of Chicago

2 Texas

A&M University

Lixin University of Accounting and Finance

1 40

Firms, as consumers, are heterogeneous

2 40

Firms, as consumers, are heterogeneous Firm 1

Firm 2

Identity

Split from former vertically integrated utility

Municipal Utility

Physical assets

13 generating units ≈ 18, 000 MW of natural gas, coal and nuclear

2 generating units ≈ 500 MW of natural gas

Trader’s previous experience

1y “Director of Energy Trading” 4ys “Energy Trader” 3ys natural gas transportation & exchange firm

2ys trading desk at another firm 10ys “Superv. of System Operations” 8ys “System Operator” 4ys “System Operations Dispatcher” 4ys “Generation Control Operator”

2 40

Firms, as consumers, are heterogeneous

2 40

Motivation Efficiency concerns from an antitrust perspective: large firms • Exercise market power • Mergers and concentration • Texas market monitor: “small fish swim free” rule.

3 40

Motivation Efficiency concerns from an antitrust perspective: large firms • Exercise market power • Mergers and concentration • Texas market monitor: “small fish swim free” rule.

Should we worry about how small firms compete? But can firms compete in a way that creates inefficiency, in addition to those related to market power? (i.e. prevents least-cost dispatch) • Can differences in sophistication of pricing strategies cause

inefficiencies?

3 40

Motivation Efficiency concerns from an antitrust perspective: large firms • Exercise market power • Mergers and concentration • Texas market monitor: “small fish swim free” rule.

Should we worry about how small firms compete? But can firms compete in a way that creates inefficiency, in addition to those related to market power? (i.e. prevents least-cost dispatch) • Can differences in sophistication of pricing strategies cause

inefficiencies? This paper: What if all real-world firms were to engage in some strategic thinking, but some “fall short” of playing Nash equilibrium? • Heterogeneity in level of strategic thinking?

3 40

Strategic Sophistication and Efficiency • (Standard) “Sophisticated” Nash equilibrium bidding leads to

inefficiency, aka “market power”. • (Less Studied) Low level strategic thinking also inefficient • Hortaçsu and Puller (2008) study electricity auctions

Rich theory/lab literature on bounded rationality theory: Level-k, Cognitive Hierarchy, QRE. • In I.O., we have seen work on demand but almost nothing on

supply. • More in general, almost no application of level-k, CH, and QRE

using field data. Why? Identification.

4 40

Strategic Sophistication and Efficiency

Consider the “normal” I.O. approach • Differentiated product industries: MC → prices • Auctions: valuations → bids

Solution: field data on marginal cost • Enter electricity markets. . .

5 40

This paper • Same context as HP: bidding in the Texas electricity market • Our strategy • Embed a Cognitive Hierarchy (CH) model into a structural model of bidding • Exploit a dataset with bids and marginal costs to estimate levels of strategic sophistication • Why? (aka, what is new relative to HP?) • How heterogeneous is sophistication? • What is the impact of strategic sophistication on efficiency? • What are the (private) returns to strategic sophistication? • Bonus: Ability to calculate counterfactuals • In multi-unit auctions, solving for Nash equilibria is difficult/impossible (fixed point in function space) • The structure of the CH model makes finding equilibrium “easy” (sequence of best-responses)

6 40

Research Questions 1

What type of strategic behavior do we observe?

7 40

Research Questions 1

What type of strategic behavior do we observe? • Small firms are less sophisticated than large firms • Significant heterogeneity in sophistication

7 40

Research Questions 1

What type of strategic behavior do we observe? • Small firms are less sophisticated than large firms • Significant heterogeneity in sophistication

2

How much would an (exogenous) increase in strategic sophistication by a firm or group of firms affect the efficiency of the market?

7 40

Research Questions 1

What type of strategic behavior do we observe? • Small firms are less sophisticated than large firms • Significant heterogeneity in sophistication

2

How much would an (exogenous) increase in strategic sophistication by a firm or group of firms affect the efficiency of the market? • Increasing sophistication of small firms increases efficiency by

9–17%. Effects are smaller for larger firms.

7 40

Research Questions 1

What type of strategic behavior do we observe? • Small firms are less sophisticated than large firms • Significant heterogeneity in sophistication

2

How much would an (exogenous) increase in strategic sophistication by a firm or group of firms affect the efficiency of the market? • Increasing sophistication of small firms increases efficiency by

9–17%. Effects are smaller for larger firms. 3

Could mergers that increase strategic sophistication, but do not create cost synergies, increase efficiency?

