultaneous multi-round ascending design proposed by Milgrom, Wilson and McAfee. Milgrom and Wilson: Nobel 2020.

Matching markets

Roth and Shapley (Nobel 2012): stable matching and exchange rules for markets that do not use prices.

The designer's tools: allocation and payment rules, information disclosure (which bids are shown, and when?), entry conditions. Designs are tested in lab experiments and with field data before deployment.

Assessment

Why Are Vickrey Auctions Rarely Used?

  • Trust: bidders must trust that the second-highest bid is reported honestly (risk of shill bids).
  • Intuition: “pay the second price” feels unnatural to most people; overbidding is common in the lab.
  • Model assumptions: with common or interdependent values, truthful bidding may no longer be dominant.
  • Multiple items: budget constraints and complementary goods complicate the payment rule (VCG).
  • Seller concerns: the risk of a low realized price or collusive bidding.
So practical designs often balance incentives for truthfulness against ease of use (e.g. the English auction reaches the same outcome transparently). Details: Rothkopf, Teisberg and Kahn (1990); Ausubel and Milgrom (2006).

Assessment

Discussion: Which Auction, and How Fair?

Which format would you rather join?
  • English: transparent, and seeing the price step by step is an advantage.
  • Vickrey: truthful bidding is best; no need to strategize.
  • Dutch / first price: fast, but you must guess how much to shade.
  • The choice depends on personal strategy and appetite for risk.
Are auctions and market design fair?
  • Rules are fixed in advance → predictability; manipulation can be reduced.
  • But deep-pocketed players have an advantage; unequal access and information asymmetry matter.
  • Seller revenue and equal opportunity are different goals; “perfect fairness for all” is not guaranteed.

Assessment

Summary: The Four Formats Side by Side

FormatPaysEquilibrium bid (IPV, U[0.1], n bidders)Equivalent to
Englishprice at which the last rival drops outstay in up to your valuesecond price
Second pricesecond-highest bidb = v (weakly dominant)English
Dutchthe accepted priceb = (n−1)v/nfirst price
First priceown bidb = (n−1)v/nDutch
Revenue equivalence

All four have expected revenue (n−1)/(n+1); 1/3 with two bidders.

Winner's curse

With common values the winner may overpay; shade more as the number of rivals grows.

Double auctions

Match bids in decreasing and asks in increasing order; the price comes from [lo, hi].

Assessment

Sample Problems

  1. Limits 40, 65, 80, 90. Under the four rules, who wins and what do they pay (everyone bids their limit)?
  2. n = 4 bidders, values U[0, 1]. What is the first-price equilibrium bid of a bidder with value 0.6?
  3. Three bidders, U[0, 1]: what is the expected revenue of a second-price auction?
  4. In Vickrey your value is 50 and the rivals' highest bid is 55. What happens if you bid 60?
  5. Buyers 100, 70, 45; sellers 30, 60, 80. How many trades, and what is the midpoint price?
  6. Two bidders, U[0, 1], second price, reserve r = 0.4. What is the expected revenue? More than with no reserve?
Show answers
  1. The 90 wins. English ≈ 80, second price 80; Dutch 90, first price 90.
  2. b = (4 − 1)/4 · 0.6 = 0.45.
  3. (n − 1)/(n + 1) = 2/4 = 1/2.
  4. You win but pay 55: utility 50 − 55 = −5. Bidding truthfully gives 0.
  5. 2 trades (100–30, 70–60); interval [max(60, 45), min(70, 80)] = [60, 70] → price = 65.
  6. 1/3 + r² − 4r³/3 = 1/3 + 0.16 − 0.0853 ≈ 0.408 > 1/3 ≈ 0.333. (Best is r = 0.5: 5/12 ≈ 0.417.)

Assessment

References (1/2)

  • V. Krishna, Auction Theory, 2nd ed., Academic Press, 2010 · P. Milgrom, Putting Auction Theory to Work, Cambridge University Press, 2004.
  • T. Roughgarden, Twenty Lectures on Algorithmic Game Theory, Cambridge University Press, 2016; course page: CS364A · N. Nisan, T. Roughgarden, É. Tardos, V. Vazirani (eds.), Algorithmic Game Theory, Cambridge University Press, 2007 (Chapters 9 and 13).
  • A. Karlin, Y. Peres, Game Theory, Alive, American Mathematical Society, 2017 (auction chapters) · D. Easley, J. Kleinberg, Networks, Crowds, and Markets, Cambridge University Press, 2010 (Chapter 9: Auctions).
  • W. Vickrey, “Counterspeculation, Auctions, and Competitive Sealed Tenders”, Journal of Finance, 1961 · R. Myerson, “Optimal Auction Design”, Mathematics of Operations Research, 1981.
  • J. G. Riley, W. F. Samuelson, “Optimal Auctions”, AER 71(3), 1981 · J. Bulow, P. Klemperer, “Auctions versus Negotiations”, AER 86(1), 1996.
  • M. H. Rothkopf, T. J. Teisberg, E. P. Kahn, “Why Are Vickrey Auctions Rare?”, J. Political Economy 98(1), 1990 · L. M. Ausubel, P. Milgrom, “The Lovely but Lonely Vickrey Auction”, Combinatorial Auctions, MIT Press, 2006.

Assessment

References (2/2)

  • R. P. McAfee, “A Dominant Strategy Double Auction”, Journal of Economic Theory, 1992 · E. Capen, R. Clapp, W. Campbell, “Competitive Bidding in High-Risk Situations”, Journal of Petroleum Technology, 1971.
  • J. H. Kagel, R. M. Harstad, D. Levin, “Information Impact and Allocation Rules in Auctions with Affiliated Private Values: A Laboratory Study”, Econometrica 55(6), 1987 · R. B. Myerson, M. A. Satterthwaite, “Efficient Mechanisms for Bilateral Trading”, Journal of Economic Theory 29(2), 1983.
  • Course pages: N. Shah, CSC304 (Toronto) · M. Dinitz (Johns Hopkins) · M. Hoefer (Frankfurt).

The End

Chapter 5: Auctions
Previous: Chapter 4: Mechanism Design · Next: Chapter 6: Online Platforms
1 / 1 Sercan KÜLCÜ. All rights reserved.