Lecture Notes

International Financial Crises

Bernardo Guimaraes

February 2024

Abstract

These lecture notes cover currency crises, banking crises and debt crises, and include quite a bit
on expectations and coordination. Its main focus is on theoretical models, but the last two sections
are on empirical methods that can be used to shed light on these (and other) issues. Comments,
suggestions, and corrections of any mistakes or typos are more than welcome.

Contents
1 Currency crises: basic models 4
1.1 The …rst generation models of currency crises . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.1 Flood and Garber (1984) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The second generation models of currency crises . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Jeanne (1997) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.2 A game theoretical representation: Obstfeld (1996) . . . . . . . . . . . . . . . . . . 12

2 Expectations and coordination 15

2.1 Expectations and higher order beliefs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.1 The common knowledge assumption . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.2 Beauty contests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.3 The model with incomplete information . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.4 Unique equilibrium: an intuition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.5 Morris and Shin (1998) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.6 The e¤ect of public information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.7 Discussion about contagion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Rational herd behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2.1 Bikhchandani, Hirshleifer and Welch (1992) . . . . . . . . . . . . . . . . . . . . . . 27
Sao Paulo School of Economics - FGV, [email protected], http://sites.google.com/site/bernaguima.

1
3 Banking crises 30
3.1 Bank runs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.1 Diamond and Dybvig (1983) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Banking crises and moral hazard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.1 The moral hazard e¤ect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.2 The Asian crisis of 1997 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4 Sovereign debt and default 37

4.1 A model with exogenous interest rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Sovereign debt sustainability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2.1 Collard, Habib and Rochet (JEEA 2015) . parei
. . . aqui
. . . . . . . . . . . . . . . . . . . . 41
4.2.2 The maximum sustainable level of debt . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Sovereign debt with no commitment to repay . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.1 Endogenous default decision (Eaton and Gersovitz, 1981) . . . . . . . . . . . . . . 43
4.3.2 A deterministic example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4 Recent models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4.1 Arellano (2008) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4.2 Cuadra, Sanchez and Sapriza (2010) . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.4.3 Which shocks matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.4.4 IMF and …scal adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5 Debt crises and coordination 62

5.1 A simple model of liquidity crisis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.1.1 IMF intervention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2 Liquidity crises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.1 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2.2 A lender of last resort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Coordination and the price of debt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.1 Morris and Shin (2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6 Bubbles 69
6.1 Speculative bubbles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2 Riding bubbles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

7 Expectations and option prices 75

7.1 Expectations implicit in …nancial prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.1.1 Rose and Svensson (1994) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.2 Option prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.2.1 The Black and Scholes model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.2.2 Campa, Chang and Refalo (2002) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.2.3 Bates (1991) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7.2.4 Guimaraes (2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

2
 7.2.5 Parametric non-parametric method . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.2.6 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

8 Identi…cation through heteroskedasticity 85

8.1 Evidence on default costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.2 US interest rates and default risk in emerging economies . . . . . . . . . . . . . . . . . . . 88
8.3 The e¤ects of commodity price shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3
1 Currency crises: basic models

1.1 The …rst generation models of currency crises

estoque esgotável
Salant and Henderson (1978) showed that if the government uses a stockpile of an ex-
haustible resource (e.g., gold) to stabilize its price, eventually a speculative attack will
occur: the private investors will suddenly acquire the entire government’s stock.
Paul Krugman (1979) showed that if a pegged exchange rate cohexists with budget
de…cits that need to be …nanced by money creation, the argument in Salant and Henderson
(1978) also applies: a speculative attack will force the government to abandon the pegged
regime.
Flood and Garber (1984) develops the concept of “shadow exchange rate”and provide
two linear examples of the logic presented by Krugman (1979). This notes covers the
deterministic model of that paper. The stochastic model of Flood and Garber (1984) is
also interesting, but not covered here.

1.1.1 Flood and Garber (1984)

The exchange rate is denoted by S. In the non-stochastic version of Flood and Garber
(1984), the exchange rate is initially pegged at S. Money demand depends negatively on
interest rates:
Mt
= a0 a1 i t (1)
Pt
Money supply equals foreign currency reserves (Rt ) plus domestic credit (Dt ).

Mt = Rt + Dt (2)

Domestic credit is expanding:

D_ t
= ; >0 (3)
Dt
The model assumes there will be no arbitrage opportunities in markets for goods and
assets. That implies interest rate parity (IRP) and purchasing power parity (PPP) (it
and Pt are constants).
Pt = St :Pt (4)
S_
it = it + (5)
S
Initially, the government has a positive stock of reserves and will keep the peg until
reserves reach a given minimum level (say, until Rt = 0). Before the peg is abandoned,
S_ = 0. By PPP (equation 4), P_ = 0, and by IRP (equation 5), it is constant, equal it .

4
Therefore Mt is also constant (equation 1). De…ne M H as the demand for money while
the peg is kept:
MH
= (a0 a1 it ) (6)
S:Pt
In the model, the expansion of domestic credit generates loss of reserves until the
moment in which the peg is abandoned. Then, it leads to an increasing trend in the
money supply and, consequently, in‡ation. Therefore, after the peg is abandoned, the
demand for real balances is smaller because the nominal interest rate is higher, due to
in‡ation (equations 5 and 1). An arbitrage condition implies that Pt and St cannot jump
up, and so the discrete reduction in money demand translates in a discrete fall of Mt .
Initially, reserves are falling steadily, at a rate . When Rt is exactly equal to the di¤erence
in money demand in both regimes, all agents exchange part of their domestic currency
for foreign currency and the government is forced to abandon the peg. De…ne M L as the
demand for money right after the peg is attacked:
!!
ML S_
mL = = a0 a1 it + (7)
S:Pt S

Now, de…ne the shadow exchange rate (S~t ) as the exchange rate that would prevail if
the currency was allowed to ‡oat (demand for real balances would be mL ) and foreign
reserves vanished (so that Mt = Dt ). PPP implies that Pt = Pt :S~t and we have:
Mt Dt
mL = = (8)
Pt S~t Pt
Equations 7 and 8 imply:
S
S~t = L Dt (9)
M
As Flood and Garber (1984) show, a speculative attack forces the abandonment of the
peg exactly when S~t = S.

The size of the attack In Flood and Garber (1984), a speculative attack is an instan-
taneous event: agents exchange some of their local currency for foreign currency (M falls)
and deplete the Central Bank stock of reserves (R falls to 0). What is the lost in reserves?

Right before the attack, we have:

M H = S:Pt : (a0 a1 i t ) (10)

Right after the attack, we have:

5
M/P M
low inflation
M=D
high inflation

t t

As the demand for

real balances falls,
if P is not to jump,
M has to fall by a
discrete amount.

P=S

Figure 1: Flood and Garber (1984)

6
 !!
L S_
M = S:Pt : a0 a1 it + (11)
S
Subtracting (11) from (10), we get:

S_
M = MH M L = S:Pt :a1 :
S
As Mt = Rt + Dt , M= R+ D. Domestic credit is growing continuously. There-
fore, Dt is the same right before and right after the devaluation ( D = 0). So, M= R:

S_ P_
R = S:Pt :a1 :
= S:Pt :a1 :
S P
Interpreting the above equations: the fall in reserves corresponds to the fall in the
demand for money. The fall in the demand for money is due to in‡ation post-devaluation.
In‡ation occurs because the Central Bank has run out of reserves (so cannot …nance the
…scal authority by selling reserves anymore) and thus starts to …nance the …scal authority
via in‡ation.

Some take home points

Inconsistency between domestic policy and exchange rate policy leads to speculative
attacks. Increases in D lead either to decreases in R (reserves dwindle) or to increases
in M (monetary expansion, that leads to in‡ation). Loss of reserves can’t go forever
(stock of reserves available to Central Banks is …nite). At some point, increases in
D lead to increases in M .

A simple demand for money relation, arbitrage in all markets (PPP, IRP) and the
increase in domestic credit lead to agents massively sell domestic currency and force
the abandonment of the peg.

A massive speculative attack is not incompatible with rational agents.

What to do about speculative attacks? The model seems to say: “don’t shoot the
messenger!”

The model predicts that crises are predictable, antecipated.

In‡ation follows the currency crises.

However...
European Exchange Rate Mechanism, 1993: the bark in the dark that was not heard
(Obstfeld, 1996).

7
1.2 The second generation models of currency crises

The weak links between changes in economic variables and speculative attacks in some re-
cent episodes (e.g., the ERM crises in 1992-3 and the contagion of 1997-8) have stimulated
the idea that bad fundamentals may be a pre-condition for a crisis, but its occurrence and
timing are somewhat random events. The so called second generation models of currency
crisis formalize this view. This literature points out that if fundamentals are not good
enough, the optimal strategy for an agent in a currency crisis game depends on expecta-
tions: if everybody is expected to attack the currency, it is optimal to attack it, but if
everybody is expected to refrain from doing so, then not attacking is the optimal choice.
Those models present multiple equilibria. Sudden and exogenous shifts on expectations
may trigger a crisis.

R = R* + Exp(∆S/S)

M/P = L(R,Y) LR<0, LY>0

M
P

Figure 2: Change in expectations

A second generation model of currency needs:

1. A reason why the government want to abandon its …xed exchange rate regime,

2. A reason why the government want to keep its …xed exchange rate regime,

8
3. Cost of defending the …xed exchange rate regime must be increasing in expectations
of devaluation.

B,C
Cost of keeping the fixed
exchange rate regime

Benefit of keeping the fixed

exchange rate regime

Expectations of devaluation

Figure 3: Government’s decision on the exchange rate regime

What are the bene…ts of keeping the …xed exchange rate regime?

Removing volatility of the exchange rate regime is good for trade, investment.

Nominal anchor - in‡ation.

Reputation.

What are the costs of keeping the …xed exchange rate regime? (See Obstfeld, 1996)

Increases in the interest rate may slow down economy, increase unemployment.

Distribution e¤ects: hikes in the interest rate make mortgages more expensive, bond
holders wealthier, indebted companies poorer.

Banks may su¤er when interest rates increases (we will discuss this issue further in
a couple of weeks).

If a country is highly indebted, its …scal burden increases if expectations of a deval-

uation push up interest rates.

Government may want to in‡ate away its debt.

Why would the costs of keeping the …xed exchange rate regime be increasing:

The higher is the expected devaluation, the higher is the hike in the interest rates.

9
 Seeing from a di¤erent perspective, if one expects a currency to depreciate, he/she
will sell it (or short it). The pressure for devaluation is proportional to this amount
sold (or short) as the government will have to buy it or to increase incentives (interest
rates) for others to hold it.

1.2.1 Jeanne (1997)

Jeanne (1997) develops a simple second generation model of currency crises. He uses
the model to execute a likelihood ratio test for the existence of multiple equilibria in the
French Franc crisis.

The currency is pegged at a …xed rate with a foreign currency.

Policymaker can defend the peg (possibly at some cost) or abandon it.

The bene…t of the pegged exchange rate The loss function of the policy maker is

Lt = u2t + t Ct

where ut is the unemployment rate, t = 1 if the peg is abandoned and 0 otherwise, and
Ct is an exogenous reputational cost.
There is also a (kind-of) Philipps curve

ut = ut 1 a (et Et 1 (et ))

where et is the exchange rate and a; > 0.

The exchange rate is …xed at e (exogenous). If there is a devaluation, it will go to
e+ e (exogenous). Let’s say the probability of a devaluation is t 1. So

E(et 1 ) = e + t 1 e

and
ut = ut 1 a (et e t 1 e)

If there is a devaluation,

udt = ut 1 a e(1 t 1)

If the currency peg is kept,

uft = ut 1 +a e t 1

= udt + a e

10
 The net bene…t of keeping the peg is given by

Bt = Ldt Lft
2 2
= [ ut 1 a e(1 t 1 )] + Ct [ ut 1 +a e t 1]

= Ct 2 ut 1 a e + (a e)2 + 2 (a e)2 t 1

Now de…ne

b t = Ct 2 ut 1 a e + (a e)2
= 2 (a e)2

and we can write that the net bene…t of mantaining the peg is

Bt = bt : t 1

so bt depends on economic fundamentals but Bt also depends on expectations.

The decision of the policy maker

The policymaker may be in:

–“soft”mood, with probability : mantains the peg if the net bene…t of doing so
is positive.
–“tough” mood, with probability 1 : maintains the peg whatever is the cir-
cunstance.

He needs this changes in moods to …t better the data, the model would go through
without it.

We assume that:
bt = t 1 + t

where t 1 = Et bt+1 and t is an error term.

Equilibria

The devaluation probability must be equal to the probability that the government is
soft and the net bene…t of mantaining the peg is negative, that is:

t = : Pr (Bt+1 < 0)
= : Pr ( t+1 < : t t)

= :F ( : t t)

11
 So t appears at both sides of the equation. There may be multiple equilibria or not,
depending on f (0).
There are multiple equilibria if:

1. the parameters satisfy f (0) > 1;

2. the fundamental lies in the multiple-equilibrium interval;

Pictures in the paper show that if F () reacts strongly enough to t, there is a region
with multiple equilibria.

1.2.2 A game theoretical representation: Obstfeld (1996)

As government’s decision depends on how many agents attack the currency, self-ful…lling
crises may occur. Everybody expects that the peg will be abandoned, so everybody attacks
the currency. And the peg is abandoned because everybody attacked the currency. This
kind of circular logic is characteristic of the second generation models of currency crises.
A simple example from Obstfeld (EER 1996) helps to clarify this point: a government
that wants to …x its currency and two private holders of domestic currency who can sell
it (attack the currency) or hold it (not attack). The government has R reserves to defend
the peg. Each trader has domestic money resources of 6 which can be sold for reserves.
To sell and take a position against the government, there is a cost of 1 (assumed to be
irrespective of the amout sold, but that is not important for the results). In the event of
giving up its peg, the government devalues by 50 percent (so, the traders get 1/2 unit of
money for each unit they bought in the event of a successful attack).
The “high-reserve”game: R = 20
Trader 2
A N
Trader 1 A 1, 1 1; 0
N 0, 1 0; 0

The “low-reserve”game: R = 6
Trader 2
A N
Trader 1 A 1=2, 1=2 2; 0
N 0, 2 0; 0

The “intermediate-reserve”game: R = 10

12
 Trader 2
A N
Trader 1 A 3=2, 3=2 1; 0
N 0, 1 0; 0

Take home points

There are self-ful…lling crises in a model with very rational investors.

Sunspots, events completely disconnected from the economy, may change expecta-
tions and trigger a currency crisis.

Fixed exchange rate regimes that would work well in the absence of the speculative
attack may fall without major fundamental imbalances.

There are strategic complementarities between agents’actions: incentives for an agent

to attack the currency are increasing in the share of agents that choose to attack the
currency.

Crises are not fully predictable.

The game with multiple equilibria works not only for crises, but for many other
things. For example, your willingness to go to a party may be an increasing function
of the others’decisions. So there might be 2 equilibria (either many people go to the
party or noone goes).

Here there are complementarities because the actions of the government depend
on what everyone does, but strategic completarities could arise for other reasons.
Example: suppose you work for an asset management company. If you lost money
in Asia and all your peers did the same, or if you all pro…t from investing there, this
is business as usual for you. If you are the only one that lost money there, your job
may be in danger. On the other hand, if you are the only one that got it right, you
may get an extra bonus. However, you may prefer the former safe choice to the latter
risky lottery, so it may be rational for you to try to do what others are doing.

However. . .

Assumption that agents know what others are doing in equilibrium is very strong (is
it?)

13
Expectations are given exogenously in the model. How would you form your own
expectations about what other players would do? Compare it with your own decision
in the game played in class.

A side point: The e¤ect of an increase in foreign interest rates (R ) can also be seen
at …gure 2 If R increases, the interest-rate-parity curve shifts up, which forces the
domestic economy to increase interest rates (if a devaluation is to be avoided). In
1994, the US Federal Reserve Bank sharply increased interest rates. Interest rates in
Mexico had to follow. Clearly, that increases the cost of keeping the …xed exchange
rate regime for the Mexican government. The currency crises occurred in Mexico in
December 1994.

14
2 Expectations and coordination

2.1 Expectations and higher order beliefs

2.1.1 The common knowledge assumption

If an event (say > 0) is common knowledge, then everybody knows it, everybody knows
that everybody knows it, everybody knows that everybody knows that everybody knows
it, everybody knows that everybody knows that everybody knows that everybody knows
it, and so on.
It is not easy to see why that would be di¤erent from the simple “everybody knows
it”. The example in Geanakoplos (1992, page 54) illustrates the strenght of the common
knowledge hypothesis.