7 40

Research Questions 1

What type of strategic behavior do we observe? • Small firms are less sophisticated than large firms • Significant heterogeneity in sophistication

2

How much would an (exogenous) increase in strategic sophistication by a firm or group of firms affect the efficiency of the market? • Increasing sophistication of small firms increases efficiency by

9–17%. Effects are smaller for larger firms. 3

Could mergers that increase strategic sophistication, but do not create cost synergies, increase efficiency? • Yes, but only if small firms involved; otherwise concentration effect

dominates.

7 40

Literature • Theory and lab: Costa-Gomez, Crawford and Broseta (2001), Crawford and Iriberri (2007), Camerer et al (2004), McKelvey and Palfrey (1995), Nagel (1995), Stahl and Wilson (1995), Gill and Prowse (2016).

• Empirical/field: Hortaçsu and Puller (2008), Gillen (2010), Goldfarb and Xiao (2011), An (2013).

• Electricity markets: Doraszelski, Lewis, and Pakes (2016), Fabra and Reguant (2014), Bushnell, Mansur and Saravia (2008), Sweeting (2007), Wolak (2003), Borenstein, Bushnell and Wolak (2002), Wolfram (1998).

• Productivity differences across firms: Syverson (2004), Hsieh and Klenow (2009), Bloom and Van Reenen (2007).

• Behavioral supply: Romer (2006), Massey and Thaler (2013), Ellison, Snyder, and Zhang (2016), DellaVigna and Gentzkow (2017).

8 40

Outline

1

Institutional setting

2

A Model of Non-Equilibrium Bidding Behavior

3

Data and Estimation

4

Counterfactuals: Increasing Sophistication

9 40

Institutional Setting

10 40

Texas Electricity Market - Early Years Timeline of Market Operations: • Generating firms sign bilateral trades with firms that serve

customers • Day-ahead: One day before production and consumption,

generating firms schedule a fixed quantity of production for each hour of the following day (‘day-ahead schedule’) • Day-of: shocks can occur (e.g. hotter July afternoon than

anticipated) • ‘Balancing Market’ to ensure supply and demand balance at

every point in time

11 40

Balancing Market Auction • Generation firms submit hourly bids to change production relative

to their ‘day-ahead schedule’ • Bids are monotonic step functions (up to 40 elbow points) for

portfolio of firm’s generators • Demand is perfectly inelastic • Uniform-price auction that clears every 15-minute interval with

hourly bids • Accounts for 2-5% of all power traded

12 40

How do firms do this?

13 40

How should firms choose price-quantity pairs? P

MCi

RD1

Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P

MCi

RD1 MR1 Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P

MCi

RD1 Q1

MR1 Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P

MCi

P1

A

RD1 Q1

MR1 Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P

MCi

A

RD2

RD1

MR2 MR1 Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P

MCi

A P2

B

RD2 Q2

RD1

MR2 MR1 Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P SBR

MCi

A P2

B

RD2 Q2

RD1

MR2 MR1 Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P SBR

MCi

RD2 QC

RD1

MR2 MR1 Can firms do this in practice?

Q

14 40

How should firms choose price-quantity pairs? P SBR

MCi

C Q

QC Can firms do this in practice?

14 40

How should firms choose price-quantity pairs? P SBR

QC

MCi

Incentives: Bid above MC for Q > QC (i.e., monopolist on residual demand) Bid below MC for Q < QC (i.e., monopsony) Q Can firms do this in practice?

14 40

Data Market Opens

8/1/2001

8,760 hourly auctions

SAMPLE PERIOD

8/1/2002

1/31/2003

15 40

Data Market Opens

8/1/2001

8,760 hourly auctions

SAMPLE PERIOD

8/1/2002

1/31/2003

For each hourly auction, we have data on: • Demand - perfectly inelastic balancing demand • Bids - each firm’s hourly firm-level (“portfolio”) bids • Marginal costs - each firm’s hourly MC of supplying balancing

power for plants that are “turned on”

MC Details

MC Figure

We focus on the 6–6:15pm periods with no transmission congestion.

