Exchange Rates in a Behavioural Finance

Framework
Marianna Grimaldi & Paul De Grauwe
Sveriges Riksbank & University of Leuven
November 28, 2003

Abstract
We develop a simple model of the foreign exchange market in which
agents optimize their portfolio and use different forecasting rules. They
check the profitability of these rules ex post and select the more profitable
one.This model produces two kinds of equilibria, a fundamental and a
bubble one. In a stochastic environment the model generates a complex
dynamics in which bubbles and crashes occur at unpredictable moments.
We also analyse the empirical relevance of the model

1
1 Introduction
The rational expectations paradigm has been important in our understanding
of financial makets. However, it has also shown its limitations. This is especially
the case in the foreign exchange market where much of the observed short-term
dynamics cannot easily be understood in the context of rational expectations
models. As will be argued in this paper the difficulties of rational expecta-
tions models (RE-models) in correctly describing the short-term dynamics of
the exchange market is related to the strong informational burden imposed on
individual agents by the these models.
In this paper we use an approach which is influenced by the behavioural
finance literature. In this literature agents use relatively simple behavioural
rules. These rules do not employ all available information, mainly because
agents find it difficult to process and to evaluate the wealth and complexity
of the available information (Shleifer(2000), Shiller(2001)). Agents are rational,
however, in that they use these rules only as long as they are profitable. We
will add to this the notion that agents compare the rule they currently use to
alternative rules and decide to switch to the alternative if it turns out that the
it is more profitable. This is tantanamount to introducing a fitness criterion
in the selection of behavioural rules. In this sense the model is also in the
tradition of the literature on evolutionary dynamics (Kirman(1993), Brock and
Hommes(1997), Lux(1998)).
The paper proceeds in different steps. We start by presenting the simplest
possible version of the model. In this model there is no feedback from the goods
market and there are no transactions costs. We show how this leads to endemic
bubbles and crashes. In the next step we introduce a feedback from the goods
market, first by endogenizing the current account, and then by introducing the
assumption that it is costly to trade goods and services internationally. It will be
shown that these complications alter substantially the nature of the dynamics
in the foreign exchange markets. Finally, we discuss the empirical implications
of the model.

2 The model
In this section we develop the simple version of the exchange rate model. The
model consists of three building blocks. First, agents select their optimal port-
folio using a mean-variance utility framework. Here we follow the tradition of
mainstream economics in that agents are utility maximizing individuals. Second,
these agents make forecasts about the future exchange rate based on simple but
different rules. In this second building block we introduce concepts borrowed
from the behavioural finance literature. Third, agents evaluate these rules ex-
post by comparing their risk-adjusted profitability. Thus, the third building
block relies on an evolutionary dynamics.

2
2.1 The optimal portfolio
We assume agents of different types i depending on their beliefs about the future
exchange rate. Each agent can invest in two assets, a domestic and a foreign
one. The agents’ utility function can be represented by the following equation:

i i 1
U (Wt+1 ) = Et (Wt+1 ) − µV i (Wt+1
i
) (1)
2
i
where Wt+1 is the wealth of agent of type i at time t+1, Et .is the expecta-
tion operator, µ is the coefficient of risk aversion and V i (Wt+1 i
) represents the
conditional variance of the wealth of agent i. The wealth is specified as follows:
i
¡ ¢
Wt+1 = (1 + r∗ ) st+1 dit + 1 + r Wti − st dit (2)

where r and r∗ are respectively the domestic and the foreign interest rates, st+1
is the exchange rate at time t + 1, di,t represents the holdings of the foreign
assets by agent of type i at time t. Thus, the first term on the right-hand side
of 2 represents the value of the foreign portfolio in domestic currency at time
t + 1, while the second term represents the value of the domestic portfolio at
time t + 1.
Substituting equation 2 in 1 and maximising the utility with respect to di,t
allows us to derive the standard optimal holding of foreign assets by agents of
type i :

(1 + r∗ ) Eti (st+1 ) − (1 + r) st
di,t = (3)
µσ 2i,t
The optimal holding of the foreign asset depends on the expected excess return
corrected for risk. The market demand for foreign assets at time t is the sum of
the individual demands, i.e.:
N
X
ni,t di,t = Dt (4)
i=1

where ni,t is the number of agents of type i.