2.1.2 Beauty contests

Consider the following game:

Player 2
A N
Player 1 A , 1; 0
N 0, 1 0; 0

If 2 (0; 1), the game has 2 Nash Equilibria: (A; A) and (N; N ). What would you
choose?
The optimal choice depends on the probability attached to the other agent choosing
A. Let’s denote this probability by p, that is, p = Pr(s2 = A). Then, my payo¤ from
choosing A is:

(s1 = A) = p: + (1 p):( 1)
= p+ 1

So, A is the optimal choice if:

p>1

A natural question is: what does p depend on? What is the probability that the other
player will choose A? Naturally, that depends on the probability player 2 assigns for
player 1 choosing A...
Keynes, in the “General Theory of Employment, Interest and Money” (1936) wrote
that: “...professional investment may be likened to those newspaper competitions in which
the competitors have to pick out the six prettiest faces from a hundred photographs, the

15
prize being awarded to the competitor whose choice most nearly corresponds to the average
preferences of the competitors as a whole; so that each competitor has to pick, not those
faces which he himself …nds prettiest, but those which he thinks likeliest to catch the
fancy of the other competitors, all of whom are looking at the problem from the same
point of view. It is not a case of choosing those which, to the best of one’s judgement, are
really the prettiest, nor even those which average opinion genuinely thinks the prettiest.
We have reached the third degree where we devote our intelligences to anticipating what
average opinion expects the average opinion to be. And there are some, I believe, who
practise the fourth, …fth and higher degrees.”
In a Nash equilibrium, an agent knows which action the other players will be choosing.
So, the above mentioned reasoning is not incorporated in the second generation models
of currency crises.

2.1.3 The model with incomplete information

Carlsson and Van Damme (Econometrica, 1993) show that multiplicity of equilibrium
comes from two modelling assumptions:

In equilibrium, agents know what others will do;

All information is common knowledge.

What happens when we remove those assumptions and agents are uncertain about
what others will do? An agent has to estimate the likelihood of other players attacking
the currency. So, they try to assess the others’information — what do they know? what
are they expecting? Like in the beauty contest example, an agent is guessing what the
others guess that she knows.
Consider that agents are playing the same game they played before:
Player 2
A N
Player 1 A , 1; 0
N 0, 1 0; 0

However, is not observed. The only information agents have about is a noisy signal
x. Formally, the prior on is uniformly distributed in the real line (or think of a normal
distribution with ! 1) For each agent i:

xi = + i

16
 The error term, i, is independent accross agents. The striking result …rst proved by
Carlsson and Van Damme is that the above game has a unique equilibrium even if the
variance of the error term is very small — for example, even if i is distributed uniformly
between -0.01 and 0.01, that is: i U ( 0:01; 0:01).
The key here is that it is never common knowledge that 2 [0; 1]. For example,
suppose agent 1 gets a signal x1 = 0:40. Then:

Agent 1 knows that 2 [0:39; 0:41],

so agent 1 knows that x2 2 [0:38; 0:42],

so agent 1 knows that agent 2 knows that 2 [0:37; 0:43],

so agent 1 knows that agent 2 knows that x1 2 [0:36; 0:44],

so agent 1 knows that agent 2 knows that agent 1 knows that 2 [0:35; 0:45],

so agent 1 knows that agent 2 knows that agent 1 knows that x2 2 [0:34; 0:46],

so agent 1 knows that agent 2 knows that agent 1 knows that agent 2 knows that
2 [0:33; 0:47], and so on.

An agent that gets a signal x1 = 0:40 thinks that perhaps = 0:39, so x2 might be
0:38, in which case agent 2 would think it is possible that = 0:37 and x1 = 0:36. So
agent 1 considers it possible that agent 2 thinks that x1 is as low as 0.36 or as high as
0.44. Likewise, if x1 = 0:36, agent 1 would consider it possible that agent 2 thinks that x2
is as low as 0.32. Combining both statements means that agent 1 with x1 = 0:40 thinks
that agent 2 might think that agent 1 might think that agent 2 might think that x1 is as
low as 0.32.
OK, but why does it matter?

2.1.4 Unique equilibrium: an intuition

Suppose that i N (0; ).

From the point of view of agent 1, that got signal x1

N (x1 ; ).
p
x2 N (x1 ; 2 ). (Why? From the point of view of agent 1, N (x1 ; ). and
x2 N ( ; ), so...)

The expected payo¤ from choosing A for an agent that got signal x1 is:

E( (s1 = A)) = p + x1 1

17
Dominant regions

If x < 0, E( (s1 = A)) < p 1 < 0. Regardless of what player 2 is choosing, player
1 is better o¤ by choosing N .

If x > 1, E( (s1 = A)) > p > 0. Regardless of what player 2 is choosing, player 1 is
better o¤ by choosing A.

Now, what if 0 < x < 1? Suppose for example that = 0:1.

If x1 = 0:1, what is the probability that agent 2 will play A?
x2 0:1 0:1
Pr(x2 < 0) = Pr p < p
0:1 2 0:1 2
1
= p = 0:24
2
where is the standard normal cumulative distribution function. So, with probability
0.24, agent 2 got a signal smaller than 0 and will play N . If agent 2’s signal is positive,
we don’t know what he will do, but we can say that:

p < 0:76

So, agent 1’s expected payo¤ of attacking is:

E( (s1 = A)) < 0:76 + 0:1 1= 0:14

so, agent 1 will not attack.

Using a similar argument, if x1 = 0:9, then:
x2 0:9 1 0:9
Pr(x2 > 1) = Pr p > p
0:1 2 0:1 2
1
= 1 p = 0:24
2
So, with probability 0.24, agent 2 got a signal bigger than 1 and will play A. If agent
2’s signal is smaller than 1, we don’t know what he will do, but we can say that:

p > 0:24

So, agent 1’s expected payo¤ of attacking is:

E( (s1 = A)) > 0:26 + 0:9 1 = 0:14

so, agent 1 will attack.

18
 Now, suppose that x1 = 0:2. Then the probability that agent 1 gets a signal smaller
than 0 is just 0.08. If all we can say is that p < 0:92, then we cannot say anything about
agent 1’s optimal decision in this case because

E( (s1 = A)) = p + x1 1)
E( (s1 = A)) < 0:92 + 0:2 1 = 0:12

that is, from this calculation, we can’t tell whether E( (s1 = A)) is positive or negative.
How can agent 1 decide in this case?
Agent 2 is also rational and is doing the same calculations agent 1 is doing. Therefore,
agent 1 knows that if x2 = 0:10, agent 2 will not attack. Why is that? If agent 2 got
signal x2 = 0:10, he will consider that the probability that agent 1 got a negative signal
is 0.24 and will not attack. Agent 1 knows that. And given that his signal x1 = 0:20, he
knows that agent 2 will have got a signal x2 0:10 with probability 0.24 because:
x2 0:2 0:1 0:2
Pr(x2 < 0:1) = Pr p < p
0:1 2 0:1 2
So, he knows that p < 0:76 and therefore, agent 1’s expected payo¤ of attacking is:

E( (s1 = A)) < 0:76 + 0:2 1= 0:04

E( (s1 = A)) is negative! If x1 = 0:2, agent 1 will choose N .

Now, suppose that we know that players 1 and 2 choose N if x xL , for some xL .
Suppose x1 = xL + , where is a positive and very small constant, << .
Then:
x2 x1 xL x1
Pr(x2 xL ) = Pr p p
2 2
= Pr z p 0:5
2
So,

p = Pr(s2 = A) 1 Pr(x2 xL )
0:5

Then,

E( (s1 = A)) = p + x1 1
x1 0:5

19
 Therefore, we can iteratively delete s = A whenever x1 < 0:5.

At the other side, suppose that we know that players 1 and 2 choose A if x xH , for
some xH .
Suppose x1 = xH , where is a positive and very small constant, << .
Then:
x2 x1 xH x1
Pr(x2 xH ) = Pr p p
2 2
= Pr z p 0:5
2
So,

p = Pr(s2 = A) Pr(x2 xH )
0:5

Then,

E( (s1 = A)) = p + x1 1
x1 0:5

Therefore, we can iteratively delete s = N whenever x1 > 0:5.

So, the unique equilibrium that survives strategically elimination of strictly dominated
strategies is:

si = A if xi > 0:5

si = N if xi < 0:5

The argument works in the same way even if the support of " is bounded, for example,
even if i U ( 0:01; 0:01). Note the higher order beliefs in action: if you get a low signal
(say x = 0:05) you will end up playing N even though you know that is positive and
you know that the other player knows that is positive. That is because he may think
that you may think that he may think that ........ that is negative. Although agents
in this case know that > 0, so the (A; A) equilibrium would yield positive payo¤s for
them, that is not common knowledge.

2.1.5 Morris and Shin (1998)

Morris and Shin (1998), building on Carlsson and Van Damme (Econometrica, 1993),
endogenize expectations in a model of currency attacks. Here I present a simpli…ed version.

20
 Consider an economy with a continuum of players and denote by l the proportion of
players that choose to attack. The government has a constant bene…t for holding the peg
and a cost that depends negatively on fundamentals ( ) and positively on the proportion
of agents that choose to attack (l). We will consider that the cost will exceed the bene…t
if l > 0. So, the government abandons the peg if l > .
The information structure is as before (xi = + i ). An agent chooses between ‘attack’
(A) and ‘not attack’(N ). If the agent chooses N , she does not win or lose anything, her
payo¤ is 0 regardless of what others do. The cost of attacking is t. If she attacks and
there is a devaluation, she gets 1. So, if she chooses A, her payo¤ if 1 t is there is a
devaluation and t otherwise. Suppose that i N (0; ).
Here, if is negative, the government abandons the peg regarless of what agents do
and if > 1, the peg if kept even if all agents decide to attack. The agent will attack the
currency if she perceives that fundamentals are weak, that is, only if she thinks that is
small.

The model with common knowledge of fundamentals Suppose for a while that all agents
observe . Then, the model is a second generation model.
Suppose everyone attacks the currency (l = 1). Then the government abandons the
peg if < 1. In this case, it is optimal for an agent to attack the currency if < 1.
Suppose that noone attacks the currency (l = 0). The the government abandons the
peg if < 0. In this case, it is optimal for an agent not to attack the currency if > 0.
Thus, if < 0, fundamentals are too weak, everybody attacks the currency and the
government leave the peg. If > 1, nobody attacks and the peg survives. But if 0 < < 1,
both equilibria exist.

The model with imperfect information about Now, let’s return to the case when is
not observed and agents have just imperfect information about it.
Morris and Shin (1998) show that there is a unique equilibrium, characterized by two
thresholds:

1. an agent attacks only if her signal xi is smaller than x .

2. the government abandons the exchange rate peg only if < .

2 conditions pin down the equilibrium:

1. An agent who gets signal x is indi¤erent between attacking or not attacking.

21
 2. When = , the fraction of agents that attack the currency is just enough to make
the government abandon the peg.

An agent that gets signal x asks himself: what is the probability that the attack will
succeed?

The question is: what is the probability that < ?

From that agent’s point of view, N (x ; ). Saying di¤erently, =x i.

So, Pr( < ) = Pr(x i < ) = Pr( i < x ) = Pr( i < x ). Thus:
x
Pr( < )= ( )

where is the cumulative standard normal distribution.

So, The expected payo¤ of attaking is:

x x
E(payA) = (1 t): ( ) t(1 ( ))
x
= ( ) t

The agent is indi¤erent when:

x
( ) t=0 (12)

The second equilibrium condition: when = , the fraction of agents that attack the
currency is just enough to make the government abandon the peg. But when = ,
what is the proportion of agents that choose to attack?

As we have many agents, the question is: what is the proportion of agents that get
a signal x such that x < x ?

If = ,x= + i.

So, Pr(x < x ) = Pr( + i < x ) = Pr( i < x ). Thus:

x
Pr(x < x ) = ( )

When = , the cost and the bene…t of government keeping the peg are the same.
So:
x
=l) = ( ) (13)

22
 Remember that:
x x
( )=1 ( )

So, the two equilibrium conditions (equations 12 and 13) yield:

1 t=0

So:
=1 t

A few take home points

If agents are trying to guess what others are trying to do in a currency crises situation,
expectations are crucial for the …nal outcome but are not disconnected from economic
fundamentals.

In this simple model, expectations depend only on prices (t) and fundamentals ( ).
However, more elaborated models using this technique are able to show other in-
teresting features of expectations (for example, the e¤ect of public information in
crises).

2.1.6 The e¤ect of public information

Let’s look at the magnifying e¤ect of public information (from Morris and Shin, 2003).
Consider again the following game:
Player 2
A N
Player 1 A , 1; 0
N 0, 1 0; 0

Suppose that agents do not observe . The information agents have about is a noisy
private signal xi and a public signal y. For each agent i 2 f1; 2g:
2
xi = + "i , " i N (0; )

and y is the same for all agents:

2
y= + , N (0; )

The private signal is the result of your own analysis of the economy. The public signal
is in the …rst page of the Financial Times and you are sure that everybody will read it.

23
 Agent i’s posterior expectation of is normally distributed with mean:
1 1 2
2 y+ 2 xi y + 2 xi
i = E( jxi ; y) = 1 1
= 2+ 2
2 + 2

and variance 2 2
1
1 1
= 2 2
2 + 2 +
Moreover, from the point of view of agent 1, the signal received by agent 2 is:

x2 = 1 + "2

Since E ("2 ) = 0, E1 (x2 ) = 1. The variance of agent 1’s estimator of x2 is:

2 2 2 2 4
2 2 +
V ar( 1 ) + V ar("2 ) = 2 2
+ = 2+ 2
+
2 2
An intuition Let’s consider the case = . For concreteness, suppose = 0:5, x1 = 0:4,
x2 = 0:4 and y = 0:6. Remember that in the case with only private information, agents
would be indi¤erent when xi = 0:5. The question is: what happens in this case?
From the point of view of agent 1:

E( ) = 0:5

E(x2 ) = 0:5

Now, what does agent 1 expect about agent 2’s expected x1 ?

Agent 1 considers that agent 2 has 2 pieces of information: x2 and y.

From the point of view of agent 1, E(x2 ) = 0:5 and E(y) = 0:6.

So, E1 (E2 (x1 )) > 0:5 (it is a (weighted) average between 0.5 and 0.6).

Notice that both agents think that is 0.5 plus an error term. But both agents think
that the other agent’s estimate of is bigger than 0.5.
In this case, agents would be indi¤erent if all their information were private. But both
expect that the other will attack because the information in the front page of the Financial
Times matters more for what I think that you think.

24
Back to the model Consider agents follow a cut-o¤ strategy: they will invest if their
posterior is larger than , which is a function of parameters and the public signal.
Agent 1 knows agent 2 will play A if her posterior is larger than , or
2
y + 2 x2 ( 2 + 2) 2
y 2

2+ 2
> , x 2 > 2
, x 2 > + 2
( y)

and given the distribution of probability of x2 (from the point of view of agent 1) the
probability x2 is smaller than + ( 2 = 2 ) ( y) is given by:
r !
2+ 2 2
1 + ( y)
2 2 2+ 4 2

Hence agent 1’s the expected payo¤ from investing is:

r !
2+ 2 2
v 1; = 1 1 + ( y)
2 2 2+ 4 2

In equilibrium, an agent with with posterior i = has to be indi¤erent, so is given by:

p
v ( ) = v( ; ) = ( ( y)) (14)

where is given by
2 2
+ 2
= 4 2+2 2

The derivative of v with respect to is given by:

dv p p
=1 ( [ y])
d
If is large enough, this derivative is negative for some values of y and we have multiple
equilibria depending on the realization of the public signal. That is illustrated in Figure 3.4
p p
of Morris and Shin (2003). Since [ y] 1= 2 , there is a unique equilibrium
p
in the model whenever 2 .

p
How does the public information a¤ect the equilibrium? Now, suppose 2 , so there
is a unique equilibrium. The equilibrium condition v ( ) = 0 in (14) for = i can be
written as: 2
y + 2 xi p 2
y + 2 xi
2+ 2 2+ 2
y =0

The question now is how much the private signal has to change to compensate for a change
in the public signal and still leave agent 1 indi¤erent. The answer can be obtained by
applying the implicit function theorem to the above equilibrium condition:
2 p
dxi 2 + (:)
= p (15)
dy 1 (:)

25
We can also calculate how much the private signal would have to change to compensate
for a change in the public signal if there was no strategic e¤ect. Fix agent i’s posterior:
2
y + 2 xi
i = 2+ 2
=c

Then the implicit function theorem yields:

2
dxi
= 2
(16)
dy
which is smaller (in absolute value) than the derivative in (15). We can de…ne the publicity
multiplier as the ratio between (15) and (16), which yields:
2p
1+ 2 (:)
= p
1 (:)
q
2 + 2
1+ 1 2 +2 2 (:)
= q
2+ 2
1 2 2 +2 2 (:)

which is larger than 1 and decreasing in . In case ! 1, ! 0. Hence the derivative

in (15) is zero (as one would expect), but the multiplier is still larger than 1. As the
precision of the public information increases ( decreases), the multiplier gets larger,
public information has a stronger e¤ect on the equilibrium.