15 40

What do we observe? Large firm

16 40

What do we observe? Large firm

Medium firm

16 40

What do we observe? Large firm

Medium firm

Consistent with best-responding to steeper RD

16 40

What do we observe? Small firm

Very Small firm

16 40

What do we observe? Small firm

Very Small firm

Can cause inefficient dispatch but not because of market power!

16 40

Summarizing Performance Across Firms Firm Reliant City of Bryan Tenaska Gateway Partners TXU Calpine Corp Cogen Lyondell Inc Lamar Power Partners City of Garland West Texas Utilities Central Power and Light Guadalupe Power Partners Tenaska Frontier Partners

Percent of Potential Profits Achieved 79% 45% 41% 39% 37% 16% 15% 13% 8% 8% 6% 5%

17 40

Ruling Out Alternative Explanations • Do bidding rules prevent firms from submitting ex post “best

response” bids? • No!

“Simple bidding rule”

• Are the dollar stakes large enough to justify the fixed costs of

submitting the “right” bids? • Money-on-the-table: between 3 and 18 million dollars per year.

• Startup costs? • All the units we consider in MC are already “on”. • Adjustment costs? • Flexible natural gas units often are marginal. • Inconsistent with Medium firm’s bid for quantities below contract position. • “Bid-ask” spread smaller for firms closer to best-response bidding despite having similar technology.

18 40

Ruling Out Alternative Explanations

• Is capacity overstated?: No, and even if it did it wouldn’t be a

problem when decreasing generation. • Transmission constraints: HP find cannot explain deviations. • Collusion: would be small players; monetary transfers unlikely.

19 40

A Model to Explain this Bidding Behavior: “Cognitive Hierarchy”

20 40

What Is “Hierarchical Thinking”? Imagine the following game: • Pick a number between 0 and 100 • Winner is player with number closest to

2 3

of average

• What is your number?

21 40

What Is “Hierarchical Thinking”? Imagine the following game: • Pick a number between 0 and 100 • Winner is player with number closest to

2 3

of average

• What is your number? • Level-1 thinking: If all other players pick 100, I should pick 67.

21 40

What Is “Hierarchical Thinking”? Imagine the following game: • Pick a number between 0 and 100 • Winner is player with number closest to

2 3

of average

• What is your number? • Level-1 thinking: If all other players pick 100, I should pick 67. • Level-2 thinking: If all other players use above reasoning, I should

pick 45.

21 40

What Is “Hierarchical Thinking”? Imagine the following game: • Pick a number between 0 and 100 • Winner is player with number closest to

2 3

of average

• What is your number? • Level-1 thinking: If all other players pick 100, I should pick 67. • Level-2 thinking: If all other players use above reasoning, I should

pick 45. • Level-3 thinking: If all other players use above reasoning.... • ...

21 40

What Is “Hierarchical Thinking”? Imagine the following game: • Pick a number between 0 and 100 • Winner is player with number closest to

2 3

of average

• What is your number? • Level-1 thinking: If all other players pick 100, I should pick 67. • Level-2 thinking: If all other players use above reasoning, I should

pick 45. • Level-3 thinking: If all other players use above reasoning.... • ... • Only rational and consistent choice is to choose 0 • People playing a game can have different levels of strategic

thinking

21 40

Cognitive Hierarchy Applied to this Market • Relaxes Nash assumption of ‘mutually consistent beliefs’. • Players differ in level of strategic thinking. • ki ∈ {0, . . . , K} • Level-0 players are non-strategic (Important assumption, I’ll

discuss it in detail in a couple of minutes)

22 40

Cognitive Hierarchy Applied to this Market

• Players level-1 to level-k are increasingly more strategic • level 1: assume all rivals are level 0. Best-respond to these beliefs. • level 2: assume rivals are distributed between level 0 and level 1.

Best respond to these beliefs. • ... • level k: assume rivals are distributed between level 0 and level k − 1.

