You are on page 1of 20

Introduction to Financial Crises

Lecture 3: Global Games

University of Vienna – Masters in Banking and Finance

Winter Term 2022


Background

Coordination Games and Equilibrium Multiplicity: Games in which players’ actions


exhibit strategic complementarities/feedback effects are prone to multiple equilibria.

Some Common Examples drawn from Finance:


• Bank Runs – e.g., Diamond & Dybvig (1983)
• Balance-of-Payment (Currency) Crises – e.g., Obstfeld (1986)
• Sovereign Debt Crises – e.g., Calvo (1988)

Models exhibiting multiple equilibria have some appeal insofar as they can be
interpreted as reflecting the fragility of financial markets and institutions.

1 / 19
Global Games
Reference: Morris and Shin (2001)

Equilibrium multiplicity does, however, pose some serious shortcomings:


• Equilibrium switches are driven by shifts in beliefs (which remain unexplained).
• There is no correlation between equilibrium outcomes and fundamentals.
• Difficult to draw policy implications as their effects vary across equilibria.

Multiplicity arises because agents’ beliefs give rise to actions that lead to outcomes
that vindicate agents’ initially held beliefs.

More specifically, equilibrium multiplicity follows from the assumptions that:


• economic fundamentals are common knowledge: “I know that You know θ, You
know that I know that You know θ, etc.”
• there is no strategic uncertainty (i.e. agents are certain about the behaviour of
other agents in equilibrium).

Global Games (Main Idea): Break players’ ability to perfectly coordinate their actions
and beliefs by adding idiosyncratic incomplete information (i.e., noise) to the game.

2 / 19
This Lecture
Reference: Koenig (Class Notes)

1 A Simple Example of a Game of Regime Change

2 The Global Game Solution

3 Uniqueness of Global Game Solution

3 / 19
Game of Regime Change

The game is characterized by two possible regimes: status quo and alternative.

A continuum of players, indexed by i ∈ [0, 1] must choose between two actions. One
action is favourable to the status quo while the other is favourable to the alternative.

We will call these actions “not attack” and “attack,” respectively, and denote the
action chosen by player i by ai ∈ {0, 1}.

The regime outcome is denoted by R ∈ {0, 1}, where R = 0 represents the survival of
the status quo and R = 1 represents a change to the alternative regime.

The status quo is abandoned if and only if sufficiently many players attack:
Z 1
R=1 ⇔ A≡ ai di ≥ θ
0

where θ is an economic fundamental that measures the “resilience” of the status quo.

4 / 19
Payoff Matrix

Regime
Status Quo (A < θ) Regime Change (A ≥ θ)
Not Attack (ai = 0) 0 0
Player i
Attack (ai = 1) −C B

Depending on the value of θ, this game may exhibit multiple equilibria:


• If θ ≥ 1, there exists a unique equilibrium where a∗ = 0 ∀i
i
• If θ ≤ 0, there exists a unique equilibrium where a∗ = 1 ∀i
i
• If θ ∈ (0, 1), there exists two pure strategy Nash equilibria

multiple equilibria

θ
0 1

5 / 19
Global Game of Regime Change

The multiplicity of equilibria for θ ∈ (0, 1) is sustained by players’ self-fulfilling


expectations about other players’ actions in equilibrium.

The common knowledge assumption allows players to perfectly coordinate their beliefs
and actions in equilibrium.

The game therefore does not give any indication as to which outcome will obtain since
players’ beliefs are effectively undetermined.

The global game refinement shows that this indeterminacy result is not robust to
(arbitrarily) mild relaxations of the common knowledge assumption.

6 / 19
Global Game – Information Structure

To break the common knowledge assumption, assume that players do not perfectly
observe the state θ but only have a common prior about θ.

Assume here that the prior to be normally distributed such that θ̃ ∼ N (µθ , α−1 ).

Each player i only observes a noisy signal of θ:

xi = θ + ϵ i

where ϵ̃i ∼ N (0, β −1 ) and ϵ̃i and θ̃ are independent.

A strategy for player i can be defined as a mapping from signals to actions:

si : xi → ai ∈ {0, 1}

7 / 19
Threshold Strategies

For the moment, we will restrict attention to threshold strategies (we will show below
that this restriction is without loss of generality).