2.1.7 Discussion about contagion

There is large evidence that a crisis in a country makes other countries more subject to
crises. The Mexican crisis in the end of 1994 had large impacts in other Latin American
countries, especially in Argentina. The Asian crisis started in Thailand in July/1997
and was spread to many other countries in Southern Asia. Latin American countries
were also a¤ected (the Asian Flu). The Russian crisis in 1998 also had strong impacts
in other developing countries like Brazil and Mexico (the Russian virus). The Brazilian
devaluation in 1999 let Argentina in a more fragile position.
What are the possible linkages between countries? Trade links or other fundamental
links are plausible candidates, but one would be hard pressed to …nd strong commercial
links between Thailand and Brazil. Changes in “…rst world”prices might have e¤ects as
well (for example, the strong impacts in Latin America of changes in US interest rates
and the e¤ects of the devaluation of the Yen with respect to the dollar in 1995-1996).
What else?
A crisis in one country might generate information about the way the economy works
that is relevant to other countries – and there is where the magnifying e¤ects of public

26
information come in. A crisis in Thailand can be interpreted as a public signal, everyone
knows about it. Even if we think that provides little information about Brazil, we know
everyone else has got this information, and if coordination is important, that a¤ects
behaviour.
The literature also discusses a di¤erent channel, where contagion is caused by wealth
e¤ects: investors that lost money in a country need to withdraw from some other. Using
the global games methodology, Goldstein and Pauzner (2004) present a model of contagion
caused by wealth e¤ects. And now it is advertising time again: Guimaraes and Morris
(2007) extend the model in Morris and Shin (1998) in order to account for risk, wealth
and portfolio e¤ects.

2.2 Rational herd behavior

If my payo¤ depends on your actions, we may choose the same action due to the strategic
complementarities discussed above. But conformity of behavior may arise because of an
informational cascade. One agent observes actions of others and learns from it. That is
called herd behavior or informational cascades

2.2.1 Bikhchandani, Hirshleifer and Welch (1992)

Here I go through the model at section 2 of BHW with a slight modi…cation: BHW assume
that if the agent is indi¤erent, she will adopt either action with probability 50%. I will
assume that if the agent is indi¤erent, he will follow its own signal. It makes no di¤erence
in the general message but simpli…es the algebra a bit.
Suppose agent 1 receives signal H (without loss of generality). Then, from his point
of view:

Pr(V = 1jH) = p
Pr(V = 0jH) = 1 p

Since p > 0:5, he chooses to adopt. Thus, the following players know he has got signal
H.

27
 Suppose agent 2 receives signal H. From her point of view:1
Pr((V = 1) \ (H; H))
Pr((V = 1)j(H; H)) =
Pr((H; H) \ (V = 0)) + Pr((H; H) \ (V = 1))
So:
1 2
2
p
Pr((V = 1)j(H; H)) = 1 2 1
2
p + 2
(1 p)2
Similarly,
1
2
(1 p)2
Pr((V = 0)j(H; H)) = 1 2
2
p + 12 (1 p)2
And, as p > 21 , she chooses to adopt. If she receives signal L, then, from her point of
view:
1
2
p(1 p) 1
Pr((V = 1)j(H; L)) = 1 1 =
2
p(1 p) + 2
p(1 p) 2
1
2
p(1 p) 1
Pr((V = 0)j(H; L)) = 1 1 =
2
p(1 p) + 2
p(1 p) 2
She is indi¤erent. So, according to my tie-breaking convention, she follows her own
signal and decides to reject — in the paper, due to their tie-breaking rules, she takes
either action with 50% probability.
Now, consider agent 3. If agents 1 and 2 have taken di¤erent actions. Agent 3 knows
that they have got di¤erent signals. Thus, before agent 3 sees her own signal, she thinks
that: Pr(V = 0) = Pr(V = 1) = 12 . So, the game is as in the beginning.
What if agent 1 and 2 have both decided to adopt? Agent 3 knows that both have
received signal H. If she also receives signal H, she will also decide to adopt, as from
her point of view, Pr(V = 1) > Pr(V = 0) — is it clear for you without doing the
calculations? But what if she receives signal L? Then, she will think: “3 signals, 2 H and
1 L. Which is more likely: (V = 0) or (V = 1)?”
1 2
2
p (1 p)
Pr((V = 1)j(H; H; L)) = 1 2 1
2
p (1 p) + 2
p(1 p)2
1 2
2
p(1 p)
Pr((V = 0)j(H; H; L)) = 1 2 1
2
p (1 p) + 2
p(1 p)2
1 In general, given history h,
Pr(V = 1 \ h)
Pr(V = 1jh) =
Pr(V = 1 \ h) + Pr(V = 0 \ h)
and with m realizations of signal H and n signals in total, that is:
1 m
2
p (1 p)n m

1 m 1
2
p (1 p)n m + 2
(1 p)m pn m

28
 Again, as p > 12 , she chooses to adopt, although she got signal L. The intuition is:
more signals pointing to (V = 1) makes (V = 1) more likely.
What will agent 4 do? She knows that agent 3 would adopt anyway, so agent 3’s action
is uninformative. She knows that agents 1 and 2 have received signal H. Thus, she is in
the same situation of agent 3 and will adopt regardless of her signal. Do I need to repeat
this spiel for agent 5?

That characterizes an information cascade. Agents act regardless of their own infor-
mation, not because they are crazy but precisely because they are rational — they extract
information from others’actions instead of ignoring it.
A key (and interesting) feature of the equilibrium is that agents may observe many
others moving before them, but only the actions of the …rst ones are informative. The
others are just following the herd. So, the information they had is not passed ahead.

Discussion of the key assumptions:

Action: 0 or 1. If there was a continuous set of actions and agents’optimal action de-
pended on the probability they assigned to true value being 1, everybody would infer
the signal received by an agent from her action. Thus, there will be no information
cascade: individuals would always consider their own information and eventually the
true state would be known.

Exogenous order of movements (which can be relaxed, see Caplin and Leahy, 1994)

My signal and your signal have (almost) the same weight in my decisions. What do
you think of this assumption? Is it rational?

29
3 Banking crises

Banking crises are related to international …nancial crises for two reasons: because cur-
rency and debt crises are related to banking crises; and because some models used to study
international …nancial crises are based on models of banking crises, especially Diamond
and Dybvig (1983).

3.1 Bank runs

The model of Diamond and Dybvig (1983) has become the standard model of bank runs
in the literature. First, the model shows that banks issuing demand deposits can improve
on a competititve market. However, that also creates an undesirable equilibrium (bank
runs), in which banks that would have no problem if people just believed so might be
victims of self-ful…lling prophecies. The presentation here also borrows from Diamond
(2007).

3.1.1 Diamond and Dybvig (1983)

Illiquidity of some assets in this model explain both the existence of banks and their
vulnerability.

Setup There are 3 periods (t = 0; 1; 2), one homogenous good, and measure-one contin-
uum of agents.
Agents get an endowment of 1 unit at t = 0 (and nothing else).
There are 2 technologies:

1. Storage, people can store goods from one period to the other. It is liquid and not
publicly observable.

2. A productive technology: one unit invested at t = 0 yields either 1 unit at t = 1 or

R units at t = 2. At t = 1, agents have to decide whether they wait until t = 2 or
whether they get 1 at t = 1.

At t = 0, consumers are all the same. But at t = 1, a consumer learns its own type,
which might be:

1. Type 1: consumes at period one, utility is u(c1 ).

2. Type 2: consumes at period two, utility is u(c2).

30
 The type is unobservable to others. To make the problem interesting, R > 1. Moreover
the utility function is increasing and concave, the Inada conditions hold and the coe¢ cient
of risk aversion has to be large enough ( u00 (c)c=u0 (c) > 1).
A fraction 2 (0; 1) of agents is type 1.

Competitive markets can do nothing to improve allocation An agent by itself can invest
in the technology at t = 0. If the agent happens to be type 1, she consumes 1 at t = 1. If
she happens to be type 1, she consumes 1 and gets u(1), if she is type 2, she gets u(R).
Competitive markets cannot improve upon that. Period 0 price of c1 has to be 1, and
1
the price of c2 has to be R at both t = 0 and t = 1, otherwise there are arbitrage
opportunities.

If types were public observable... A central planner that can observe everyone’s type
would choose c1 and c2 in order to maximize
u(c1 ) + (1 )u(c2 ) (17)
subject to the budget constraint
(1 )c2
c1 + =1
R
hence (17) can be written as
R(1 c1 )
u(c1 ) + (1 )u
1
which yields
u0 (c1 ) = Ru0 (c2 )
For example, if
1
u(c) = 1 (18)
c
then
c2 p
= R
c1
which is di¤erent then what an agent can do by herself (and the competitive equilibrium).
Here, the type one agent gets proportionally more and the type 2 agent gets less. There
is risk sharing, more projects are interrupted at t = 1 but that is good for the agents
ex-ante. If risk aversion was low enough, that would not be true, because interrupting
more projects at t = 1 also means that the total available for consumption is smaller.
In equilibrium then,
p
R R
c1 = p , c2 = p > c1 (19)
1 + R 1 + R

31
In case of log utility... The …rst order condition is
1 R
=
c1 c2
and using the budget constraint, we get:

c1 = 1

which means agents do not want more risk sharing than what they can obtain by them-
selves.

Demand deposit contracts Now suppose there are competitive banks that can o¤er the
following contract: agents deposit their money at t = 0 at the bank. The bank invests in
the new technology at t = 0. Agents can either get c1 at t = 1 or whatever is left at t = 2.
In order to pay agents at t = 1, the bank interrupts production using the productive
technology.
Now suppose that the type 1 agents choose to withdraw at t = 1 and all the others
wait until t = 2. Since c1 and c2 solve the central planner problem assuming agents
consume at t = 1, we know that the 1 type 2 agents will get c2 . Since c2 > c1 , that
is indeed the optimal strategy for type 2 agents, so the demand deposits implement the
…rst best in this economy. That shows how banks can improve on a competitive market
by issuing demand deposits.
But now suppose everyone decides to withdraw at t = 1. There will be no money left
at t = 2. So for type 2 agents it is optimal to try to get their money at t = 1 (it is a
choice between the possibility of getting c1 and the certainty of getting zero). That means
that if everyone expects all agents will withdraw early, withdrawing early is the optimal
strategy. A bank run is thus a self-ful…lling prophecy.
Multiple equilibria arise here because the amount of resources left to t = 2 is increasing
on the number of agents that leave their money at the bank, and leaving the money at the
bank pays o¤ only if there are enough resources at the bank at t = 2. The fundamental
frictions in the model are the illiquidity of the productive technology and the fact that
types are unobservable.

32
Example Using the utility in (18), with = 1=4 and R = 2 an agent’s utility in the
equilibrium without banks would be:
1 3
u(1) + u(2)
4 4
1 1 3 1
= 1 + u 1
4 1 4 2
= 0:375

Using the equilibrium condition in (19), we get that:

c1 = 1:28 , c2 = 1:813

and
1 3
u(1:28) + u(1:813)
4 4
= 0:391 > 0:375

How do we implement this equilibrium?

Instead of buying the risky asset, we can leave enough to pay the agents at t = 1 in
the illiquid asset. Since at t = 1 we will need an amount equal to:
1
1:28 = 0:32
4
we can leave 32% of the economy’s resources for the type 1 agents and invest 68% of the
resources in the illiquid technology. Then at t = 2, we will get, for each agent:
0:68 2
3 = 1:813
4

Hence the bank is not generating more units of consumption than the average agent in
autarky, it is actually investing less in the illiquid technology. The bene…ts come from
risk sharing.

Extensions and policies

When the illiquid asset yields < 1 instead of 1 at t = 1, agents in autarky are
worse o¤, but the bank is not, because it can buy the liquid asset (that yields 1 with
certainty) for the type 1 consumers.

Instead of consumers with risk aversion, liquidity could be demanded by entrepre-

neurs who may or may not get a pro…table project at t = 1 (see Diamond, 2007).

33
 Options considered in Diamond and Dybvig (1983) to deal with bank runs include
the suspension of convertibility: the idea is that if I know not many agents will be
able to withdraw at t = 1 and there will be money left for type 2 agents, I don’t
have incentives to withdraw at t = 1 (if I am a type 2 agent). The role of govenment
policy in their paper is to prevent a bad equilibrium rather than moving an existing
equilibrium.

What else? Using an approach similar to Morris and Shin (1998), Goldstein and Pauzner
(2005) …nd a unique equilibrium in this model and analyze the e¤ect of parameters on
the equilibrium and the e¤ect of di¤erent policies.
Chang and Velasco have a few papers with similar versions of a model of …nancial crises
in emerging markets based on Diamond and Dybvig (1983), with banks in the domestic
economy pooling resources and providing liquidity, but borrowing from abroad. See, e.g.,
Chang and Velasco (1998).

3.2 Banking crises and moral hazard

3.2.1 The moral hazard e¤ect

The game: ‘heads I win, tails the taxpayeres lose’:

Suppose that there is a potential investment that will cost $70 million up front. If all
goes well, the project will yield $90 million. That will occur with probability 1=2. But
with probabilty 1=2, the return will only be $30 million. The expected payo¤ then is
(1=2 $90m) + (1=2 $30m) = $60 million. Ordinarily, this investment would never be
made.
However, bailout garantees change the result. Suppose that an investor is able to
borrow the entire $60 million because everyone (including him and the lenders) knows
that the government will protect them if his project fails and he cannot repay. Then,
from his point of view, he will make $20 million (= $90m $70m) with probability half
and walk away with nothing with probability half.

The solution seems to be simple: the government cannot bailout such projects. But that
is not simple. Diaz-Alejandro (1985), studies the example of Chile in the 80’s. “Good-
bye …nancial repression, hello …nancial crash”, that is what happened. Lots of bank
regulations were removed with the aim of ending …nancial repression and the government
of Chile had pledged not to bail-out banks if they crashed. However...

34
 N

good - ½ ½ - bad

A A

careful risky careful risky

(3,0) (1,0)
(1,0) G

bail-out
not

(0,-5) (-10,-10)

Figure 4: The moral hazard game

The externality A bank is a borrower and a lender. Its assets are bonds, loans, other
…nancial assets. Its liabilities are loans, bonds and other …nancial instruments. If a bank
crashes and it cannot pay its debts, its bankrupcy can lead to the failure of other banks
and companies — it may be the …rst domino to fall — because its liabilities are other
banks and companies assets.
So, bailing-out a failing bank is costly. But not bailing-out may be worse, due to the
negative externalities spread to the rest of the economy.

The credibility issue Consider the game shown at …gure 4.

Players:

N: nature,

A: agent (banks, big companies, lenders...),

G: government.

When the government says: “I will not bail out insolvent banks/companies”, the gov-
ernment is saying: “I will play not”. If the government is indeed playing not, the agent
gets payo¤ of 1 if he plays careful and expected payo¤ of -3.5 if he plays risky. Thus, the
agent playing careful and the government playing not is a Nash equilibrium: nobody has
any incentive to deviate.

35
 However, this Nash equilibrium is not credible (it is not a Subgame Perfect Nash
Equilibrium). If we get into a situation in which the government is called to action,
choosing not yields -10, while choosing bail-out yields -5. None is great, but bail-out
causes less damage. So, that is what the government chooses. Now, the agent knows this,
the threat of not bailing-out is not credible. Thus, the agent gets payo¤ of 1 if he plays
careful and expected payo¤ of 1.5 if he plays risky. Thus, the agent plays risky and, in
the event of a bad state of nature, hello …nancial crash.
The key assumption in this game is that, once a banking crisis takes place, the payo¤
of not bailing out the banks is smaller than the payo¤ of helping them. Diaz Alejandro
(1985) argues that is true in reality: promises to play not at the last node are not credible.
Policy implication: regulation, ex-ante actions are needed.

3.2.2 The Asian crisis of 1997

Many analysts have considered that moral hazard — the game ‘heads I win, tails the
taxpayers lose’— played an important role in the Asian crisis of 1997 (see, e.g., Corsetti,
Pesenti and Roubini, 1999). In Flood and Garber (1984), the speculative attack results
from expectations of in‡ationary …nancing. This pushes up the “shadow exchange rate”
(in Flood and Garber, as PPP is assumed, that occurs instantaneously) and that leads to
a speculative attack. But expectations of money creation come not through the expansion
of domestic credit but from the fragility of the economy and the banking system (Corsetti,
Pesenti and Roubini (1999) point to the expectations of bail-outs as a key factor for that).
Other analysts will argue that those Asian countries had not bad economic fundamen-
tals and that a self-ful…lling crisis in the form described by the second generation models
happened. Radelet and Sachs (1998) is a well known example of a paper defending such
view.