Best respond to these beliefs. • Firms beliefs about their rivals’ level of strategic thinking is a

function of characteristics of those rivals (e.g. size)

23 40

Our model in pictures Assume F2 believes F1 to be type-0 P

Firm 1

P MC1

Firm 2 MC2

q1

q2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-0 P

Firm 1

P MC1

Firm 2 MC2

q1

q2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-0 P

S01

Firm 1

P MC1

QC

Firm 2 MC2

q1

q2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-0 P

S01

Firm 1

P

Firm 2

MC1

QC

MC2

q1 MR

RDq 2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-0 P

S01

Firm 1

P MC1

QC

Firm 2 S2 MC2

q1 MR

RDq 2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-1 P

Firm 1

P MC1

Firm 2 S2 MC2

q1 MR

RDq 2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-1 P

S11

Firm 1

P MC1

QC

Firm 2 S2 MC2

q1 MR

RDq 2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-1 P

S11

Firm 1

P MC1

QC

Firm 2 S2 MC2

q1

MR′

MR RD′

RDq 2

Model in Math

24 40

Our model in pictures Assume F2 believes F1 to be type-1 P

S11

Firm 1

P MC1

QC

q1

Firm 2 S2 ′ S2

MC2

q2

Model in Math

24 40

Our model in pictures Higher-type rivals rotate RD and induce more competitive bidding P

S11

Firm 1

P MC1

QC

q1

Firm 2 S2 ′ S2

MC2

q2

Model in Math

24 40

Identification Suppose larger firms are higher types (γ > 0) Rival-Small/Low Type Rival-Large/High Type P P S MC S MC

q P

q

Firm i’s RD

q

25 40

Identification Suppose larger firms are higher types (γ > 0) Rival-Small/Low Type Rival-Large/High Type P P S MC S MC

q P

q

Firm i’s RD RDγ>0

q

25 40

Identification Suppose larger firms are lower types (γ < 0) Rival-Small/High Type Rival-Large/Low Type P P S MC S MC

q P

q

Firm i’s RD

q

25 40

Identification Suppose larger firms are lower types (γ < 0) Rival-Small/High Type Rival-Large/Low Type P P S MC S MC

q P

q

Firm i’s RD

RDγ<0 q

25 40

Identification Suppose larger firms are lower types (γ < 0) Rival-Small/High Type Rival-Large/Low Type P P S MC S MC

q P

q

Firm i’s RD RDγ>0 Is i’s bid more consistent with RDγ>0 or RDγ<0? RDγ<0 q

25 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be

P S0i

QC

SBR i

MCi

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

QC

SBR i

MCi

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

QC

SBR i

MCi

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

QC

SBR i

• not observed

MCi

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

SBR i

• not observed

MCi • Bid marginal costs

QC

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

SBR i

• not observed

MCi • Bid marginal costs

QC

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

SBR i

• not observed

MCi • Bid marginal costs • bids would have to be flatter

than BR, not observed

QC

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

SBR i

• not observed

MCi • Bid marginal costs • bids would have to be flatter

than BR, not observed • Bid vertical

QC

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

SBR i

• not observed

MCi • Bid marginal costs • bids would have to be flatter

than BR, not observed • Bid vertical

QC

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i SM i

SBR i

• not observed

MCi • Bid marginal costs • bids would have to be flatter

than BR, not observed • Bid vertical

QC

Q

26 40

More on level-0 firms In general, level-0s are non-strategic players. In our setting, this can be • Bid randomly

P S0i

SBR i

• not observed

MCi • Bid marginal costs • bids would have to be flatter

than BR, not observed • Bid vertical

QC

Q

• higher types would bid

flatter and approach BR from the left, as we observe

26 40

Corroborating “Reduced-Form” Evidence of Non-strategic Behavior Publicly Observable Shock – Nuclear Generator Went Off-line

Descriptive regressions find: • Large firms respond to own cost shocks and cost shocks of

competitors • Small firms only respond to own cost shocks

27 40

Corroborating “Reduced-Form” Evidence of Non-strategic Behavior

Outage

Largest Six -26.27* (4.69)

Smallest Six -0.64 (0.42)

3.75* (0.32)

Largest Six -9.80* (2.92) 0.27* (0.03) 2.82 (2.41)

Smallest Six 0.4 (0.38) 0.18* (0.02) 0.19 (0.37)

Largest Six -8.40* (2.05) 0.30* (0.03) -21.13* (6.55)

Smallest Six -0.03 (0.25) 0.11* (0.02) 0.76* (0.21)

40.28* (4.49) No

No

No

No

Yes

Yes

378 0.09

378 0.01

378 0.40

378 0.31

378 0.67

378 0.68

Own MC Constant

Bidder Fixed Effects N R2

Note: Each column reports estimates from a separate regression of the slope of a firm’s bid function on an indicator variable that the auction occurred during the fall 2002 nuclear out∂S

age. An observation is a firm-auction. The dependent variable is the slope ( ∂pit ) of firm i’s bid in auction t where the slope is linearized plus and minus $10 around the market-clearing price. Own MC is the slope of the firm’s own marginal cost function linearized plus and minus $10 around the market-clearing price. White standard errors are reported in parentheses. + p<0.05, * p<0.01

28 40

Estimation

29 40

Estimation: Information Firm type: ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ). • ki is private information • τi is public information.