Market equilibrium implies that the market demand is equal to the market
supply Zt which we assume to be exogenous1 . Thus,

Zt = Dt (5)
Substituting the optimal holdings into the market demand and then into the
market equilibrium equation and solving for the exchange rate st yields the
equilibrium exchange rate:
1 The market supply is determined by the net current account and by the sales or purchases

of foreign exchange of the central bank. We assume both to be exogenous here. In section
we will endogenize the current account.

3
 µ ¶ "N #
1 + r∗ 1 X E i (st+1 ) Zt
st = wi,t t 2 −µ (6)
1+r P
N
wi,t σ i,t 1+r
i=1
σ 2i,t
i=1
where wi,t . = ni,t /N is the weight (share) of agent i.
Thus the exchange rate is determined by the expectations of the agents, Eti ,
about the future exchange rate. These forecasts are weighted by their respective
variances σ 2i,t . When agent’s i forecasts have a high variance the weight of this
agent in the determination of the market exchange rate is reduced.

2.2 The forecasting rules

We now specify how agents.form their expectations of the future exchange rate
and how they evaluate the risk of their portfolio.
We start with an analysis of the rules agents use in forecasting the exchange
rate. We take the view that individual agents are overwhelmed by the complex-
ity of the informational environment, and therefore use simple rules to make
forecasts. Here we describe these rules. In the next section we discuss how
agents select the rules.
We assume that two types of forecasting rules are used. One is called a
”fundamentalist” rule, the other a ”chartist” (technical trading) rule2 . The
agents using a fundamentalist rule, the ”fundamentalists”, base their forecast
on a comparison between the market and the fundamental exchange rate, i.e.
they forecast the market rate to return to the fundamental rate in the future.
In this sense they use a negative feedback rule that introduces a mean reverting
dynamics in the exchange rate. The speed with which the market exchange
rate returns to the fundamental is assumed to be determined by the speed of
adjustment in the goods market which is assumed to be in the information set
of the fundamentalists (together with the fundamental exchange rate itself).
Thus, the forecasting rule for the fundamentalists is :
¡ ¢
Etf (∆st+1 ) = −ψ st−1 − s∗t−1 (7)
where s∗t is the fundamental exchange rate at time t , which is assumed to follow
a random walk and 0 < ψ < ∞. The fundamental exchange rate is the value of
the exchange rate that equilibrates the current account. At this stage, however,
we do not model the goods market. As a result, the fundamental exchange rate
is exogenous. We return to this issue in a later section.
The agents using technical analysis, the ”chartists”, forecast the future ex-
change rate by extrapolating past exchange rate movements. Their forecasting
rule can be specified as :
T
X
Etc (∆st+1 ) = β αi ∆st−i (8)
i=1
2 This idea of distinguishing between fundamentalist and chartist rules was first introduced

by Frankel and Froot().

4
Thus, the chartists compute a moving average of the past exchange rate changes
and they extrapolate this into the future exchange rate change. The degree of
extrapolation is given by the parameter β. Note that chartists take into account
information concerning the fundamental exchange rate indirectly, i.e. through
the exchange rate itself. In addition, chartist rules can be interpreted as rules
that attempt to detect ”market sentiments”. In this sense the chartist rule can
be seen as reflecting herding behaviour3 .
It should be stressed that both types of agents, chartists and fundamentalists,
use partial information. Thus our approach differs from the approach in the
tradition of rational expectations models in which it is assumed an asymmetry
in the information processing capacity of agents. In the latter approach some
agents, the ”rational” agents, are assumed to use all available information, while
other agents, ”noise traders”, do not use all available information. Such an
asymmetry it is interesting in order to facilitate the mathematical analysis of
the models. However, the basis on which such an asymmetry can be invoked
remains unclear. In contrast with this tradition, we take the view that the
informational complexity is similar for all agents, and that none of them can be
considered to be superior on that count.
We now analyse how fundamentalists and chartists evaluate the risk of their
portfolio. The risk is measured by the variance terms in equation 6, which we
define as the weighted average of the squared (one period ahead) forecasting
errors made by chartists and fundamentalists, respectively. Thus,
∞
X £ i ¤2
σ i,t = γ k Et−k (st−k+1 ) − st−k+1 (9)
k=1

where γ k are geometrically declining weights, and i = f, c.