36
4 Sovereign debt and default

4.1 A model with exogenous interest rates

Before we go to models of sovereign default, it will be helpful to study a model with an

exogenous relation between the interest rate and level of debt and no default.

Model Consider a small open economy with constant output y. Time is continuous and
there is a measure-one continuum of agents. The economy can borrow or lend from abroad
and b is the level of assets ( b is the level of debt). Consumption of the representative
agent c is given by
b_ = y + r(b)b c (20)
where r () is a function that yields the interest rate. For b 0, r = r , we’ll say r = 0.
For b < 0, interest rates increase with the level of debt (so r0 (b) < 0, since debt is b). The
idea here is that a high level of debt makes default more likely. This will be formalized
in the next section.
The representative agent maximizes.
Z 1
t
U= e u(ct )dt
0

where u (c) is the instantaneous utility of the representative agent. Let’s suppose >r ,
so that we get positive debt (negative assets) in equilibrium.

Equilibrium The current value Hamiltonian is

H = u(ct ) + t (y + r(b)b c)

The optimization conditions are

@H
= 0 ) u0 (ct ) = t
@c
and
@H _t
= t
@b
) (r0 (b)b + r(b)) = t
_t

The …rst condition implies

u00 (ct )ct = _ t
Hence
u0 (ct ) (r0 (b)b + r(b) )= u00 (ct )ct

37
Thus
ct r0 (b)b + r(b)
= (21)
ct
where is the relative risk aversion coe¢ cient
u00 (ct )ct
=
u0 (ct )

Equations (20) and (21) characterize the equilibrium. In steady state, b_ = 0 and c = 0.
Hence

= r0 (b)b + r(b) (22)

c = y + r(b)b

The …rst equation shows that the steady state level of assets is negative, b < 0. Since
> r , we need r0 (b)b > 0 which only happens if b < 0. The idea is that the country
borrows up to the point that the e¤ect of an increase in the level of debt on the amount
it has to pay to foreign borrowers (r0 (b)b) equates r(b). The second equation shows
that steady state consumption is given by output minus debt service.
Note that the expression in (22) can be written as:
@r b
= r+r
@b r
which yields
= r(1 + ) =) r =
1+
where
@r b
=
@b r
is the elasticity of interest rates with respect to debt.

An example To get intuition, let’s say

r(b) = b
2
for b < 0, hence measures how interest rates are sensitive to the level of debt. Then
r0 (b) = =2, and we get
b=

the level of debt is smaller if interest rates are very sensitive to the level of debt. In
equilibrium, r(b) is given by:
r=
2

38
(which makes sense since the elasticity = 1). Thus
2
c=y
2
which implies that steady state consumption is actually larger for high values of : if bor-
rowing is too expensive, the domestic country borrows less, so steady state consumption
is larger. In sovereign default models, it is essential to look at incentives to contract debt
ex-ante, this already shows up here.

The e¤ect of unexpected shocks The model can be used to analyse how the economy
works when hit by unexpected shocks. That is not how we do research nowadays. The
model has been solved assuming y and the function r () are …xed, so technically a shock
to parameters is a zero probability event in the model. So what is the point of studying
0-probability events? Well, even if agents attached a positive (and sizable) probability to
shocks, any changes in y and r () would still have an unexpected component and the e¤ect
of the unexpected component on the model is very similar to what we have here. Thus
the intuition we get from this model carries on for more complicated models.
The dynamic system can be represented in a phase diagram (Figure 5). We see the loci
b_ = 0 and c = 0, the saddle path and the direction of the movement around the diagram.

c=0
c

b=0

Figure 5: Phase diagram

Figure 6 shows the e¤ect of a reduction in y: the locus b_ = 0 shifts down. Figure 7
shows the e¤ect of an increase in world interest rates (say r goes up but its derivative is
unchanged): the locus b_ = 0 shifts down and the locus c = 0 shifts to the right.
We can then analyze what happens when a temporary unexpected shock hits the
economy.

39
 c=0
c

b=0

Figure 6: E¤ect of a reduction in y

c=0
c

b=0

Figure 7: E¤ect of an increase in r(b)

Consider an unexpected and temporary negative shock to y that will last until time
T : then the locus b_ = 0 shifts down. The shock is temporary so eventually the
economy will get back to the saddle path depicted at Figure 5. Since b is a stock, it
cannot jump, so c jumps down (but not all the way to the steady state equilibrium
at Figure 6, it stays above that). Hence the economy moves to the left and up, and
meet the original saddle path at time T (anticipated jumps to consumption do not
happen, they are not optimal). Then the economy goes back to its steady state.
Intuitively, as the unexpected negative shock hits, consumption jumps down, but the
economy starts to accumulate debt and consumption slowly comes back. When y goes
back to its previous level, the economy repays the extra debt it had accumulated.
More debt implies higher interest rates, which can be loosely interpreted here as

40
 higher default risk. Note that this is without making any assumption about how the
level of output a¤ects incentives for repaying.

Consider an unexpected and temporary positive shock to r that will last until time T
(say a constant is added to interest rates): the locus b_ = 0 shifts down and the locus
c = 0 shifts to the right. The shock is temporary so eventually the economy will get
back to the saddle path depicted at Figure 5. The e¤ect is qualitatively similar to
what we had before. The level of debt is a stock, it does not jump, but c jumps down
and the economy starts moving up and to the left until it meets the original saddle
path at time T . Then the economy goes back to its steady state.
The intuition is similar to what we have before, interest rate rates are high so the
economy starts to consume less and goes back slowly to the previous level of con-
sumption. So it initially accumulates more debt and then pays it back while returns
to its steady state.

4.2 Sovereign debt sustainability

The simple model above provides some intuition on the dynamics of an economy that
wants to smooth consumption, can borrow from abroad but faces higher interest rates
when the level of debt goes up. But how can we think about the relationship between
economic variables (such as the level of debt), default risk and interest rates?
Here we explore a framework based on the idea that countries repay their debt whenever
they can –the literature often use the term ability-to-pay for this class of models. In the
next section we assume countries cannot commit to future repayments and optimally
choose between repaying and defaulting –the literature often labels these as willingness-
to-pay models.

4.2.1 Collard, Habib and Rochet (JEEA 2015)

The model assumes there is an exogenous maximum primary surplus, GDP growth is
stochastic, and the country will pay its debt whenever it is possible to do so.
Notation:

yt : GDP

: maximum primary surplus as a fraction of yt

bt : proceeds from issuing debt at t as a fraction of yt

41
 dt : face value of debt as a fraction of yt , payable at t + 1

gt : gross rate of GDP growth, i.i.d., with cdf F () and pdf f ()

r: risk-free world interest rate

Lenders are competitive and risk-neutral, there is no bailout, no debt recovery (haircut
is 100%). Since the country pays its debt whenever possible, the maximum borrowing
proceeds at t are
dt yt
bt yt = Pr [( + bt+1 ) yt+1 > dt yt ]
1+r
If the probability of repayment is 100%, bt = dt =(1 + r). Since lenders are risk-neutral and
competitive, borrowing proceeds are proportional to the probability of debt payment.
Rearraging the expression for bt yt , we get
yt+1 dt dt
bt = Pr >
yt + bt+1 1 + r
dt dt
= Pr gt >
+ bt+1 1 + r
dt dt
= 1 F
+ bt+1 1+r
The maximum amount the government can borrow at t is
dt dt
bt = max 1 F
dt + bt+1 1+r
De…ne
dt
xt =
+ bt+1
Then,
+ bt+1
bt = max [1 F (xt )] xt
xt 1+r
or
+ bt+1
bt =
1+r
where
max [1 F (x)] x = [1 F (xM )] xM
x

Iterating forward the expression for bt , we get

bt = + + + :::
1+r 1+r 1+r 1+r
or
bt =
1+r

42
where
2
=1+ + + :::
1+r 1+r

4.2.2 The maximum sustainable level of debt

Suppose gt = g. Then = g:
With a non-degenerate distribution,
Z 1
= [1 F (xM )] xM < gf (g)dg g
xM

Variance in the rate of growth reduces .

If > 1 + r, ! 1.
If < 1 + r, then !
1
=
1 1+r

and the maximum sustainable level of debt is

!
1
bM = =
1+r 1 1+r
1+r

4.3 Sovereign debt with no commitment to repay

We now turn to willingness-to-pay models. Eaton and Gersovitz (1981) propose a dynamic
model of a small open economy that can borrow from abroad but cannot commit to repay
debt that has become the standard framework for quantitative macro models of sovereign
debt and default. We will now go through some of the key ideas in Eaton and Gersovitz
(1981).

4.3.1 Endogenous default decision (Eaton and Gersovitz, 1981)

Sovereign debt is di¤erent from …rms’or consumers’debt because the rules of the game
are usually less clear. Thus it becomes important to understand incentives for a country
to default or repay its sovereign debt.
The basic feature of models following Eaton and Gersovitz (1981) are:

Time is discrete.

Small open economy that can borrow from abroad but cannot commit to repay debt.
In most cases, a planner maximizes utility of individuals.

43
 Default or debt payment is a choice. In the simplest cases, default means debt
becomes zero.

Markets are often incomplete, the sovereign country cannot issue debt contingent on
the realization of a shock.

There are foregin creditors, who are often risk-neutral investors, so they lend to the
country and pin down the value of a bond, q:
1
q=
1+r
where is the probability of default (in this case default leads to zero debt) and r
is the risk-free world interest rate. Naturally, has to be endogenous.

We can thus write value functions:

Vp (d; X) = max
0
fup () + E(V (d0 ; X 0 )jX)g
d ;W

Vd (X) = max fud () + E(Vd (d0 ; X 0 )jX)g

W
V (d; X) = max fVp (d; X); Vd (X)g

where d is the level of debt, is the time discount factor, X is a vector of state variables
and W is a vector of control variables. Vp (d; X) is the value function in case the country
chooses to repay. In that case, the domestic planner chooses debt next period (d0 ) and
whatever else (W ). The value function next period will be V (d0 ; X 0 ) which is the maximum
between Vp (d; X) and Vd (X), the domestic country will then choose whatever is larger.
Vd (X) is the value function in case of default, which in the simplest case is not a function
of some level of debt because previous debt is wiped out in case of default. Note that
here, once the country defaults, it is always in default state (next period value function is
Vd (d0 ; X 0 )). Most quantitative models assume an exogenous probability that the country
regains access to international credit markets at every period. Note also that utility in
case of default and debt payment (ud () and up ()) are typically di¤erent, in most models
default entails some punishment that a¤ects something like output or trade.
The country is subject to a budget constraint. A country that chooses to pay debt has
to pay d but can borrow qd0 , both q and d0 are endogenous.
One immediate implication of this model is that Vp (d; X) is decreasing in d (since the
country has to pay debt d) and Vd (X) is independent of d (since default implies all debt is
wiped out). That implies there exists a level of debt d (X), which is typically a function
of the state variables, such that it is optimal to repay debt if it is below that level. Figure
8 sums up this idea.

44
 V

Vd

Vp

Figure 8: Choice of default

Shocks to the economy will a¤ect the vector X and thus a¤ect both Vp (d; X) and Vd (X),
so both curves will cross each other at di¤erent points implying a di¤erent maximum level
of debt compatible with repayment.

4.3.2 A deterministic example

For concreteness, let’s analyze the simplest possible case, there is one good, the economy
receives a constant endowment y in every period and world interest rates are constant
and equal to r . If the country has ever defaulted, a fraction of the output is lost
forever and the country loses access to international capital markets, so consumption is
y forever (there is no investment or savings and no government spending). The time
discount factor is < (1 + r ) 1 . Households get utility u(c) from consumption c.
Consumption in case of repayment is given by

cp = y d qd0

and in case of default is given by

cd = (1 )y
hence value functions are

Vp (d) = max
0
fu(cp ) + V (d0 )g
d
Vd = u((1 )y) + Vd
V (d) = max fVp (d) ; Vd g

The only state variable is d0 , which is known in advance. There is no uncertainty.

Hence the value functions Vp (d0 ) and Vd are known one period in advance, so it is known
whether the country will default or not.

45
 Since there are no state variables, the level of debt d that equates Vp (d) and Vd is
constant. So the bond price schedule faced by the domestic economy is:
(
1
1+r
if d0 d
q= 0
0 if d > d
since is either 0 or 1. It makes no sense to issue bonds at price 0 (the domestic country
gets nothing and then is punished because it defaulted).
What is d? Consider a country that always issues debt d. Then
1
Vp d = u y d+ d + Vp d
1+r
which yields
r
u y 1+r
d
Vp d =
1
The value function of a country that default on its debt implies
u((1 )y)
Vd =
1
Equating Vp d and Vd and doing algebra leads to
d (1 + r )
=
y r
The maximum level of debt a country is willing to repay is proportional to its output
because default punishment is proportional to output (by assumption); it is proportional
to the default punishment; and it depends negatively on r since high interest rates imply
that servicing debt is expensive, default becomes relatively more attractive.

4.4 Recent models

4.4.1 Arellano (2008)

Arellano’s quantitative dynamic stochastic small open economy model builds on Eaton
and Gersovitz (RES 1981).

Setup of the model

Small open economy

Households are identical and have preferences given by

X
1
t
Eo u(ct )
t=0

46
and u(ct ) respects the Inada condition. In the numerical simulation:

ct1 1
u(ct ) =
1
Stochastic endowment: y

Goverment is benevolent and maximizes utility of households.

One international asset: one period discount bonds (B).

–Price of bond: q
–Borrowing qB 0 implies repayment of B 0 next period
–notation: B > 0 means positive assets, B < 0 means the country is indebted.

Government chooses between defaulting or repaying

–no commitment to repay

If government chooses to repay, the resource constraint is:

c=y+B qB 0

Default implies:

–current debts are erased;

–temporary exclusion from international …nancial markets (borrowing and lend-
ing);
–output costs: endowment is lower, equal to y def = h(y) y, h is an increasing
function;
–so:
c = y def

Country re-enters the …nancial market with an exogenous constant probability .

Probability of default:

Creditors are risk-neutral and behave competitively.

–International risk-free interest rate: r

47
 –An arbitrage condition pins down q:

Rf = Rd
1
1+r = (1 )
q
1
) q=
1+r
Timing:

–government starts with assets B, observes realization of endowment y and decides

about repaying or not;
–if the governent decides to repay, borrowing takes place.
–consumers consume.

The value function depends on whether the country repays or defaults on its debt.
Denote by v c the value function condtional on repayment and by v d the value function
conditional on default.
The government chooses what is best for the agents. Suppose we start the period at
state (B; y). Denote the value function by v o (B; y). Then:

v o (B; y) = max v c (B; y); v d (y)

If the country decides to repay, than the value function is given by:
Z
c
v (B; y) = max
0
u(c) + v o (B 0 ; y 0 )f (y 0 jy)dy 0
B y0

where c = y + B qB 0 .
If the country chooses to default, the value function is given by:
Z
d def
v (y) = u(y ) + v o (0; y 0 ) + (1 )v d (y 0 ) f (y 0 jy)dy 0
y0

Equilibrium In a recursive equilibrium:

1. Risk-neutral creditors are indi¤erent between lending to the domestic country or not
(so q is obtained through the above arbitrage condition).

2. Government maximizes.

3. Households eat.

All the action comes from the government’s decision.

48
 Define parameters
and grids

Pick some initial function q (say q = q*)

Pick some initial v’s

Choose default
and B’

Calculate v’s
Iterate

Iterate

Has it No
converged?

Yes

Calculate probabilities
of default

Calculate function q

Has q No
converged?

Yes

We are done

Figure 9: Flow diagram for Arellano (2008)

49
Computation of equilibrium The problem here is partial equilibrium in the sense that we
do not model the outside world (foreign interest rates are taken as given). However, prices
depend on the probabilities of default (arbitrage condition for the risk-neutral creditors)
and although the government internalizes the e¤ects of its actions on the price of debt,
we can’t solve for it directly. We have to …nd q as a …xed point — q is a function.
Figure 9 shows the basic structure of an algorythm to compute the equilibrium of the
model.

Results Negative shocks to y might prompt default or might increase default risk. So
default occurs in bad times, interest rates increase in recessions (owing to the default
premium). See the …gures in the paper.

4.4.2 Cuadra, Sanchez and Sapriza (2010)

The point this paper wants to explain is the following: emerging market economies typ-
ically exhibit a procyclical …scal policy, expenditures fall and taxes rise in recessions.
Moreover, they experience countercyclical default risk. Cuadra, Sanchez and Sapriza
(2010) develop a dynamic model with endogenous …scal policy and sovereign default to
account for that. Here I present a simpli…ed version of that model.