A threshold strategy is defined as follows:


(
1 if xi ≤ x̂i
si : ai =
0 if xi > x̂i

where the strategy is symmetric if x̂i = x̂ for all i.

To derive the global game solution, we proceed as follows


1 Assume that players adopt a symmetric threshold strategy around x̂.
2 Based on this assumption, show that there exists a unique threshold equilibrium.
3 Finally, show that there do not exist other equilibria in non-threshold strategies.

8 / 19
Step 1 – The Probability of an Attack

Notice that:
x̃i |θ ∼ N (θ, β −1 )
Given the threshold strategy x̂, the size of an attack given some realized value of θ
equals the mass of players receiving a signal below the threshold; that is:
p
A(θ, x̂) = Pr(xi < x̂|θ) = Φ( β(x̂ − θ))

where Φ(·) denotes the distribution function of the standard normal distribution.

What is the critical value of θ such that the status quo fails to survive?

Since A(θ, x̂) is decreasing and continuous in θ, there exists a unique value θ̂ such
that regime change obtains if and only if θ < θ̂. This value solves:

θ̂ = A(θ̂, x̂) (1)

9 / 19
Step 2 – Posterior Beliefs

How likely does a player i observing signal xi expect the status quo to break down?

By Bayes’ rule, the posterior distribution of θ given xi is:

αµθ + βxi 1
 
θ̃|xi ∼ N ,
α+β α+β

Hence, the posterior belief of regime change conditional on xi is:

αµθ + βxi
p  
Pr(θ ≤ θ̂(x̂)|xi ) = Φ α + β θ̂(x̂) −
α+β

which is monotonically decreasing in xi .

It follows that player i only finds it optimal to attack if xi < x̂, where x̂ solves the
following indifference condition:

Pr(θ ≤ θ̂(x̂)|x̂)B + (1 − Pr(θ ≤ θ̂(x̂)|x̂))(−C) = 0 (2)

10 / 19
Step 3 – Threshold Equilibrium

Equations (1) and (2) jointly characterize the threshold equilibrium {x̂∗ , θ̂∗ }.

Solving (1) for x̂ we obtain:

1
x̂(θ̂) = θ̂ + p Φ−1 (θ̂)
β

Substituting this equation into (2) and simplifying yields:


r
α α + β −1 C
 
F (θ̂; α, β) ≡ p (θ̂ − µθ ) − Φ−1 (θ̂) − Φ =0
β β B+C

In order for the equilibrium to be unique, F (·) must be monotonically decreasing in θ̂.

Proposition p √
There exists a unique threshold equilibrium {x̂∗ , θ̂∗ } if and only if α/ β≤ 2π.

11 / 19
Public versus Private Information

p
The magnitude α/ β reflects the relative precision of public versus private
information.

The previous result tells us that if agents’ private info is sufficiently precise (high β)
relative to agents’ public info (low α), the global game selects a unique equilibrium.

A convenient way to characterize the global game solution is to focus on the limit case
where players’ private signals become arbitrarily precise (i.e. β → ∞).

Proposition
B
As β → ∞, the threshold equilibrium x̂∗ → θ̂∞
∗ where θ̂ ∗ =
∞ B+C
∈ (0, 1).

∗ , no player attacks and the status quo survives.


For realizations of θ > θ̂∞
∗ , all players attack and the regime changes.
For realizations of θ < θ̂∞

12 / 19
Generic Uniqueness Result

So far, we have restricted attention to equilibria in threshold strategies, and shown


that there exists only one.

But what about equilibria in non-threshold strategies?

For example, an equilibrium of the form {x1 , x2 } with x1 < x2 where agents attack if
xi < x1 , refrain from attacking if xi ∈ (x1 , x2 ) and attack again if xi > x2 .

The global game framework can be used to show that the threshold equilibrium is
indeed the unique equilibrium (i.e. there are no equilibria in non-threshold strategies).

The proof follows from the iterated elimination of strictly dominant strategies.

13 / 19
Dominance Regions

The proof relies on the fact that for very small (high) signals, attacking (refraining)
becomes a dominant strategy; these are often referred to as dominance regions.