Costs: public information. ki and size−i determine i’s beliefs about −i’s types. i best-responds to those beliefs. We compute i’s best response for each k and minimize the distance between predicted bids and the data.

30 40

Estimation: Minimum-distance approach P S0i

S1i

SData

S2i S3i

MCi

QC

Q

31 40

Estimation: Minimum-distance approach P S0i

S1i

SData

S2i S3i

P1

MCi

P0 QC

Q

31 40

Estimation: Minimum-distance approach P S0i

S1i

SData

S2i S3i

P1

MCi

P0 QC

Q

31 40

Results

32 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Larger Firms Are Higher Type ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei ) 1

0.9

0.8 Installed capacity

Probability

0.7

relative to largest firm 11% 22% 28% 36% 44% 56% 54% 69% 78% 80% 87% 100%

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

33 40

Manager Training Matters (1)

(2)

(3)

Constant

-0.726 (0.087)

-0.749 (0.106)

-3.493 (0.414)

Size

14.594 (1.027)

13.619 (1.188)

3.090 (0.755)

AAU School

0.376 (0.065)

Econ/Business/Finance degree Number of auctions

5.626 (1.188) 99

Note: Bootstrapped standard errors using 45 samples. Model fit

34 40

Learning? 1

0.9

0.8

Probability

0.7

0.6

0.5

First week Last week

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

Small Firm - Estimated Type Distribution with Learning (Size and time trend specification)

35 40

Learning? 1

0.9

0.8

Probability

0.7

0.6

0.5

First week Last week

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

12

14

16

18

20

Type

Big Firm - Estimated Type Distribution with Learning (Size and time trend specification) More on learning: Quantity offered did not change over time

36 40

Simulations of Changes in Sophistication 1

“Consulting Firm”

2

Merger

37 40

Increasing Sophistication Decreases Costs Changes in average generating costs:

Counterfactual

INC side Public Private

DEC side Public Private

Small firms to median Above median firms to highest Three smallest to median

38 40

Increasing Sophistication Decreases Costs Changes in average generating costs:

Counterfactual Small firms to median Above median firms to highest Three smallest to median

INC side Public Private

DEC side Public Private

-9.43%

38 40

Increasing Sophistication Decreases Costs Changes in average generating costs:

Counterfactual Small firms to median Above median firms to highest Three smallest to median

INC side Public Private

DEC side Public Private

-9.43% -4.49%

38 40

Increasing Sophistication Decreases Costs Changes in average generating costs:

Counterfactual

INC side Public Private

Small firms to median Above median firms to highest Three smallest to median

-9.43% -4.49% -6.97%

DEC side Public Private

38 40

Increasing Sophistication Decreases Costs Changes in average generating costs:

Counterfactual

INC side Public Private

Small firms to median Above median firms to highest Three smallest to median

-9.43% -4.49% -6.97%

DEC side Public Private

-7.70% -3.72% -6.04%

38 40

Increasing Sophistication Decreases Costs Changes in average generating costs:

Counterfactual

INC side Public Private

Small firms to median Above median firms to highest Three smallest to median

-9.43% -4.49% -6.97%

-7.70% -3.72% -6.04%

DEC side Public Private -16.49% -7.96% -10.70%

-13.91% -6.50% -9.94%

38 40

Mergers that Increase Sophistication

Mergers only reduce generation costs when small firms are involved

Smallest and largest firms Median and largest firms Two largest firms

INC side

DEC side

-3.2% +9.0% +17.3%

-15.4% +21.9% +56.3%

39 40

Conclusions and Takeaway Messages Does heterogeneity in strategic sophistication affect market performance? • Context: bidding into electricity auctions in Texas. • First paper using field data to study pricing decisions. • To model pricing decisions, we embed a CH model into a

structural model of bidding. Takeaways: 1

2

Significant heterogeneity in sophistication. Larger firms are more sophisticated than smaller firms. Does sophistication matter? Yes! • Increasing sophistication improves efficiency. • Most of the gains come from smaller firms.