The model A small open economy can borrow from abroad but cannot commit to repay
debt. A representative agent gets endowment yt , with probability density function f and
support [yL ; yH ]. The representative agent maximizes:
X1
t
E0 (u (ct ) + v (gt ))
t=0

where < 1. Private consumption (ct ) and government spending (gt ) are given by:
ct = (1 t )yt

gt = t yt + qt dt+1 dt
where t is the tax rate, dt is debt maturing at t (contracted at t 1) and qt is the
endogenous price of debt. Output Yt is given by:
(
yt if default has never happened
Yt =
g(yt ) otherwise
where g() is such that default happens in bad times (there is less punishment for default
when Yt is low). The government chooses t and dt+1 , where
t
qt =
1+r

50
and t is the probability of repayment in the following period.
The value function if the country repays debt is:

Vp (d; y) = max
0
fu ((1 )y) + v ( y + qd0 d) + EV (d0 ; y 0 jy)g
;d

and if the country defaults, it is

Vd (y) = max fu ((1 )g(y)) + v ( g(y)) + EVd (y 0 jy)g

and
V (d; y) = max fVp (d; y) ; Vd (y)g

Now let’s take for granted that default occurs if and only if yt < Y (dt ), where Y (d)
is the value of Yt that makes the government indi¤erent between defaulting and paying
the debt, Y (d) is increasing in d (in the paper that occurs in equilibrium owing to the
assumptions on the function g). Since at Y (d) the country is indi¤erent between repaying
and defaulting,
Vp d; Y (d) = Vd Y (d)

Since default occurs next period if and only if Yt < Y (d0 ),

Z Y (d0 ) Z yH
0 0 0 0
EV (d ; y jy) = Vd (y) f (y jy)dy + Vp (d0 ; y) f (y 0 jy)dy 0
yL Y (d0 )

First order conditions Consider a country that has not defaulted (so risk of default is an
issue). The …rst order condition with respect to implies
@u @v
=
@c @g
and the …rst order condition with respect to d0 yields
Z yH
@v @(qd0 ) @Vp (d0 ; y) 0 0 @ Y (d0 )
+ f (y jy)dy + Vd Y (d0 ) Vp d0 ; Y (d0 ) f (Y (d0 )jy) = 0
@g @d0 0
Y (d ) @d 0 @d 0

Since Vd Y (d0 ) = Vp d0 ; Y (d0 ) (because the country is indi¤erent between paying and
defaulting in state d0 ; Y (d0 ) ), that becomes
Z yH
@v @Vp (d0 ; y)
q (1 ") + 0
f (y 0 jy)dy 0 = 0
@g 0
Y (d ) @d

where " is the absolute value of the elasticity of the price of debt with respect to d0 ,
@q d0
"=
@d0 q

51
and since
@Vp (d; y) @v
=
@d @g
we get Z yH
@v @u
(1 ") = f (y 0 jy)dy 0
@g 1 + r Y (d0 ) @g g= 0 y 0 d0 +q 0 d00
where the integral term is the expected value of the marginal utility of government spend-
ing conditional on repayment times the probability of repayment. I will write that as
Z yH
@v @v 0
f (y 0 jy)dy 0 = Ey0 y Y (d0 )
0
Y (d ) @g 0 0 0 0
g= y d +q d00 @g
so
@v @v 0
(1 ") = (1 + r ) Ey0 y Y (d0 )
@g @g
Note that the probability of default per se has no …rst order impact in the sense that
a lower makes debt more expensive but debt is also less likely to be repayed so the
at both sides of the equation cancels out. But " has an important e¤ect.
Suppose the case where default is not possible, which means = 1 and " = 0. Then
@u @v
= (1 + r ) Ey0
@g @g
which is a pretty standard result. But now suppose " is large (1 " is small). Then
the government wants a large present marginal utility of government spending (and so a
large marginal utility of consumption as well). That means small levels of consumption
and government spending and, consequently, little borrowing. The reason is that bor-
rowing one extra unit decreases the price of debt (increases the interest rate paid by the
government), so it is better to decrease domestic expenditure and borrow little.
At times when default risk is more important, the negative e¤ect of the level of debt on
the price of debt kicks in, then it is optimal for the government to choose small g and high
(low consumption). Procyclical …scal policy is thus a consequence of counter-cyclical
default risk.

4.4.3 Which shocks matter?

Guimaraes (2011) analyses whether sovereign default episodes can be seen as contingencies
of optimal international lending contracts. The model also considers a small open economy
without commitment to repay debt. The point is that shocks to world interest rates have
strong e¤ects on the incentives for default –much stronger than output shocks if the cost
of default in terms of output loss is proportional to the level of output. Here I present a
simpli…ed version of the argument.

52
The model An open economy can borrow from abroad, but cannot commit to repay
its debts. The economy is populated by a continuum of in…nitely lived agents whose
preferences are aggregated to form the usual representative agent utility function:
X
1
t
u(ct )
t=0

Default implies a permanent output cost, . The fraction of output lost due to default
is , so production is given by:
(
yt , if no default
yt =
(1 ):yt , if default
Default also implies permanent exclusion from international capital markets. Default
costs are permanent, which captures the loss that a country su¤ers by taking an antago-
nistic position towards the rest of the world and never repaying its debts. In the model
this is out-of-equilibrium behaviour, which corresponds to never observing such action in
reality.
There is a continuum of risk-neutral lenders that, in equilibrium, lend to the country
as long as the expected return on their assets is not lower than the risk-free interest rate
in international markets, r . The price of a bond that delivers one unit of the good next
period with certainty, (1 + r ) 1 , will be denoted q . There is a maximum amount of
debt the country can contract that prevents it from running Ponzi schemes but it is never
reached in equilibrium.
The economy’s ‡ow budget constraint is then given by:
(
yt dt + qt dt+1 , if it has never defaulted
ct =
(1 )yt , if it has ever defaulted
There are two states, s 2 fh; lg. We will explicitly condition debt repayment on the
state.

Stochastic interest rates Let’s analyze the case of stochastic q . The price of a riskless
h
bond in international markets is q in the high state and q l in the low state, q h
> q l.
The economy switch to the other state with probability , so the probability st = st 1 is
1 ; with probability , st 6= st 1 .
We want to understand what is the maximum incentive compatible level of debt at
each state? If the country borrows up to the maximum it can, and lenders can always
receive the maximum payment the country is willing to pay, how much debt falls when
the economy goes from the high to the low state?

53
 A risk-neutral creditor that lends q d0 must get an expected repayment equal to d0 .
Denote by dh and dl the repayment conditional on high and low state, respectively, and
d = dh dl . If st 1 = h, a country that borrowed q h d0 has debt dh if st = h and dl if
st = l such that dh (1 ) + dl = d0 .2 If st 1 = l, a country borrowing q l d0 has debt dh
if st = h and dl if st = l such that dl (1 ) + dh = d0 .3
In each period, the central planner chooses between repaying or defaulting. Each option
yields a di¤erent value function and the planner chooses the maximum of the two:

V (q ; d) = max fVpay (q ; d); Vdef g

The value functions conditional on repayment are:

h h
Vpay (d) = max u(y d+q (1 )dh + dl ) + (1 h
)Vpay (dh ) + Vpay
l
(dl )
dh
l l
Vpay (d) = max u(y d+q (1 )dl + dh ) + (1 l
)Vpay (dl ) + Vpay
h
(dh )
dh

where dh and dl are the levels of debt in the high and low state, respectively.
In case of default, the value function in both states is:
u((1 )y)
Vdef = u((1 )y) + Vdef =
1
It makes no di¤erence whether foreign interest rates are low or high if the country is
excluded from international …nancial markets.

Equilibrium The incentive compatible level of debt at the high state dh is given by
h
Vpay (dh ) = Vdef , and the incentive compatible level of debt at the low state dl is given by
l
Vpay (dl ) = Vdef . Hence, Vpay
h
(dh ) = Vpay
l
(dl ). Thus

u(y dh + q h
(1 )dh + dl ) = u(y dl + q l
(1 )dl + dh )

and doing algebra we get to

dh dl q h
q l
= (23)
d 1 q(1 2 )

where d = dh + dl =2 and q = q h + q l =2.

Stochastic technology Now …x the world interest rates and allow y to ‡uctuate between
y in the high state and y l in the low state, y h > y l . The value functions conditional on
h

2 Hence, dh = d0 + d and dl = d0 (1 ) d.
3 Hence, dl = d0 d and dh = d0 + (1 ) d.

54
repayment are:
h
Vpay (dh ) = max u(y h d+q (1 )dh + dl ) + (1 )V h (dh ) + V l (dl )
dh
l
Vpay (dl ) = max u(y l d+q (1 )dl + dh ) + (1 )V l (dl ) + V h (dh )
dh

Should the country go into default, the value functions are:

h
Vdef = u((1 )y h ) + (1 h
)Vdef l
+ Vdef
l
Vdef = u((1 )y l ) + (1 l
)Vdef h
+ Vdef
h
Making Vpay (dh ) = Vdef
h
, we get:

dh q dl + (1 )dh = y h (24)
l
Analogously Vpay (dl ) = Vdef
l
yields:

dl q dh + (1 )dl = y l (25)

Subtracting (25) from (24), we get:

dh dl [1 q (1 2 )] = (y h yl ) (26)

Summing (25) and (24) and manipulating, we get:

d
= (1 q )
y
Using this value of on the above equation yields:
dh dl (1 q ) yh yl
= (27)
d 1 q (1 2 ) y

Stochastic interest rates stochastic technology Compare the expressions for dh dl =d

in (23) and (27): debt relief generated by reasonable ‡uctuations in productivity is at least
an order of magnitude below that generated by shocks to world interest rates.

4.4.4 IMF and …scal adjustment

Gonçalves and Guimaraes (2015) study time consistency of …scal policy and the role of
the IMF in a model of sovereign default. I present here a simpli…ed version of that model.

55
The model An open economy can borrow from abroad, but cannot commit to repay
its debts. The economy is populated by a continuum of in…nitely lived agents whose
preferences are aggregated to form the usual representative agent utility function:
X
1
t
(u(ct ) + gt )
t=0

subject to gt 0 (and ct 0).

In each period, the country’s representative agent receives a stochastic endowment y
drawn from a probability density function f , which is continuous with full support in
the [yL ; yH ] interval, with yH > yL > 0 and iid. It is also assumed that u0 (0) ! 1 and
u0 (yL ) < 1.
Default implies a permanent output cost, . The fraction of output lost due to default
is , so production is given by:
(
yt , if no default
yt =
(1 ):yt , if default

Default also implies permanent exclusion from international capital markets.

There is a continuum of risk-neutral lenders that, in equilibrium, lend to the country
as long as the expected return on their assets is not lower than the risk-free interest rate
in international markets, r . The price of a bond that delivers one unit of the good next
period with certainty, (1 + r ) 1 , will be denoted q .
The sequence of events in a given discrete time period is the following:

1. y is revealed;

2. decisions about defaulting on maturing debt d and ‡oating new debt obligations d0
are made;
0
3. the tax rate prevailing next period is chosen.

Taxes are chosen one period in advance. The private sector has no access to external
capital markets and hence private consumption equals net income:

c = y^(1 )

Government spending is given by:

g = y^ d + qd0 if the debtor has access to capital markets

g = y^ otherwise

56
 0
Value functions Debt and default decisions are made after y is observed but before
has been chosen, re‡ecting the idea that the country cannot commit to a certain level of
taxes when it issues debt.
The value function associated with repaying debt is:

Vp ( ; d; y) = max
0
fu ([1 ] y) + y d + qd0 + EV ( 0 ; d0 ; y 0 )g (28)
d
0
where maximizes

u ([1 ] y) + y d + qd0 + EV ( 0 ; d0 ; y 0 ) (29)

taking d0 as given. V is the maximum of two value functions:

V ( 0 ; d0 ; y 0 ) = max fVp ( 0 ; d0 ; y 0 ); Vd ( 0 ; y 0 )g

If the government opts to default, the value function is:

Vd ( ; y) = max
0
fu ([1 ] y) + y + EV0 ( 0 ; y 0 )g (30)

where
V0 ( 0 ; y 0 ) = max
00
fu ([1 0
] y^0 ) + 0 y^0 + EV0 ( 00 ; y 00 )g
and y^0 = (1 )y 0 . The default punishment kicks in one period after the default decision.
As a tie-breaking convention, we will suppose that the country repays debt if indi¤erent.

Default when tax revenues are low When the constraint g 0 is slack, the derivatives of
the value functions with respect to y are
dVp dVd
= = u0 (y(1 ))(1 )+
dy dy
Hence when the constraint g 0 is not binding, there is never default.
The constraint g 0 can be written as y d + qd0 0. Since , y and d are
predetermined, only qd0 can a¤ect the constraint. Di¤erentiating qd0 with respect to d0
yields
@ (qd0 ) @q d0
= q (1 ) where =
@d0 @d0 q
The probability of repayment is (weakly) decreasing in the level of debt. Hence extra
debt d0 will tend to increase the in‡ow of resources by less than q d0 .
Default only occurs when the constraint g 0 is binding. That happens when tax
revenues are low. In particular,

1. Default can only happen in bad times: for given and d, if there is default for some
y 2 [yL ; yH ] then there is default if and only if y < y~ for some y~ 2 [yL ; yH ].

57
 2. Default is more likely when is smaller: for given y and d, if there is default for
some 2 [0; 1] then there is default if and only if < ~ for some ~ 2 [0; 1].

3. If the probability of default next period is positive, than it is strictly decreasing in

0
.

The proof for the …rst 2 statements is basically the same.The key point is that when
the constraint g 0 is locally binding,
dVp ( ; d; y)
> u0 (y(1 ))(1 )+ (31)
dy

I will show that. Since the constraint on g is binding,

Vp ( ; d; y) = max
0
fu (y(1 )) + E [V ( 0 ; d0 ; y 0 )]g , with qd0 = d y
d

thus
@Vp ( ; d; y) @E [V ( 0 ; d0 ; y 0 )] @d0
= u0 (y(1 ))(1 )+ (32)
@y @d0 @y d0 =( d y)=q

We need show that the second term is larger than . Note that
@(qd0 ) @(qd0 ) @d0
=
@y d0 =( d y)=q @d0 @y d0 =( d y)=q

and since
@(qd0 ) @(qd0 )
= and = q (1 )
@y d0 =( d y)=q @d0
we have
@d0
= (33)
@y d0 =( d y)=q q (1 )
Moreover, since the constraint is locally binding, d0 = ( d y) =q and the derivative of
the value function with respect to d0 is negative (the country would prefer a smaller d0
leading to a smaller g but has to borrow enough to get g = 0),
@(qd0 ) @E [V ( 0 ; d0 ; y 0 )]
+ < 0
@d0 @d0
@E [V ( 0 ; d0 ; y 0 )]
q (1 ) <
@d0
Combining this with (33) yields:
@E [V ( 0 ; d0 ; y 0 )] @b
>
@d0 @y d0 =( d y)=q

Using (32), that implies the condition in (31).

58
 The reasoning leading to the second statement is essentially the same.
The intuition is the following: tax revenues are given by y. When they are low, the
government constraint binds with g = 0, which leads to a larger reduction in Vp in com-
parison to Vd . The optimal unconstrained borrowing d0 is not feasible. The constrained
level of borrowing is suboptimally large in case of repayment. Higher …nancing needs
translate into a larger probability of default, which drives down the price of debt.
That does not imply that the debtor will be transferring more money to creditors,
because creditors always break-even in expected terms. From an accounting perspective,
the increase in interest rates and the decrease in the probability of repayment cancel
each other. However, from an economic perspective, the decrease in the probability of
repaying increases the probability of facing output costs in the future. Hence an increase
in …nancing needs e¤ectively renders the option of repaying relatively more costly.

The time inconsistency problem Focus on the case the country has access to capital
0
markets. The optimal tax rate in equilibrium comes from:
@EV ( 0 ; d0 ; y 0 )
=0 (34)
@ 0
Now suppose the government can commit to some tax rate equal to ^ when it issues debt
d0 at the current period.4 The value function becomes

VpC ( ; d; y) = max
0
fu ([1 ] y) + y d + qd0 + EV (^; d0 ; y 0 )g (35)
^;d

where q is now a function of both d0 and ^.

One implication of the model is that if the probability of default next period is positive
and the debtor can commit to a tax rate ^ when issuing debt d0 > 0,
@q
>0
@^
An increase in ^ increases the probability of repayment and that is re‡ected in the price
of debt q. Higher taxation means less re…nancing needs and thus smaller chances of
defaulting in the future.

We now show that if the debtor could commit to a certain tax rate ^, the chosen ^
0
would be larger than given by (34). Now, ^ maximizes the expression for VpC ( ; d; y)
in (35), taking into account the e¤ect of ^ on q.
The solution to the maximization problem for ^ depends on whether the constraint
g 0 binds. Suppose …rst it does not bind. Then y d + qd0 > 0 and the …rst order
4 For expositional simplicity, I assume there will be no opportunity for commitment in the future.