For example, consider the case where some agent k believes that no other agent
attacks the regime; will agent k ever want to attack in this case?

Answer: Yes, if the private signal xk is sufficiently small; i.e. for all xk < x agent k
attacks the regime regardless of the actions of the other agents.

Similarly, consider the case where some agent k believes all the other agents attack
the regime; will agent k ever want to refrain from attacking?

Answer: Yes (again), this time if the private signal xk is sufficiently large; i.e. for all
xk > x agent k refrains from attacking regardless of the actions of the other agents.

14 / 19
Infection Argument

Recall that due to strategic complementarities in actions, agent k’s incentive to attack
the regime is increasing in the fraction of other agents attacking.

Suppose agent k observes a signal x1 slightly below x (the upper dominance region).

In this case, agent k must believe that the fundamentals (θ) are strong and that for
the regime to change many other agents must attack.

But given that x1 is high, agent k must believe that most other agents also observe
signals close to (or above) x.

Hence, agent k must believe that other agents are unlikely to attack and consequently
will refrain from attacking as well.

A symmetric argument holds for signals slightly above x (the lower dominance region).

Any equilibrium must therefore be in threshold strategies.

15 / 19
Discussion: Laplacian Beliefs
Reference: Morris and Shin (2001)

Global games resolve the equilibrium multiplicity problem in coordination games by


introducing strategic uncertainty.

By adding noise, agents’ become uncertain about other agents’ payoffs, and hence the
actions they will choose in equilibrium.

While the global game framework may seem to suggest that agents’ posses
unreasonably sophisticated reasoning capabilities, this is not in fact the case.

In fact, it can be shown that agents’ actions can be rationalized by beliefs which
resemble Laplace’s principal of insufficient reason.

In other words, an agent’s actions are based on a uniform belief about the behaviour
of other agents, implying a simple heuristic underlying agents’ choice of actions.

16 / 19
Discussion: Laplacian Beliefs (cont.)
To see this formally, assume α → 0; in this case, economic fundamentals are drawn
from an (improper) uniform density over the real line: θ ∼ U [R].

Now consider the question: "my signal is xk ; what is the likelihood that less than a
fraction z ∈ [0, 1] of the other agents have received a signal higher than mine?"

Since xi |θ ∼ N (θ, β −1 ), the fraction of agents receiving a signal higher than xk is


p
Pr(xi > xk |θ) = 1 − Φ( β(xk − θ))

This is less than some fraction z if


p 1
1 − Φ( β(xk − θ)) ≤ z ⇔ θ ≤ θ∗ ≡ xk − p Φ−1 (1 − z)
β

Since θ|xk ∼ N (xk , β −1 ), the probability of this event is given by


!!

p 1 −1
Pr(θ < θ |xk ) = Φ β xk − p Φ (1 − z) − xk =z
β

17 / 19
Applying Global Games

We have seen that global games can be used as a selection device to overcome the
equilibrium multiplicity problem that plagues coordination games.

Global games techniques, however, can only be used if the underlying game satisfies a
number of specific properties.

While these may sometimes be relaxed, the following properties must generally hold:
1 Action Monotonicity: Payoffs must reflect strategic complementarities in actions.
2 State Monotonicity: Payoffs must be increasing in economic fundamentals.
3 Continuity: Payoffs must be “weakly” continuous in actions and states.
4 Limit Dominance: There exists well-defined upper- and lower-dominance regions.

18 / 19
Summary

Global games permit a more rigorous analysis of financial fragility.

By resolving the equilibrium multiplicity problem afflicting coordination games, they


allow to perform well-defined comparative static and policy analysis.

They have been applied to numerous settings, including:


• Bank Runs: Rochet & Vives (2004), Goldstein & Pauzner (2005)
• Currency Attacks and Twin Crises: Morris & Shin (1998), Goldstein (2006)
• Debt and Credit Crises: Corsetti et. al. (2006), Goldstein & Bebchuk (2011)

Next week we will discuss the Rochet-Vives model of bank runs, and use it to study
the design and implementation of lender of last resort (LoLR) policies.

19 / 19

You might also like