2.3 Fitness of the rules

The next step in our analysis is to specify how agents evaluate the fitness of
these two forecasting rules. The general idea that we will follow is that agents
use one of the two rules, compare their (risk adjusted) profitability ex post and
then decide whether to keep the rule or switch to the other one. Thus, our
model is in the logic of evolutionary dynamics, in which simple decision rules
are selected. These rules will continue to be followed if they pass some ”fitness”
test (profitability test). Another way to interpret this is as follows. When great
uncertainty exists about how the complex world functions, agents use a trial
and error strategy. They try a particular forecasting rule until they find out
that other rules work better. Such a trial and error strategy can be considered
to be a rational strategy when agents cannot understand the full complexity of
the underlying model.
3 There is a large literature on the use of chartist analysis (technical analysis). This literature

makes clear that chartism is widely used in the foreign exchange markets. See Cheung and
Chinn(1989), Taylor and Allen(1992), Cheung et al(1999), Mentkhoff(1997) and (1998).

5
 In order to implement this idea we use an approach proposed by Brock and
Hommes(1997) which consists in making the weights of the forecasting rules a
function of the relative profitability of these rules, i.e. 4 :
£ ¤
exp γπ0c,t−1
wc,t = £ ¤ h i (10)
exp γπ 0c,t−1 + exp γπ 0f,t−1
h i
exp γπ0f,t−1
wf,t = £ ¤ h i (11)
exp γπ0c,t−1 + exp γπ0f,t−1

where π 0c,t−1 and π0f,t−1 are the risk adjusted net profits made by chartists’
and fundamentalists’ forecasting the exchange rate in period t − 1, i.e. π 0c,t−1 =
π c,t−1 − µσ 2c,t−1 and π0f,t−1 = π f,t−1 − C − µσ 2f,t−1 . We assume that the funda-
mentalists make a fixed cost C for the collection and processing of fundamental
information, while the collection of information by chartists is assumed to be
costless5 .
Equations 10and 11 can be interpreted as follows. When the risk adjusted
profits of the chartists’ rule increases relative to the risk adjusted net profits
of the fundamentalists rule, then the share of agents who use chartist rules
increases in period t increases, and vice versa. This parameter γ measures the
rate with which the chartists and fundamentalists revise their forecasting rules.
With an increasing γ agents revise their forecasts very frequently. In the limit
when γ goes to infinity agents revise the forecasting rules instantaneously. When
γ is low, chartists and fundamentalists revise their forecasts relatively slowly.
When γ is equal to zero they do not revise their rules. In the latter case the
fraction of chartists and fundamentalists is constant and equal to 0.5. Thus, γ
is a measure of inertia in the decision to switch to the more profitable rule. As
will be seen, this parameter is of great importance in generating bubbles.
Chartists and fundamentalists make a profit when they correctly forecast
the direction of the exchange rate movement. They make a loss if they wrongly
predict the direction of its movements. The profit (the loss) they make equals
the one-period return of investing $1.

3 Solution of the model

In this section we investigate the properties of the solution of the model. We first
study the deterministic solution of the model. This will allow us to analyse the
characteristics of the solution that are not clouded by exogenous noise. Then,
we analyse the stochastic version of the model, i.e. we introduce stochastics in
4 This specification of the decision rule is often used in discrete choice models. For an ap-

plication in the market for differentiated products see Anderson, de Palma, and Thisse(1992).
The idea has also been applied in financial markets, by Brock and Hommes (1997) and by
Lux(1998).
5 This asymmetry in the treatment of the cost of information for fundamentalists and

chartists is not crucial for our results.