59
condition with respect to ^ yields:
@q @EV (^; d0 ; y 0 )
d0 + =0
@^ @^
0
The choice of with no commitment is given by (34). When the government can commit
to a tax rate ^ before issuing debt, ^ has an extra positive e¤ect on the value function,
since it positively a¤ects the debt price q.
Now suppose the constraint g 0 binds so that y d+qd0 = 0. Then d0 = ( d y) =q
and
@d0 ( d y) @q
= 2
@^ q @^
hence the …rst order condition with respect to ^ implies
@EV (^; d0 ; y 0 ) @d0 @EV (^; d0 ; y 0 )
+ = 0
@d0 @^ @^
Z yH 0 0
( d y) @q @Vp (^; d ; y ) 0 0 @EV (^; d0 ; y 0 )
f (y )dy + = 0
q2 @^ y~ @d0 @^
In both cases, we are adding a positive term to the marginal e¤ect of increasing ^. If
commitment is feasible, a pledge to implement a certain tax rate ^ will be fully incorpo-
rated in bond prices, which increases the incentives for additional …scal adjustment. The
debtor bene…ts from higher taxes since these mean less frequent defaults.

A commitment device A contingent debt contract that entitles the debtor to receive a
tranfer T contingent on choosing + improves welfare in the debtor country and
leaves the lenders una¤ected. We can see that as a loan of T =(1 + r ) from the borrower
to the lender at the risk-free rate, with a condition that the borrower will only be repaid
if it chooses a tighter …scal stance.
That has a positive e¤ect: it works as a commitment device for the debtor. It also has
a negative e¤ect: since the debtor receives T only in the following period, it might have
to borrow more in the current period to roll over its debt (in case the g 0 constraint is
binding), and borrowing is expensive. However, for a small , the transfer T is of second
order of magnitude because the debtor is close to indi¤erence between and + ,
so the negative e¤ect is dominated by the positive e¤ect.
A private contract along these lines requires that lenders commit to a transfer con-
tingent on a certain level of …scal adjustment, which has to be chosen and then veri…ed.
That requires some degree of commitment and a large amount of e¤ort. Individual credi-
tors would have incentives to free ride on others’monitoring e¤orts. Overcoming the time
inconsistency problem discussed above through private contracting seems to be unfeasible.

60
 However, the problem can be solved by a big long-term player that is able to (i) choose
and verify the optimal level of …scal adjustment; (ii) commit to transfer resources to the
debtor after observing the implementation of the chosen level of …scal policy; and (iii)
take resources from the lenders to fund this transfer.

61
5 Debt crises and coordination

The models on sovereign default along the lines of Eaton and Gersovitz (1981) are all
about solvency and incentives for repayment. However, sometimes countries might go
through liquidity crises: they want to roll over their debts and will be able and willing
to pay the debt later, but cannot roll it over. We start with a simple model of liquidity
crises that illustrates the problem. Then we discuss models where coordination plays a
key role.

5.1 A simple model of liquidity crisis

There is a measure-one continuum of foreign investors, each one has invested 1 dollar in
a small open economy at time t = 0. The deal is the following: they can withdraw their
money at t = 1 and get their dollar back or can wait until t = 2. The resources are
invested in a technology that yields R for every unit invested.
However, withdrawing their investments at t = 1 reduces the return to investment in
this economy. Let x be the fraction of investors that withdraw their money at t = 1. The
gross return in the economy at t = 2 is given by:

(1 x)R + x R

where < 1. So if noone withdraws at t = 1 (x = 0), the return is R indeed, but if

everyone withdraws early (x = 1), early liquidation of projects reduces returns to R
only.
The amount of resources available at t = 2 is given by

maxf0; (1 x)R + x R xg

which is divided among the remaining 1 x agents. For simplicity, the time discount
factor is 1.

Equilibrium Suppose x = 0. Then each agent receives R in the second period. Thus it
is optimal to withdraw early only if R 1.
However, suppose x = 1. An individual waiting until t = 2 will receive 0 if R < 1.
Hence it is optimal to withdraw early if R < 1= .
Consequently:

For R < 1, x = 1 in the unique equilibrium.

For R > 1= , x = 0 in the unique equilibrium.

62
 For R [1; 1= ], there are multiple equilibria. In one equilibrium, x = 1, in the other
equilibria, x = 0.

5.1.1 IMF intervention

Now suppose the IMF can lend any amount to this economy.
Suppose the IMF lends x to this economy at t = 1. Then there are no liquidation
costs. At t = 2, the economy gets R and needs to pay those who had not withdrawn early
(1 x agents) plus the IMF (that has lent x). Suppose the IMF gets the same as each
creditor per dollar it lent. Then each creditor gets R and the IMF gets Rx.
If R < 1, the IMF loses money, so it will not lend. But for R > 1, the IMF lends to the
economy and avoids a liquidity crisis. Anticipating that, private agents have no incentives
to withdraw early in case R [1; 1= ]. The “bad equilibrium” has been eliminated by the
existence of the IMF. Interestingly, the IMF does not need to lend anything, as long as
agents know it will be there to prevent a liquidity crises.
This simple model makes some very strong assumptions:

The IMF has deep pockets. If the IMF cannot cover all early withdrawals, the bad
equilibrium might still exist in a range of parameters.

The IMF and private agents have perfect information about fundamentals.

5.2 Liquidity crises

This is the model of Rochet and Vives (2004).

The balance sheet ot the bank comprises:

Assets: investment (I) and cash (M )

Liabilities: deposits (D) and equity (or long-term debt) (E)

But the model can be interpreted as referring to a country with:

Assets: investment in local projects (I) and reserves (M )

Liabilities: short term debt (D) and local resources available of investment (or long-
term debt) (E)

63
Timing At t = 0, this is the situation, taken as given.
At t = 1, early withdrawals can occur: depositors can withdraw D at no cost. Banks
might have to liquidate some of their investment. In this case, each unit of investment
yields R=(1 + ), with > 0.
At t = 2, the remaining units of investment yield R.
The bank fails (or the country defaults) if it does not have enough resources to cover
D (either at t = 1 or at t = 2).

Payo¤s The relevant decision in the model is the decision about early withdrawals at
t = 1. It is taken by fund managers, and the payo¤s are very simple. Early withdrawals
yield C if the bank does not fail or B if the bank fails. B and C are known positive
constants.

5.2.1 Equilibrium

Let p be the probability the bank fails. Then withdrawing early is the optimal choice if

pB + (1 p)( C) > 0

which implies
C
p> (36)
B+C
Now suppose a proportion x withdraws. The bank has to pay xD. If xD > M , it has
to liquidate projects.
Let’s say y projects are liquidated at t = 1:

1. The bank fails at t = 1 if y > I

2. The bank fails at t = 2 if R(I y) < (1 x)D

At t = 1, each unit of project yields R=(1 + ), so

Ry
= [xD M ]+
1+
where [xD M ]+ means maxfxD M; 0g. Hence

[xD M ]+ (1 + )
y=
R

Case 1: xD M

64
 The bank fails at t = 2 if

RI + (M xD) < (1 x)D

which occurs if and only if

D M
R< Rs
I

Case 2: xD > M
The bank fails at t = 1 if y > I, which means if
[xD M ] (1 + )
>I
R
leading to
[xD M ] (1 + )
R< Rec
I
In case x = 1, the condition becomes
[D M ] (1 + )
R< = Rs (1 + )
I
Moreover, the bank fails at t = 2 if
[xD M ] (1 + )
R I < (1 x) D
R
[xD
M ] (1 + ) + (1 x) D
R <
I
[xD M ]
R < Rs + Rf
I
De…ne m = M=D. Using I = (D M ) =Rs , we get
[x m]+
Rec = Rs (1 + )
1 m
[x m]+
Rf = Rs 1+ (37)
1 m

The case of complete information With complete information about R, there may be
multiple equilibria.
There is an equilibrium with bank failure if R < Rs (1 + ).
There is an equilibrium with no bank failure if R > Rs .

65
The case with incomplete information At t = 0, agents have some prior on R. At t = 1,
they receive a signal si = R + "i . Each agent’s signal is independent from the others.
We’ll consider the case V ar("i ) ! 0, for simplicity. For expositional simplicity only, we’ll
2
assume a normal distribution (0; ).
There is a unique equilibrium where

fund managers withdraw if si < s

the bank fails if R < R

First equilibrium condition: at s , the agent must be indi¤erent between withdrawing

or not. Hence, from (36), p must be equal to . The probability of bank failure given s
is the probability that R is smaller than R , with R = s "i . Hence
R s
p= =

Second equilibrium condition: R is given by Rf from (37). At R = R , the share of

agents with a signal smaller than s is
s R R s
x= =1 =1

Hence, using (37),

[1 m]+
R = Rs 1 +
1 m
Hence R is decreasing in m and in and increasing in .

5.2.2 A lender of last resort

Now, there is an international lender of last resort. It will get a very accurate signal
S = R+ and will lend and amount L to the bank (or country) at t = 1 if S > Rs . It
gets back at t = 2 (if the bank does not fail).
Now the amount that has to be liquidated is
[xD M L]+ (1 + )
y=
R
and the bank fails at t = 2 (in case [xD M L] > 0) if

R(I y) < (1 x)D + L

RI < [xD M L] (1 + ) + (1 x)D + L

66
which leads to
D M [xD M L]
R< +
I I
Let ` = L=D. Using m = M=D and Rs = (D M )=I, the bank fails if
[x m `]+
R < Rs 1 +
1 m
Repeating the steps above, we get to
[1 m `]+
R = Rs 1 +
1 m
The lender of last resort a¤ects the threshold even if it is relatively small (cannot
eliminate the run equilibrium) because if a¤ects coordination.

Extension Corsetti, Guimaraes and Roubini (2006) extend this reasoning considering
that the IMF is a large player. Like in Rochet and Vives (2004), the IMF has only a
limited amount of resources and agents have incomplete information about fundamentals.

5.3 Coordination and the price of debt

Now let’s study a coordination problem that arises when investors are deciding whether
they roll over debt or not. The model follows section 2.3.1 of Morris and Shin (2003)
which is a simpli…ed version of Morris and Shin (2004).

5.3.1 Morris and Shin (2004)

A measure-one continuum of investors chooses whether or not to roll over debt. An

investor that does not roll over receives 2 (0; 1) at t = 1, can be interpreted as
collateral of a …rm’s debt. Rolling over yields 1 if the …rm (or country) is successful able
to repay its debt and 0 otherwise. The …rm is successful and able to repay if l =z,
where l is the fraction of individuals that withdraw early, is the level of ‘fundamentals’
and z is a positive constant.
At t = 1, agents receive a signal xi = + "i , where "i N (0; ) and for agents i and
j, E("i :"j ) = 0. We will think of ! 0 so that we don’t need to worry about their prior
information.
Equilibrium is characterized by 2 thresholds:

An agent chooses to roll over if xi > x .

The …rm is successful if .

67
 An agent with signal xi = x considers that

Pr( > ) = Pr(x "i > )

x
= 1

x
=

The agent is indi¤erent between withdrawing and rolling it over if

x
=

Moreover consider the case = . Then

x
l = Pr(xi < x ) =

At the equilibrium threshold , we know that

l=
z
Hence
=z

Now let’s say that G () is the prior cdf on , g () is the cdf. Then the ex-ante value of
debt as a function of is given by

V( ) = G( ) + 1(1 G( ))
= 1 (1 )G(z )

which implies
@V
= G(z ) z(1 )g(z )
@
Hence an increase in a¤ects price of debt in two ways:

1. It increases the value of debt if happens to be low and the agent chooses not to
roll over. That is the term G(z ).

2. However, it also increases incentives for early withdrawal and hence increases the
threshold = z , making it more di¢ cult for agents to coordinate on rolling over
the debt.

For low values of , the overall e¤ect of on the ex-ante value of debt is actually
ambiguous owing to the coordination channel.

68
6 Bubbles

The so-called “rational” bubbles arise in OLG settings with dynamic ine¢ ciencies and
may also appear in models with in…nitely-lived agents with …nancial frictions. In these
models, the bubble is worth more than the asset’s fundamental value (like money), and
everyone knows this. In these models, equilibria with bubbles typically coexist with an
equilibrium with no bubbles.
This section covers di¤erent models of asset overpricing that are also called bubbles.

6.1 Speculative bubbles

I will go through the numerical example in the seminal paper of Harrison and Kreps (QJE
1978).
Time is discrete. There is an asset that at every t pays dividend dt 2 f0; 1g. Dividends
are paid early in the period, before any trade takes place. Short-selling is not permitted.
There are two types of agents who disagree on the probability that dividends will be
paid in future periods. Both types believe the economy follows a Markov Chain, so that
the probability that dividends will be paid at t + 1 depend solely on whether dividends
were paid at t. De…ne the transition matrix for type a:
" #
Pr (d t+1 = 0jdt = 0) Pr (d t+1 = 1jd t = 0)
Qa =
Pr (dt+1 = 0jdt = 1) Pr (dt+1 = 1jdt = 1)
We assume " # " #
1=2 1=2 2=3 1=3
Q1 = , Q2 =
2=3 1=3 1=4 3=4
Type-2 agents believe the process for dividends is persistent. Type-1 agents don’t. They
are sure about these probabilities and never learn.
The time discount factor is = 0:75.

Present value of future dividends Let p1 (0) and p1 (1) be the present value of future
dividends for type-1 agents when the current dividend is 0 and 1, respectively. We can
write
1 1 1
p1 (0) = p (0) + 1 + p1 (1)
2 2
3 1 3
= p (0) + 1 + p1 (1)
8 8
which yields
5 1 3 3
p (0) = 1 + p1 (1) =) p1 (0) = 1 + p1 (1)
8 8 5

69
 We can also write
32 1 31
p1 (1) = p (0) + 1 + p1 (1)
43 43
and using the expression for p1 (0), we get
13 1
p1 (1) = 1 + p1 (1) + 1 + p1 (1)
25 4
which leads to
11
p1 (1) = = 1:22
9
Then
3 3 11 4
p1 (0) = 1 + p1 (1) = 1+ = = 1:33
5 5 9 3
Repeating the same steps for type-2 agents we get p2 (0) and p2 (1), the present value
of future dividends for type-2 agents when the current dividend is 0 and 1, respectively.
We obtain
4 11
p1 (0) = 3
= 1:33 p1 (1) = 9
= 1:22
2 16 2 21
p (0) = 11
= 1:45 p (1) = 11
= 1:91
Type-2 agents always value the asset more than type-1 agents. One could conjecture
that type-2 agents will always hold the asset and the equilibrium price would be p2 (d). Is
that correct?

Equilibrium prices Suppose asset prices are at states 0 and 1 are p (0) = p2 (0) = 1:45
and p (1) = p2 (1) = 1:91, respectively. A type-1 investor considers buying the stock
when dt = 0 to sell it to type-2 agents whenever the asset yields dividends. Since
Pr (dt+1 = 1jdt = 0) = 1=2 for type-1 agents, this strategy would yield
" #
2 2 3 3
13 1 3 1 3 21
+ + + ::: 1 +
24 2 4 2 4 11
3 1 32
=
8 1 38 11
3 8 32 96
= = = 1:75
8 5 11 55
But this implies that p (0) = 1:45 cannot be an equilibrium. If the price when dt = 1 is
1:91 or more, type 1 agents would be willing to pay 1:75 when dt = 0.
In turn, type-2 agents that buy the asset when dt = 1 know they can sell it to type-1
agents whenever dividends drop to 0 at a price larger than what they would be willing to
pay.

70
 Since short-selling is not permitted, those more optimistic about the asset at every
state hold it. Asset prices are given by
31 31 32 31
p (0) = max p (0) + [1 + p (1)] ; p (0) + [1 + p (1)]
42 42 43 43
32 31 31 33
p (1) = max p (0) + [1 + p (1)] ; p (0) + [1 + p (1)]
43 43 44 44
Conjecturing that p (1) > p (0) yields
3 3
p (0) = p (0) + [1 + p (1)]
8 8
3 9
p (1) = p (0) + [1 + p (1)]
16 16
and doing the algebra we get to
24 27
p (0) = = 1:85 and p (1) = = 2:08
13 13
The asset is always worth more than anyone’s valuation because agents understand
they can sell the stock to others at an in‡ated price (according to their own beliefs). This
is a speculative bubble.

6.2 Riding bubbles

Abreu and Brunnermeier (JFE 2002) propose a model of traders that know an asset is
overpriced but choose to go long anyway. “Riding the bubble” is pro…table unless the
bubble bursts.
Time is continuous. There is one risky asset. At some time t0 (unknown), it becomes
mispriced (by assumption). Arbitrageurs learn that the asset is mispriced at di¤erent
points. Before they learn, they believe the asset is at its correct price.
The asset’s fundamental value is:
(
ert for t < t0
vt =
(1 )ert for t t0
t0
Noone knows t0 , t0 2 (0; 1), and arbitrageurs know that F (t0 ) = 1 e . The
exponential distribution is convenient for several reasons. In particular, it implies that all
arbitrageurs will face the same problem.
It takes time for arbitrageurs to learn about the mispricing. The time an arbitrageur
learns about the bubble is uniformly distributed in [t0 + t0 + ]. It is important that they
don’t learn at the same time and don’t know their rank (whether they learned earlier or
later than others).

