27
PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague, 2012 PDF with color figures at www.econ.ohio-state.edu/jhm/papers/PQNash.pdf

PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Embed Size (px)

Citation preview

Page 1: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

PQ-Nash Duopoly:A Computational Characterization

J. Huston McCullochOhio State University

18th Int’l. Conf. on Computing in Economics and FinancePrague, 2012

PDF with color figures at www.econ.ohio-state.edu/jhm/papers/PQNash.pdf

Page 2: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Cournot vs Bertrand Duopoly

Cournot (1838): Each chooses P and Q, taking other’s Q as given

Bertrand (1883): Each chooses P and Q, taking other’s P as given

Page 3: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Cournot vs Bertrand Duopoly

Cournot (1838): Each chooses P and Q, taking other’s Q as given

Bertrand (1883): Each chooses P and Q, taking other’s P as given

Neither takes other’s full (P, Q) strategy as given: Both half right, half wrong!

Page 4: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Duopoly Model:

2 firms produce Q1, Q2 of identical good

Increasing Marginal (Variable) Cost schedules MC1(Q1), MC2(Q2)Decreasing Demand function D(P)

Atomistic, price-taking consumersOutput perishable at end of period, but freely disposableProducers can’t price discriminateFull information – each knows D(P), other’s MC“Parallel Rationing”:

If Pi < Pj , Firm i sells out first to highest-demand consumers, Firm j gets residual demand. (as if curb market with free entry efficiently allocates total Q to consumers)

Page 5: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Simple illustrative straight-line model:

MC1(Q1) = Q1/a, MC2(Q2) = Q2/(1-a), a (0, .5],

so a is share of smaller firm in horizontally summed MC

D(P) = 1 + e – e P,

so P = 1, Q = 1 is quasi-competitive (P-taking) equilibrium

and e > 0 is absolute price elasticity of D at P = 1, Q = 1.

Page 6: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Eg a = 1/3, e = 1:

With 2-plant monopoly (+) and quasi-competitive () equilibria.

Page 7: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Some well-known Duopoly models:

But none is a 1-stage Nash equilibrium

Page 8: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Cournot (1838) popularly said to be Q-Nash

eg by Singh and Vives (1984)

Firm 1’s decision, taking Q2 given

Page 9: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

But in fact firms can and do set P as well as Q. Firm 1 would undercut P2 and raise Q1, so not simple Nash.

Cournot is, however, outcome of 2-stage Q-then-P Nash game – Kreps and Scheinkman (1983).

Page 10: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Bertrand (1883) popularly thought to be P-Nash

Firm 1’s decision, taking P2 as given

Page 11: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

But in fact, firms can and do set Q as well as P. Firm 1 would reduce Q and thereby raise P1 above P2, so Bertrand not simple Nash

Is there a 1-stage PQ Nash equilibrium?

Page 12: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

No PQ-Nash equilibrium in pure strategies:

Can find P2, Q2 so 3 intersections of MC1 and MR1: Middle one clears market

Page 13: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

No PQ-Nash equilibrium in pure strategies:

Can find P2, Q2 so 3 intersections of MC1 and MR1: Middle one clears market, but is local profit minimum. Left and right intersections are maxima but don’t clear market.

Is there an equilibrium in mixed strategies?

Page 14: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Nash (1950):

Every finite bimatrix game has at least 1 equilibrium in pure or mixed strategies.

Page 15: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Computing in Economics and Finance:

One equilibrium of discrete approximation to bimatrix game may be found numerically using algorithm of Mangasarian & Stone (1964).

(Matlab thanks to Bapi Chatterjee 2008)

With n strategies each, takes about n3 flops:n iterations, n2 flops each

But with h x h grid of (P, Q) strategies, each has n = h2 strategies

So takes about h6 flops!

h = 11: several secondsh = 15: several minutesh = 21: several hours

Page 16: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

With mixed strategies, combined output may exceed or fall short of D.

Let qi = amount sold by firm i, Qi = amount produced, as above.

If Pi < Pj, “parallel rationing:” qi = min(Qi, D(Pi)), qj = min(Qj, max(0, D(Pj) – qi)).

In rare event Pi = Pj and Q1 + Q2 > D(P), assume sales proportional to production: qi = min(Qi, D(P) Qi / (Qi + Qj)).

Then i = Pi qi – Ci(Qi).

Page 17: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Solution generally is high “regular” price with low Q, plus string of low “sale” prices with higher Q’s:

2-step 15x15 grid for each with ½ step offset in P to reduce ties.Probabilities proportional to area of symbols, = ¼ in legend.