3

Could mergers that increase sophistication, but do not create cost synergies, increase efficiency? • Yes, but only if small firms are involved.

40 40

Thank you

Appendix

Main players in generation Firm TXU Reliant City of San Antonio Central Power & Light City of Austin Calpine Lower Colorado River Authority Lamar Power Partners Guadalupe Power Partners West Texas Utilities Midlothian Energy Dow Chemical Brazos Electric Power Cooperative Others Back

% of installed capacity 24 18 8 7 6 5 4 4 2 2 2 1 1 16

Can Firms Do This in Practice? • Grid operator reports aggregate bid function with a 2 day lag • Simple trading rule • Download bid data from 2 days ago • Assume rivals do not change their bids • Calculate best response to lagged rivals’ bids

• Does this outperform actual bidding? • Answer: Yes and it yields almost the same profits as best response

to current rivals’ bids Back

Firm performance relative to best-responding Percent achieved by Actual bids BR to lagged bids Reliant City of Bryan Tenaska Gateway TXU Calpine Cogen Lyondell Lamar Power Partners City of Garland West Texas Utilities Central Power and Light Guadalupe Power Partners Tenaska Frontier

Source: Hortaçsu and Puller (2008).

79% 45% 41% 39% 37% 16% 15% 13% 8% 8% 6% 5% Back

98.5% 100% 99.6% 96.7% 97.9% 100% 99.6% 99.6% 100% 98.7% 99% 99.3%

Measuring Marginal Cost • Each unit’s daily capacity & day-ahead schedule • Marginal Costs for each fossil fuel unit • Fuel costs – daily natural gas spot prices (NGI) & monthly average coal spot price (EIA) • Fuel efficiency – average “heat rates” (Henwood) • Variable O&M (Henwood) • SO2 permit costs (EPA) • Use coal and gas-fired generating units that are “on” that hour and

the daily capacity declaration (Nukes, Wind, Hydro may not have ability to adjust) • Calculate how much generation from those units is already

scheduled == Day-Ahead Schedule

Measuring Marginal Cost P

MW

Back

Measuring Marginal Cost P

Total MCi

MW

Back

Measuring Marginal Cost P

Total MCi

MW Day ahead schedule Back

Measuring Marginal Cost MCi Auction

P

Total MCi

MW Day ahead schedule Back

Model: Details • Market clearing price pct : N

∑ Sit (pct , QCit ) = Dt (pct ) + ε t

(1)

i= 1

• Three sources of uncertainty • Demand shock (ε t ) • Rival Contract positions (QC−it ) • Rival Types (k−i )

Hit (p, Sˆ it (p); ki , QCit ) ≡ Pr(pct ≤ p|Sˆ it (p), ki , QCit ) Back

(2)

Model: Details • Market clearing price pct : N

∑ Sit (pct , QCit ) = Dt (pct ) + ε t

(1)

i= 1

• Three sources of uncertainty • Demand shock (ε t ) • Rival Contract positions (QC−it ) • Rival Types (k−i )

Hit (p, Sˆ it (p); ki , QCit ) ≡ Pr(pct ≤ p|Sˆ it (p), ki , QCit ) Back

(2)

Model: Details Combining (1) and (2) and denoting i’s private information Ωit ≡ {ki , QCit }: Hit (p, Sˆ it (p); Ωit ) = Z

QC−it ,l−i ,ε t

aggregate supply }| { z l ˆ ˆ 1 ∑ Sjt (p, QCjt ; ki ) + Sit (p) ≥ Dt (p) + ε t dF(QC−it , l−i , ε t |Sit (p), Ωit )

j6 =i

F(QC−it , l−i , ε t |Sˆ it (p), Ωi ): the joint density of each source of uncertainty from the perspective of firm i. Let θi ≡ ∑j6=i Sljt (·; ki ) − ε ∼ Γi .