71
 An arbitrageur that learns at ti knows that t0 2 [ti ; ti ]:
Arbitrageurs decide whether to sell or not.
They assume that prices only react when the mass of sellers is su¢ ciently high. If a
smaller mass of arbitrageurs sell the asset, they assume that prices do not move because
behavioral traders buy the asset and o¤set any possible price impact. But if the mass of
agents selling short is su¢ ciently large, then behavioral traders learn something is going
on and refrain from buying. This sounds like a very strange assumption. Doblas-Madrid
(2012) presents a model that tells a very similar story without this kind of assumption,
everyone is rational.
The assumption in the paper is that mispricing is completely corrected when aggregate
sales by arbitrageurs exceed
1
K(t t0 ) = K0 1 (t t0 )

and not corrected at all otherwise. The parameter is a ‘large’number.

The cost of a short position is cpt . If the bubble bursts agents get pt .

Equilibrium The equilibrium features trigger strategies. Arbitrageurs start to short at

ti + . Intuitively, as time goes by, the moment the bubble will burst gets closer.
0
Say arbitrageurs trade with delay of . The price correction occurs at t0 + '( 0 ). From
the distribution of t,
'( 0 ) = 0
+ K('( 0 ))

From the condition for bursting the bubble:

1
K('( 0 )) = K0 1 '( 0 )

so

'( 0 ) = 0
+ K('( 0 ))

and
1
K('( 0 )) = K0 1 '( 0 )

yield
'( 0 )
'( 0 ) = 0
+ K0 1

Doing algebra,
0
+ K0
'( 0 ) =
+ K0

72
 Arbitrageurs informed at ti know that:

t0 + '( 0 ) 2 (ti + '( 0 ) ; ti + '( 0 ))

and since distribution of t0 is exponential,

t0 + '( 0 ) Exponential( )

Since
t0 + '( 0 ) 2 (ti + '( 0 ) ; ti + '( 0 ))

for t < ti + '( 0 ) , there is no chance of bursting.

For t ti + '( 0 )
e t
f (t) =
CDF (ti + '( 0 )) CDF (t)
so t
e
f (t) = (ti +'( 0 )) t)
= (ti +'( 0 ) t)
1 e (1 e 1 e
which is increasing in t.
Cost of holding a short position (time interval ):

cpt

Expected bene…t of holding a short position

(ti +'( 0 ) t)
pt
1 e
Hence it is worth holding a short position if

(ti +'( 0 ) t)
>c
1 e
and indi¤erence occurs when t ti = , so

('( ) )
=c
1 e
which leads, after some algebra, to
1 c
'( ) = ln
c
Recall
+ K0
'( ) =
+ K0

73
hence we get
+ K0 1 c
= + ln
+ K0 c
Boring algebra leads to

+ K0 c
= ln
K0 c
so is increasing in , increasing in c, and decreasing in .

74
7 Expectations and option prices

Currency crises are rare events, so data explicitly relating to them are relatively scarce.
But that problem may be overcome using …nancial price data, which are abundant and
re‡ect expectations about currency devaluations.

7.1 Expectations implicit in …nancial prices

7.1.1 Rose and Svensson (1994)

Rose and Svensson (EER 1994) estimate expectations of changes in the exchange rate
before the ERM crisis in 1992. Key questions are: (i) was the crisis expected? (ii) how
much of the change on expectations are explained by macroeconomic variables? Their
answers are: (i) basically no and (ii) basically nothing.

Measuring credibility Notation: t is the interest rate di¤erential between a given country
and Germany and st is the exchange rate (price of a Deutschemark in domestic currency).
Uncovered interest parity implies:
E [ st ]
t =
t
Now, the exchange rate was allowed to ‡uctuate inside a band. Thus, s may be written
as s = x + c, where c is the (log of) the central parity and x denotes deviations of the
spot rate from c. Expected depreciation can be separated in 2 parts:
E [ st ] E [ x t ] E [ ct ]
+
t t t
They use 2 measures of realignment expectations:

1. measure is t. It assumes E [ xt ] = t = 0.

2. Estimates E [ xt ] = t with a linear regression:

xt P
= i + :xt + : t + ut0
t
How to interpret it? g = 10% may mean an expected realignment of 10% with hazard
rate 1/year or an expected realignment of 1% with hazard rate 10/year...

Fundamentals? Rose and Svensson try to assess the impacts of macroeconomic funda-
mentals on some macroeconomic variables with:

75
 a regression of g in a set of macroeconomic variables (with country and time dum-
mies).

the estimation of a VAR.

They …nd that macroeconomic variables have very little impact on credibility.

What to take out of this? Positive results would be hard to be explained given the lack of
a structural model and, especially, the endogeneity of many of the regressors. The authors
claim that it is more di¢ cult to dismiss negative results — if there was some important
relation between those macroeconomic variables and the credibility of the exchange rate
parity, their analysis should have detected it.
The analysis is very simple. Both the estimation of credibility and the assessment of
the importance of fundamentals are quite atheoretical. How can it be improved?

7.2 Option prices

The exchange rate risk in a pegged regime depends on the probability that the peg will
be abandoned and on the expected size of a consequent currency devaluation. To a …rst
order approximation, the forward premium (or the interest rate di¤erential) is roughly
the product of these two variables. However, observing the forward premium alone does
not permit individual identi…cation of the probability of a devaluation and its expected
magnitude: a forward premium of 3% a year may refer to an expected devaluation of 30%
with probability 10% a year, or an expected devaluation of 5% with probability 60% a
year, and so on.
Options are a richer source of data because they provide information about the proba-
bility density of the exchange rate at di¤erent points. So it is possible to disentangle the
“thickness of the tail of the distribution”(probability of a devaluation) and the “distance
from the tail to the center”(the expected magnitude of a devaluation).
To give a simple intuition for identi…cation, suppose the price of an asset tomorrow will
be 1 with probability 1 p and 3 with probability p. In a risk-neutral world, a call option
with strike price 1 costs 2p, a call option with strike price 2 costs p. If the probability of
a devaluation (p) increases, both options get more expensive but the ratio of their prices
remains equal to 2. If the magnitude of the devaluation increases from 3 to 4, the option
with strike price 1 will cost 3p, a call option with strike price 2 will cost 2p — the ratio
changes.

76
7.2.1 The Black and Scholes model

Before we start studying the options, let’s take a look at the basic asset pricing model.
The Black and Scholes model assumes that the asset follows:
dS
= :dt + :dZ
S
where:

is the instantaneous expected return on the asset (ex-dividend);

is the instantaneous variance conditional on no jumps and

Z is a standard Wiener process;

The B&S model is the benchmark model in …nancial applications. According to the
model, the distribution of returns on an asset is log-normal. The data, however, does not
fully comply with the B&S formula. In particular, the tails of the distribution are too
thick.
The B&S price of an (European) option depends on observed variables (interest rate,
spot value of the asset, strike price, time to maturity) and the volatility. With the price
of an option, we can calculate its implicit volatility. Usually, we observe that the implicit
volatility increases with jS Xj. This generates the so called volatility smiles.

7.2.2 Campa, Chang and Refalo (2002)

Campa, Chang and Refalo (JDE 2002) use options to measure the credibility of Brazilian
exchange rate regime. Among …nancial prices, options are better sources of information on
the expectations about a peg than future prices (or the interest rate di¤erential) because
their value at maturity is nonzero only if the exchange rate goes beyond a certain level (the
strike price). So, if there is data on options of di¤erent strike classes, there is information
about the probability density of the exchange rate at di¤erent points, and it is possible to
uncover more information about the expectations on the path of the exchange rate. For
example, it is possible to identify the probability that the currency peg will be abandoned
and the expected magnitude of a devaluation (conditional on its occurrence).
Campa et al employ an interesting non-parametric approach that builds on the seminal
contribution of Breeden and Litzenberger (1978).

Market expectations and option prices Under risk neutrality, the price of a call with
strike price K and expiration date T is:

77
 Z 1
1
CK;T = (ST K):f (ST ):dST
1 + it K
Di¤erentiating if we respect to K, we get:
Z 1
@CK;T 1 1
= : f (ST ):dST = : [1 F (K)]
@K 1 + it K 1 + it
Di¤erentiating again, we get:
@ 2 CK;T 1
2
= :f (K)
@K 1 + it
Intuition:
A one-dollar increase in the strike price decreases the value of the call by (the present
value of) an amount equal to the probability that the option will …nish “in-the-money”.
The higher the probability the option …nishes in the money, the more likely a one-dollar
increase in the strike price will matter to the option holder and the greater the decrease
in the option price.
Thus, the second partial derivative of the option price with respect to K yields the
probability density function of the exchange rate at date T .
The probability functions derived from option prices are the so called “risk-neutral”
probabilities. They can di¤er from the real pdf’s due to risk considerations, but never-
theless they reveal important information about expectations about an asset.

Estimation If we had a continuous of options (or, if we had a lot), we could just evaluate
numerically the derivatives. The available data is de…nitely not enough for that.
We could do some interpolation (e.g., spline) and calculate the pdf. Problem: the call
prices are not always a convex function (even without interpolation). We do not want
negative pdf’s.
Methodology Campa et al employ: they obtain the implied volatility of each option
as a function of the strike price. That yields a “volatility smile”. Then, they transform
it into a continuous call price function that is twice-di¤erential in strike — which can
be done either by …tting the implied volatility as a quadratic function or by some cubic
spline interpolation. Having the price of a call option as a continuum function of the
strike prices, we apply the formula and get the densities.

Again, fundamentals? They ask the question: can realignment “intensity” be explained
by the usual macro variables? They regress their measure of intensity of devaluation in a

78
bunch of macro variables. The intensity measure is the following:
Z 1
G(T ) = (ST S):f (ST ):dST
S

which implies:
G(T ) = CS;T :(1 + iT )
Results: the level of international reserves is the only signi…cant variable in the regres-
sion (and it is endogenous). They conclude that results are consistent with past evidence:
“macroeconomic variables are largely unable to explain intertemporal movements in re-
alignment risk”.
There is no theory behind their regression. Should the macro variables impact proba-
bility, expected magnitude, or both?

7.2.3 Bates (1991)

The question David Bates is asking is: was the crash of ’87 expected? That is: were put
options too expensive prior to the crash?
In the …rst section of the paper, Bates examines the skewness of the implicit distri-
bution. He …nds that in the year leading up to the crash the probability of a fall was
higher than the probability of a large increase. I will jump to the second section in which
he presents a model (actually very similar to Merton, 1976) and estimates its parameters
implicit in the prices of options.

The model The asset (in this speci…c case, the S&P 500 index) is assumed to follow:
dS
=( :k) :dt + :dZ + k:dq
S
where:

is the instantaneous expected return on the asset (ex-dividend);

is the instantaneous variance conditional on no jumps;

Z is a standard Wiener process;

k is the random percentage jump conditional on its occurrence. It is lognormally

distributed:
2 2 2
–ln(1 + k) N( =2; ) N ( 0; ):
–E(k) k=e 1

79
 is the hazard rate of the Poisson event and

dq is the Poisson counter: Pr(dq = 1) = :dt

His proposition 2 shows that contingent claims are priced as if investors were risk-
neutral and the asset price followed a similar jump di¤usion (page 1025) with di¤erent
parameters. Saying di¤erently, by estimating the above model, we are obtaining the
risk-adjusted parameters.

Figure 10: Example of a di¤usion path

The asset follows a Black and Scholes di¤usion path almost all the time. Sometimes,
a Poisson event happens (hazard rate is ) and there is a discrete jump.
The parameters of this model are:

the volatility ;

the hazard rate ;

the mean of the jump ; and

the standard deviation of the jump .

How changes in each parameter changes the distribution of probability of an European

option:

80
 : increases standard deviation of the no-jump scenario;

: increases the weight of the jump scenario;

: pushes the part of the pdf that refers to the sump scenario further away from the
no-jump scenario;

: increases standard deviation of the jump scenario.

How each pararameter in‡uences the prices of call options?

Estimation The model yields a closed form solution for the price of a call option (equation
11 in the paper). Then, it is assumed that the observed price of an option equals the
corresponding model price plus an additive distrubance term (it could be a multiplicative
error term also, both have some inconveniences).
Then, a cross-sectional data sample with identical maturities was used and implicit
parameters were estimated via non-linear least squares for all days in the sample. The
parameters are not constrained to be constant over time.

Results It is hard to argue that the estimation succeeded in making a clear distinction
between the probability and the expected magnitude of the devaluation (…gure 6). It
seems that the estimation is picking up something happening at the tail, but it really
can’t tell what is probability and what is magnitude.
Figure 7 shows the “price of risk” ( :k). We can see that :k < 0 especially from
June-1987 to August-1987 (dates are not shown in the horizontal axis, June to August is
the period with higher risk of a jump down.
Interestingly, in the 2 months right before the crash, the risk of a jump implicit in
option prices is much smaller. Even in the Friday right before the crash, there is no sign
of risks of an immediate collapse of stock prices.

7.2.4 Guimaraes (2007)

Time for me to talk about my own work!

Guimaraes (2007) presents procedure for testing whether currency crises depend on
sunspots or are triggered when the overvaluation hits a threshold.
The idea is the following: if crises are triggered by currency overvaluation crossing a
threshold, the expected magnitude of a devaluation, conditional on its occurrence, is equal
to the threshold value, which may di¤er substantially from the unconditional expected

81
currency overvaluation. On the other hand, if crises are triggered by sunspots, uncorre-
lated with the economic variables that determine the exchange rate in a ‡oating regime,
then the expected magnitude of a devaluation conditional on its occurrence is similar to
the unconditional expected currency overvaluation.
The probability and the expected magnitude of a devaluation are not observable but
can be estimated using data on exchange rate options. Guimaraes(2007) identi…es the
probability and expected magnitude of a devaluation of Brazilian Real in the period lead-
ing up to the end of the Brazilian pegged exchange rate regime and contrasts the estimates
to the predictions from a simple model of currency crises under di¤erent assumptions
about the trigger.

The model As in Bates (1991), a parametric approach is employed. The asset pricing
model is the following:
Denote by S the exchange rate and s its logarithm. Initially, the exchange rate follows
a standard Brownian motion with low volatility:

ds = 1 dt + 1 dX

The pegged regime may be abandoned at any time. The interruption is a Poisson event
with hazard rate . It leads to a discrete jump in the exchange rate and to a new di¤usion
process, assumed to last forever.
The jump is constant (k):
S af ter
= (1 + k)
S bef ore
The ‡oating regime is described by a Brownian motion with drift and higher volatility:

ds = 2 dt + 2 dX

Results The empirical results unveil completely di¤erent patterns for the probability and
expected magnitude of a devaluation (conditional on its occurrence). The probability was
volatile and mostly driven by contagion from external crises, as the Asian and Russian
crises triggered by far the greatest increases in the probability that the peg would be
abandoned. In contrast, the expected magnitude was stable and entirely una¤ected by
the Russian episode.
In addition, these data suggest that the Asian and Russian crises negatively impacted
the Brazilian shadow exchange rate. They explicitly show that the crises coincided with
both the greatest increases in the risk of a devaluation in Brazil and the largest depreci-
ations of other Latin American currencies, like the Mexican Peso. Since the crises were

82
fairly exogenous to the Latin American economies, it is natural to assume that if the
Brazilian currency were allowed to ‡oat, it would also have depreciated.

Conclusion The empirical …ndings favour thresholds and learning over sunspots.

7.2.5 Parametric non-parametric method

Some advantages of the parametric method:

it imposes a structure that makes sense economically.

interpretation of the results is immediate.

in the non-parametric case, you need to impose some structure anyway (quadratic,
interpolation).

it is good if there is not much data (market is not so liquid), and if the model is
appropriate.

Some disadvantages of parametric method:

it imposes a structure that may di¤er from the true data generating process.

the parametric model is not exactly what you end up estimating (as you vary ’s
and ’s).

the non-parametric approach may yield accurate results if we have very good data
(or if we do not have a good model).

7.2.6 Extensions

Jondeau and Rockinger (2000) describe alternative methods to infer risk-neutral densities.
Their paper brings a collection of techniques. It is worth discussing a few examples.

Stochastic volatility The di¤usion process is the following:

dSt p
= :dt + t::dW1
St
p
d t = ( t )dt + : t :dW2

Parameters:

: long-run volatility;

83
 : mean-reversion speed;

: volatility of the volatility.

The instantaneous volatility t is not exactly a parameter, is the realization of a random

variable, but it is natural to estimate it as well.