Page 18: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

D moreelastic

D less elastic

Firm 1verysmall

Firms equal in size

Same story with other parameter values:

Note that max P is always less than Cournot, while min P always > Bertrand.But sometimes unsold Q. Which more efficient?

Page 19: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Robt. Gertner (unpublished 1986 MIT dissertation!)

• Finds FOC for PQ-Nash problem!• Support for Pi

is [Pmin, Pmax], same for both firms• Qi conditional on Pi is unique• Mass point at Pmax, continuous distribution below• If both charge Pmax market clears; else, overproduction• Gives differential eq’n for density, but “intractable”• Gives no solution, even for linear or const. elast. schedules• Only treats symmetric case MC1(Q) = MC2(Q)

Now at Chicago Booth School My nomination for next Nobel Prize!

Page 20: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Valid simultaneous move Nash Duopoly models are now: • Gertner PQ 1-stage accounts for random sales!• Cournot-Kreps-Scheinkman Q-then-P 2-stage

generalized by Wu, Zhu & Sun IJIO 2012• Production-to-Order P-Qmax 1-stage (next slide)

Levitan and Shubik (1978) anticipate Gertner with zero MC and costly disposal (or storage)

Kreps and Sheinkman (1983) anticipate Gertner with vertical MC in 2nd stage game à la Edgeworth (1925) but KS requires parallel rationing – Davidson & Deneckere (1986)

Singh and Vives (1984) – firms artificially constrained to either pick Pi and meet all D at that Pi or pick Qi and sell at D-1(Q1+Q2)

Page 21: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Production-to-Order (PTO) PQ Duopoly

Each firm simultaneously sets Pi and Qimax

Then passively takes orders qi up to Qimax

Produces Qi = qi, so never overproductionCould be underproduction in mixed equilibrium

If Pi < Pj, qi = min(Qi

max, D(Pi)) qj = min(Qj

max, max(D(Pj)-qi, 0))

If Pi = Pj, assume (less compellingly) qi = min(Qi

max, D(P) Qimax / (Qi

max + Qjmax)

Then i = Pi qi – Ci(qi).

Page 22: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Computational example of PTO Duopoly with a = 1/3

Primary strategy of smaller firm now lowest price charged by either, not highest price as in Gertner PQ Duopoly.(There are also pure NE with P = 1, but these are dominated.)

Page 23: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

PTO Duopoly with a = ½:

Multiple pure Nash Equilibria arise, with P1 = P2 = 1 up to P1 = P2 = 1.065 (with e = 1)

High price equilibrium is “Superlative Nash Eq” ie the optimal one for both firms, so would be chosen

I conjecture this pattern is valid in general: Mixed strategies if a < ½ with smaller asymmetrically favoring lowest price Converges on pure strategies as a ½ Lower avg. P than Gertner

Page 24: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Unsolved issues: 1. Solve Gertner differential eq’n for simple cases:

Linear MC, affine D Qi|Pi affine?Constant elasticity D with Cobb-Douglas or CES MVC (w/ 2nd factor fixed)Or at least quick, accurate algorithm

2. FOC for asymmetric MCsThen solve for simple cases

3. Find FOC for PTO duopolyThen solve for simple cases

4. Solve PQ-Stackelberg modelApproximates Gertner?

Application: (McC 1993) Costs vs benefits of antitrust with increasing returns to scale, sunk factor, duopoly alternative

Page 25: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Q: What is a duopolist’s favorite music genre?

Page 26: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Q: What is a duopolist’s favorite music genre?

A: Doo-wop!

Page 27: PQ-Nash Duopoly: A Computational Characterization J. Huston McCulloch Ohio State University 18 th Int’l. Conf. on Computing in Economics and Finance Prague,

Stackelberg ModelsPseudo-Nash arguments invalid

Q-Stackelberg – 3-stage sequential game valid“Leader” chooses QL, then“Follower” chooses QF, then P’s determined by Kreps-Scheinkman game

P-Stackelberg (McC 1993)Leader chooses PL, QL, while Follower chooses PF, QF as a fn of PL only. inconsistent perception of Leader’s strategy space

PQ-Stackelberg (I haven’t seen this proposed or implemented)Leader chooses PL, QL, then Follower chooses PF, QF as a fn of both PL, QL.However, Leader can induce Follower to be P- or Q-taker -- High QL and/or PL – F acts as P-taker (as in Fig. 4) Low QL and/or PL – F acts as Q-taker (as in Fig. 6) At boundary, F is indifferent (as in Fig. 7), so easy to find. So outcome conceivably Q- (or P-) Stackelberg!?

But might induced zone changes alter outcome? Does Leader always choose same zone?

Do both firms agree on who the “Leader” should be?? If so, how?