Back

Model: Details

The firm’s problem max

Z p

Sˆ it (p) p

U p · Sˆ it (p) − Cit Sˆ it (p) − (p − PCit )QCit dHit p, Sˆ it (p); Ωit

Necessary condition for optimality: p − Cit′ (Sit∗ (p))

Back

=

(Sit∗ (p) − QCit )

Hs p, Sit∗ (p); ki , QCit Hp p, Sit∗ (p); ki , QCit

(3)

Model: Details

The firm’s problem max

Z p

Sˆ it (p) p

U p · Sˆ it (p) − Cit Sˆ it (p) − (p − PCit )QCit dHit p, Sˆ it (p); Ωit

Necessary condition for optimality: p − Cit′ (Sit∗ (p))

Back

=

(Sit∗ (p) − QCit )

Hs p, Sit∗ (p); ki , QCit Hp p, Sit∗ (p); ki , QCit

(3)

Why is Assumption 1 important? 1

It implies that residual demand is flatter for higher type.

2

No more assumptions needed about how private information enters the bid functions.

Why? Consider a level-1 bidder

where θit ≡ ∑j6=i QCjt − ε t .

Back

Why is Assumption 1 important? 1

It implies that residual demand is flatter for higher type.

2

No more assumptions needed about how private information enters the bid functions.

Why? Consider a level-1 bidder Hit (p, Sˆ it (p); k = 1, QCit ) =

Z

QC−it ,l−i ,ε t

1(∑ S0jt (p, QCjt ) + Sˆ 1it (p) ≥ j6 =i

Dt (p) + ε t )dF(QC−it , l−i , ε t |Sˆ 1it (p), ki = 1, QCit )

where θit ≡ ∑j6=i QCjt − ε t .

Back

Why is Assumption 1 important? 1

It implies that residual demand is flatter for higher type.

2

No more assumptions needed about how private information enters the bid functions.

Why? Consider a level-1 bidder Hit (p, Sˆ it (p); k = 1, QCit ) =

Z

QC−it ,l−i ,ε t

1(∑ S0jt (p, QCjt ) + Sˆ 1it (p) ≥ j6 =i

Dt (p) + ε t )dF(QC−it , l−i , ε t |Sˆ 1it (p), ki = 1, QCit ) Assumption 1

=

Z

QC−it ,l−i ,ε t

1( ∑ j6 =i

z}|{ QCjt

− εt ≥

Dt (p) − Sˆ 1it (p))dF(QC−it , l−i , ε t |Sˆ 1it (p), ki = 1, QCit )

where θit ≡ ∑j6=i QCjt − ε t .

Back

Why is Assumption 1 important? 1

It implies that residual demand is flatter for higher type.

2

No more assumptions needed about how private information enters the bid functions.

Why? Consider a level-1 bidder Hit (p, Sˆ it (p); k = 1, QCit ) =

Z

QC−it ,l−i ,ε t

1(∑ S0jt (p, QCjt ) + Sˆ 1it (p) ≥ j6 =i

Dt (p) + ε t )dF(QC−it , l−i , ε t |Sˆ 1it (p), ki = 1, QCit ) Assumption 1

=

Z

QC−it ,l−i ,ε t

1( ∑ j6 =i

z}|{ QCjt

− εt ≥

Dt (p) − Sˆ 1it (p))dF(QC−it , l−i , ε t |Sˆ 1it (p), ki = 1, QCit )

=

where θit ≡ ∑j6=i QCjt − ε t .

Z

1(θit ≥ QC−it ,l−i ,ε t Dt (p) − Sˆ 1it (p))dF(QC−it , l−i , ε t |Sˆ 1it (p), ki

Back

= 1, QCit )

We can do the same for type 2 But now Hit (p, Sˆ it (p); ki = 2, QCit ) =

Z

QC−it ×l−i × ε t

∑

1(

∑

QCjt +

j6 =i∈l0

j6 =i∈l1

S1jt (p, QCjt ) − ε t ≥

Dt (p) − Sˆ 2it (p))dF(QC−it , l−i , ε t |Sˆ 2it (p), ki = 2, QCit )

=

(4)

Z

1(θit ≥ QC−it ×l−i × ε t Dt (p) − Sˆ 2it (p))dF(QC−it , l−i , ε t |Sˆ 2it (p), ki

where, θit = ∑j6=i∈l0 QCjt + ∑j6=i∈l1 S1jt (p, QCjt ) − ε t . We can do this recursively for all types.