Sum of log-normals Another way to imposing some structure in the probability density
functions is to consider that the option price is a mixture of M log-normal densities. That
is: Z 1
rT
P
M
Ct = e i (S K):l(S ; i; i ):dS
i=1 S =K

That gives us a formula not-much more complicated than the Black-Scholes formula
and we can estimate the implicit parameters.
Now, given that we want to impose some structure, what is good about this one? Or,
which kind of model could yield some distribution similar to this?

Stochastic interest rate Bakshi, Cao and Chen (1997) present a model with stochastic
volatitly, jumps (hazard rate is given by Poisson distribution and the size of the jump is
logn-normal) and stochastic interest rates. They estimate the implicit parameters of the
model and check what matters.

84
8 Identi…cation through heteroskedasticity

Rigobon and Sack (2004) use identi…cation through heteroskedasticity to estimate the
impacts of monetary policy on asset prices.5 We have to deal with two main problems:

1. endogeneity: the policy decision may be in‡uenced by simultaneous changes in asset

prices.

2. omitted variables: a number of other variables may have an impact on the policy
decision and on asset prices.

The following system of equations captures these features.

st = it + zt + t (38)
it = s t + z t + "t (39)

Where it is the change in the 1-year interest rate, st is the change in the spot
exchange rate, zt is an omitted variable, "t is a monetary policy shock and t is a shock
to the asset price.
If we run OLS on equation 38, we get the following expression for the expected ^ :
+( + ) z
E(^ ) = + (1 ) 2
" + + ( + )2 z

which means that the estimator for is biased unless "= z ! 1 and "= ! 1.
Now, lets’s assume that zt , "t and t have no serial correlation and are uncorrelated
with each other. Let’s divide the sample, consisting on values of s and i, in two
subsamples: the subset F corresponds to the dates where the FOMC meets, and the
subset N corresponds to dates with no meeting. The number of observations in each
of the two sets is denoted by TF and TN , respectively. The key assumption in Rigobon
and Sack (2004) is that the variance of the shock to the interest rate ("t ) in the dates
belonging to the set F is higher than the variance of the shock to the interest rate in the
dates belonging to the set N , while the variances of t and zt are the same:
F N
" > "
F N
=
N
F = z

The traditional way of analyzing the impact of monetary policy decisions (the so-called
event study approach) is to consider that unexpected changes in the policy rate are exoge-
nous and use those to estimate equation 38 (which basically means assuming "= z !1
5 See also Rigobon (2003).

85
and "= ! 1). Rigobon and Sack (2004) show that such strong assumptions are not
needed: with the above assumptions on heteroskedasticity, we can consistently estimate
. The intuition is the following: in dates where the FOMC meets, there is a shock
to equation 39, " increases, but there are no shocks to other variables. So, the overall
relation between s and i should be di¤erent between the two subsamples, F and N .
The …gures at page 1557-1559 illustrate this idea.
Solving for the reduced form of equations 38 and 39, we reach:
1
it = [( + ) zt + t + "t ] (40)
1
1
st = [(1 + ) zt + t + "t ] (41)
1
De…ne xt = [ it ; st ]0 . The covariance matrices in each subsample are F =
E ( xt : x0t jt 2 F ), and N = E ( xt : x0t jt 2 N ) : De…ning F N; we get:
" #
F N
" " 1
= (42)
(1 )2 2

F N
In words, because " 6= " ; the covariance matrix of s and i is not the same in
subsets F and N and therefore it is possible to identify by looking at the di¤erence in
the covariance matrix in the two di¤erent subsamples.
Now, consider the following variables:
0
i0 i0N
I pF ; p
TF TN
0
s0 s0N
S p F ; p
TF TN
0
i0 i0
wi pF ; p N
TF TN
0
s0 s0N
ws p F ; p
TF TN
A major result in Rigobon and Sack (2004) is that can be consistently estimated
by a standard instrumental variables approach with those novel instruments, wi and
ws . Why are wi and ws valid instruments for I at equation 38? First, the instrument
wi is correlated with I because the bigger variance in the Copom dates implies that
p p
the positive correlation between i0F = TF and i0F = TF more than outweights the

86
 p p
negative one between i0N = TN and i0N = TN . Formally:
1 1 1
plim wi0 I = i0F : iF i0N : iN
T TF TN
1 F 1 N
= 2 " 2 " >0
(1 ) (1 )
On the other hand, wi is not correlated with neither z nor because the positive and
negative correlation of each part of the vector exactly cancel each other out:
1 1 1
plim wi0 z = i0F :zF i0 :zN
T TF TN N
+ F + N
= z z = 0
1 1
1 1 1
plim wi0 = i0F : F i0 :
T TF TN N N
F N
= =0
1 1

The paper also shows how we could implement this methodology through GMM
Then, it shows how we could test the following null hypotheses:

1. The assumed information structure is correct. For this, note that equation 42 yields
F N
3 moment conditions and 2 parameters ( and (1
" "
) 2 ).

2. The event study is unbiased (this is a Hausman test).

Results are presented on pages 1567-1572.

8.1 Evidence on default costs

Hebert and Schreger (2017) employ the method of identi…cation through heteroskedastic-
ity to estimate the e¤ects of changes in the probability of default on the value of assets
in Argentina.
The source of exogenous variation is the legal battle between creditor holdouts and
Argentina in US courts, described in details in Hebert and Schreger (2017). It all started
in December 2001 when Argentina defaulted on over $100 billion in external sovereign
debt. The exchange rate was devalued by around 75 percent. In January 2005, Argentina
presented an o¤er to creditors that was accepted by holders of $62 billion of debt. S&P
declared the end of Argentina default in June. Argentina fully repaid the IMF in 2006.
In December 2010, Argentina o¤ered another bond exchange. Many creditors agreed to

87
the exchange. Remaining holdouts were owed around $11 billion. Argentina had then
structured 90% of its debt.
Creditors led by NML started a legal battle based on the pari passu clause (equal
treatment). There were numerous rulings between December 2011 to July 2014 in the
Court for the Southern District of New York, the Court of Appeals for the Second Circuit
and the Supreme Court. These rulings are shocks to the probability that Argentina will
default on its debt, but are pretty exogenous to the Argentinian economy. The method
of identi…cation through heteroskedasticity …ts very well here. With the dates of the legal
rulings issued by those 3 courts, we can estimate the e¤ect of exogenous shocks to the
default probability on asset prices in Argentina.
They …nd large e¤ects. A 10 percent increase in the probability of default causes a
6 percent decline in the value of Argentine equities and a 1 percent depreciation of a
measure of the exchange rate.

8.2 US interest rates and default risk in emerging economies

Increases in US real interest rates raise the opportunity cost of money for investors and
thus lead to higher interest rates for emerging economies. Simple intuition and the models
in Section 4 imply that this increase in interest rates raises the probability of default, which
leads to further increases in the cost of debt. However, identifying this e¤ect empirically
is not a trivial task because any changes in the world economy a¤ect both variables. For
example, negative shocks to economic activity in the world tend to lead to lower interest
rates in the US and higher odds of default.
Foley-Fisher and Guimaraes (2013) investigate the impact of changes in US real interest
rates on sovereign default risk in emerging economies (measure by the EMBI+ premium)
using the method of identi…cation through heteroskedasticity. The identi…cation hypoth-
esis is that on dates that the FOMC meets, there is an extra shock to US interest rates –
an assumption similar to the one employed in Rigobon and Sack (2004).
Indeed, the raw correlation between real interest rates and the EMBI+ premium is
negative: a 2 basis point increase in the 10-year US real interest rate is on average related
to a 1 basis point decrease in the EMBI+. Intuitively, high real rates typically re‡ect
favorable external conditions for emerging markets, which reduce the risk of default.
The causal e¤ect, however, is very di¤erent. A 1 basis point increase in 10-year US
interest rates raises EMBI+ premia by around 1 basis point.

88
8.3 The e¤ects of commodity price shocks

Many papers study how shocks to commodity prices a¤ect a bunch of economic and
…nancial variables. However, most studies simply assume that commodity prices are
exogenous and the primary source of external shocks. For a small open economy, this
seems to be a natural assumption. However, many global factors are likely to a¤ect
commodity prices and the small open economy, blurring the identi…cation. For example,
a stronger world economy will lead to higher commodity prices but also more capital ‡ows
to emerging countries and less default risk, for example.
Melo (2024) employs the method of identi…cation through heteroskedasticity to es-
timate the e¤ect of commodity price shocks on default risk (from CDS), stock market
indices and exchange rates. He exploits a novel source of price heteroscedasticity, the
USDA Grain Stocks report. The report provides estimates of the domestic volume of
grain stored in on-farm and o¤-farm storage facilities. The information in the report
induces large commodity price changes (the …rst stage works) and is plausibly unrelated
to any relevant confounder (the exclusion restriction is respected).
The estimated coe¢ cients are signi…cantly lower than OLS estimates –typically used
in the literature. Some are close to zero. For example, higher grain prices decrease risk
premiums in emerging markets, but the e¤ect is perhaps a third of OLS estimates. In
the US, increases in grain prices positively in‡uence stock markets, while oil price shocks
increase in‡ation.

89
References

[1] Abreu, Dilip and Brunnermeier, Markus K, 2002, “Synchronization risk and delayed
arbitrage”, Journal of Financial Economics 66, 341-360.

[2] Arellano, Cristina, 2008, “Default risk and income ‡uctuations in emerging
economies”, American Economic Review 98, 690-712.

[3] Bakshi, Gurdip; Cao, Charles and Chen, Zhivu, 1997, “Empirical performance of
alternative option pricing models”, Journal of Finance 52, 2003-2049.

[4] Bates, David S., 1991, “The crash of ’87: Was it expected? The Evidence from
Options Markets”, Journal of Finance 46, 1009-1044.

[5] Bikhchandani, Sushil; Hirshleifer, David and Welch, Ivo, 1992, “ A theory of fads,
fashion, custon, and cultural change as informational cascades”, Journal of Political
Economy 100, 992-1026.

[6] Breeden, Douglas and Litzenberger, Robert, 1978, “Prices of State-Contingent Claims
Implicit in Option Prices”, Journal of Business 51, 621-651.

[7] Campa, José M.; Chang, P. H. Kevin and Refalo, James F., 2002, “An options-based
analysis of emerging market exchange rate expectations: Brazil’s Real plan, 1994 -
1997”, Journal of Development Economics 69, 227-253.

[8] Caplin, Andrew and Leahy, John, 1994, “Business as usual, market crashes and
wisdom after the fact”, American Economic Review 84, 548-565.

[9] Carlsson, Hans; and van Damme, Eric, 1993, “Global games and equilibrium selec-
tion”, Econometrica 61, 989-1018.

[10] Chang, Roberto and Velasco, Andres, 1998, “Financial crises in emerging markets”,
NBER Working Paper 6606.

[11] Collard, Fabrice; Habib, Michel and Rochet, Jean-Charles, 2015, “Sovereign debt sus-
tainability in advanced economies”, Journal of the European Economic Association
13, 381-420.

[12] Corsetti, Giancarlo; Pesenti, Paolo and Roubini, Nouriel, 1999, “Paper tigers? A
model of the Asian crisis”, European Economic Review 43, 1211-1236.

90
[13] Corsetti, Giancarlo; Guimaraes, Bernardo and Roubini, Nouriel, 2006, “International
Lending of Last Resort and Moral Hazard: A Model of IMF’s Catalytic Finance”,
Journal of Monetary Economics 53, 441-471.

[14] Cuadra, Gabriel; Sanchez, Juan M. and Sapriza, Horacio, 2010, “Fiscal policy and
default risk in emerging markets”, Review of Economic Dynamics 13, 452-469.

[15] Diamond, Douglas, 2007, “Banks and liquidity creation: a simple exposition of the
Diamond-Dybvig model”, Economic Quarterly 93, 189-200.

[16] Diamond, Douglas and Dybvig, Philip, 1983, “Bank runs, deposit insurance and
liquidity”, Journal of Political Economy 91, 401-419.

[17] Diaz-Alejandro, Carlos, 1985, “Good-bye …nancial repression, hello …nancial crash”,
Journal of Development Economics 19, 1-24.

[18] Doblas-Madrid, Antonio, 2012, “A robust model of bubbles with multidimensional

uncertainty”, Econometrica 80, 1845-1893.

[19] Eaton, Jonathan and Gersovitz, Mark, 1981, “Debt with potential repudiation: the-
oretical and empirical analysis”, Review of Economic Studies 48, 289-309.

[20] Flood, Robert and Garber, Peter, 1984, “Collapsing exchange-rate regimes: some
linear examples”, Journal of International Economics 17, 1-13.

[21] Foley-Fisher, Nathan and Guimaraes, Bernardo, 2013, “US real interest rates and
default risk in emerging economies”, Journal of Money, Credit and Banking 45, 967-
975.

[22] Geanakoplos, John, 1992, “Common Knowledge”, Journal of Economic Perspectives

6, 53-82.

[23] Goldstein, Itay and Pauzner, Ady, 2004, “Contagion of self-ful…lling …nancial crises
due to diversi…cation of investment portfolios”, Journal of Economic Theory 119,
151–183.

[24] Goldstein, Itay and Pauzner, Ady, 2005, “Demand–Deposit Contracts and the Prob-
ability of Bank Runs”, Journal of Finance 60, 1293–1327.

[25] Gonçalves, Carlos and Guimaraes, Bernardo, 2015, “Sovereign default risk and com-
mitment for …scal adjustment”, Journal of International Economics 95, 68-82.

91
[26] Guimaraes, Bernardo, 2007, “Vulnerability of currency pegs: evidence from Brazil”,
mimeo.

[27] Guimaraes, Bernardo, 2011, “Sovereign default: which shocks matter?”, Review of
Economic Dynamics, 14, 553-576.

[28] Guimaraes, Bernardo and Morris, Stephen, 2007, “Risk and wealth in a model of
self-ful…lling currency attacks”, Journal of Monetary Economics 54, 2205-2230.

[29] Harrison, J. Michael and Kreps, David, 1978, “Speculative investor behavior in a
stock market with heterogeneous expectations”, Quarterly Journal of Economics 92,
323-336.

[30] Hebert, Benjamin, and Jesse Schreger, 2017, “The costs of sovereign default: Evi-
dence from argentina”, American Economic Review 107, 3119-45.

[31] Jeanne, Olivier, 1997, “Are currency crises self-ful…lling? A test”, Journal of Inter-
national Economics 43, 263-286.

[32] Jondeau, Eric and Rockinger, Michael, 2000, “Reading the smile: the message con-
veyed by methods which infer risk neutral densities”, Journal of International Money
and Finance 19, 885-915.

[33] Krugman, Paul, 1979, “A Model of Balance-of-Payments Crises”, Journal of Money,

Credit and Banking 11, 311-325.

[34] Melo, Alisson, 2024, “Essays on the impacts of commodity price ‡uctuations”, Tese
de Doutorado, Sao Paulo School of Economics –FGV.

[35] Merton, Robert C., 1976, “Option pricing when underlying stock returns are discon-
tinuous”, Journal of Financial Economics 3, 125-144.

[36] Morris, Stephen and Shin, Hyun, 1998, “Unique equilibrium in a model of self-
ful…lling currency attacks”, American Economic Review 88, 587-597.

[37] Morris, Stephen and Shin, Hyun, 2003, Global games: theory and applications, in
Advances in Economics and Econometrics (Proceedings of the Eighth World Congress
of the Econometric Society), edited by M. Dewatripont, L. Hansen and S. Turnovsky;
Cambridge University Press.

[38] Morris, Stephen and Shin, Hyun, 2004, “Coordination risk and the price of debt”,
European Economic Review 48, 133-153.

92
[39] Obstfeld, Maurice, 1996, “Models of currency crises with self-ful…lling features”, Eu-
ropean Economic Review 40, 1037-1047.

[40] Radelet, Steven and Sachs, Je¤rey, 1998,.“The East Asian Financial Crisis: Diagno-
sis, Remedies, Prospects”, Brookings Papers on Economic Activity, 1998 #1, 1-90.

[41] Rigobon, Roberto, 2003, “Identi…cation through heteroskedasticity”, Review of Eco-

nomics and Statistics 85, 777-792.

[42] Rigobon, Roberto, and Brian Sack, 2004. “The impact of monetary policy on asset
prices”, Journal of Monetary Economics 51, 1553-1575.

[43] Rochet, Jean-Charles, and Xavier Vives, 2004. “Coordination failures and the lender
of last resort: was Bagehot right after all?” Journal of the European Economic As-
sociation 2, 1116-1147.

[44] Rose, Andrew and Svensson, Lars, 1994, “European exchange rate credibility before
the fall”, European Economic Review 38, 1185-1216.

[45] Salant, Stephen and Henderson, Dale, 1978, “Market antecipation of government
policy and the price of gold”, Journal of Political Economy 86, 627-648.

93