Back

= 2, QCit )

Model: Details Let

Γ(·): the conditional distribution of θit (conditional on N − 1 type draws). ∆(l−i ): the marginal distribution of the vector of rival firm types. Then H (·) becomes Hit (p, Sˆ it (p); ki , QCit ) = And

HS Hp

Z

l−i

h

i 1 − Γ Dt (p) − Sˆ kit (p) · ∆(l−i )

becomes

R ˆk Hs p, Sit∗ (p); ki , QCit l−i γ Dt (p) − Sit (p) · ∆(l−i ) = R . Hp p, Sit∗ (p); ki , QCit − l γ Dt (p) − Sˆ kit (p) Dt′ (p)∆(l−i ) −i

Back

Model: Details Assumption 2: ∆(·) is an independent multivariate Poisson distribution truncated at k − 1, as given by Poisson Cognitive Hierarchy model. Assumption 3: Γi is a uniform distribution. (We can relax but adds to computational burden) First-order condition simplifies to the “inverse elasticity rule”: p − Cit′ Sˆ kit (p) =

i i h h 1 1 ∗ Sˆ kit (p) − QCit = ∗ Sˆ kit (p) − QCit , ′ ′ −Dt (p) −RDt (p)

where the second equality follows from the fact that RD(p) = D(p) + ε − ∑j6=i Sjt (p) = D(p) + ε − ∑j6=i QCjt . Hence, RD′ (p) = D′ (p) for all p. Back

Objective function

ω (γˆ ) =

"

h

∑∑ ∑ ∑ i

t

k

p

bdata (p) − bmodel (p|k) 2 it it bmodel (p|K) − bmodel ( p |0 ) it it

i × P (p) Pi (k| |K|, γˆ )

P (p) → price points weighted by triangular distribution centered at market-clearing price Pi (k| |K|, γˆ ) → weight by probability of a firm being each type Back

#

Estimated Type Distributions ki ∼ Poisson(τˆi ), τˆi = exp(γˆ 0 + γˆ 1 sizei + γˆ 2 size2i ) 1

0.9

0.8 Installed capacity

Probability

0.7

relative to largest firm 11% 22% 28% 36% 44% 56% 54% 69% 78% 80% 87% 100%

0.6

0.5

0.4

0.3

0.2

0.1

0 0

2

4

6

8

10

Type Back

12

14

16

18

20

Model fit: CH vs. Unilateral Best-Response Dependent Variable: Profits from Actual Bids Profits under Cognitive Hierarchy

Profits under Best-Response

Constant

Observations R2

(1) CH Model

(2) Best-Response

(3)

0.803 (0.069)

– –

0.642 (0.127)

– –

0.428 (0.044)

0.137 (0.062)

-328.17 (141.976)

-241.74 (120.722)

-374.167 (125.785)

1058 0.67

1058 0.49

1058 0.69

Note: This table reports results from a regression of observed profits from actual bidding behavior on either firm profits as predicted by the Cognitive Hierarchy model (column 1), firm profits that would be achieved from a model of unilateral bestresponse to rival bids (column 2), or both. An observation is a firm-auction. Standard errors clustered at the firm-level are reported in parentheses. Back

More evidence on no learning Offered Quantities into Market in Year 2 vs Year 1

Year 2

Firm Fixed Effects INC Fixed Effects Day of Week Fixed Effects Observations R2 + p<0.05; ∗ p<0.01.

All Firms (1)

All Firms (2)

All Firms (3)

Small Firms (4)

-34.76 (42.42)

-15.85 (34.24)

-16.15 (34.70)

1.52 (2.90)

Yes No No

Yes Yes No

Yes Yes Yes

Yes Yes Yes

2264 0.01

2264 0.03

2264 0.04

1029 0.09

The dependent variable Participation Quantityit is the megawatt quantity of output bid at the market-clearing price relative to the firm’s contract position in auction t, i.e. |Sit (pmcp ) − QCit |. The sample period is the first 1.5 years of the market and Year 2 is a dummy variable for the second year. Standard errors clustered at the firm-level are reported in parentheses. Back

Diminishing Returns to Sophistication

400

50

200

Incremental profits, US dollars

100

Small firm Medium firm 0 0.02

0.04

0.06

0.08

0.1

0.12

x-axis includes range from smallest to largest firm Back

0.14

0.16

0.18

0 0.2

Incremental profits, US dollars

INC side

Diminishing Returns to Sophistication DEC side

Incremental profits, US dollars

500

Incremental profits, US dollars

200

Small firm Medium firm 0 0.02

0.04

0.06

0.08

0.1

0.12

x-axis includes range from smallest to largest firm Back

0.14

0.16

0.18

0 0.2