6
the exogeneous variables. We use simulation techniques since the non-linearities
do not allow for a simple analytical solution. We select ”reasonable” values of
the parameters, i.e. those that come close to empirically observed values. As we
will show later these are also parameter values for which the model replicates
the observed statistical properties of exchange rate movements. We will also
subject such results to an extensive sensitivity analysis.
We start with an analysis of the deterministic model. In figure 1 we show
the solutions of the exchange rate for different initial conditions. These are
fixed-point solutions (attractors). We plot such solutions as a function of the
different initial conditions6 . On the horizontal axis we set out the different initial
conditions. These are initial shocks to the deterministic system. The vertical
axis shows the solutions corresponding to these different initial conditions. The
fundamental exchange rate was normalized to 0. We find two types of fixed
point solutions. First, for small disturbances in the initial conditions the fixed
point solutions coincide with the fundamental exchange rate. We call these
solutions the fundamental solutions. Second, for large disturbances in the initial
conditions, the fixed point solutions diverge from the fundamental. We will call
these attractors, bubble attractors. It will become clear why we label these
attractors in this way. The larger is the initial shock (the noise) the farther
the fixed points are removed from the fundamental exchange rate. The border
between these two types of fixed points is characterised by discontinuities. This
has the implication that in the neigborhood of the border a small change in the
initial condition (the noise) can have a large effect on the solution.
The different nature of these two types of fixed point attractors can also be
seen from an analysis of the chartists’ weights that correspond to these different
fixed point attractors. We show these chartists weight as a function of the initial
conditions in figure ??.
We find, first, that for small initial disturbances the chartists’ weight con-
verges to 50% of the market. Thus when the exchange rate converges to the
fundamental rate, the weight of the chartists and the fundamentalists are equal
to 50%. For large initial disturbances, however, the chartists’ weight converges
to 1. Thus, when the chartists take over the whole market, the exchange rate
converges to a bubble attractor. The meaning of a bubble attractor can now be
understood better. It is an exchange rate equilibrium that is reached when the
number of fundamentalists has become sufficiently small (the number of chartists
has become sufficiently large) so as to eliminate the effect of the mean reversion
dynamics. It will be made clearer in the next section why fundamentalists drop
out of the market. Here it suffices to understand that such equilibria exist. It is
important to see that these bubble attractors are fixed point solutions. Once we
reach them, the exchange rate is constant. The chartists’ expectations are then
model consistent, i.e. chartists who extrapolate the past movements, forecast no
change. At the same time, since the fundamentalists have left the market, there
is no force acting to bring back the exchange rate to its fundamental value.Thus
6 These fixed point solutions of the exchange rate where obtained by running simulations

of 100,000 periods. Each time the exchange rate converged to a fixed point.

7
Figure 1:

Figure 2:

8
two types of equilibria exist: a fundamental equilibrium where chartists and
fundamentalists co-exist, and a bubble equilibrium where the chartists have
crowded out the fundamentalists7 . In both cases, the expectations of the agents
in the model are consistent with the model’s outcome.
These two types of equilibria differ in another respect. The fundamental
equilibrium can be reached from many different initial conditions. It is locally
stable, i.e after small disturbances the system returns to the same (fundamental)
attractor. In contrast there is one and only one initial condition that will lead
to a particular bubble equilibrium. This implies that a small disturbance leads
to a displacement of the bubble solution. Note again that the border between
these two types of equilibria is characterized by discontinuities and complexity,
i.e. small disturbances can lead to either a fundamental or a bubble equilibrium.
It is useful to compute the attractors for different values of the fundamental
exchange rate while keeping initial conditions constant. We show such an exer-
cise in figure 3. We now present different fundamental values of the exchange
rate on the horizontal axis while keeping the initial condition unchanged. We
have set the initial condition for the exchange rate equal to 48 . We obtain the
following results. First when the fundamental shock and the initial condition
are opposite in sign, the exchange rate converges to its fundamental value. This
can be seen by the fact that for negative values of the fundamental shocks, the
attractors are on a 45 line so that the equilibrium exchange rate equals its
fundamental value. In the range of fundamental shocks between 0 and 4 we ob-
tain bubble equilibria. This is the range in which the initial shock (noise) has
the same sign as the fundamental shock. When the positive fundamental shock
becomes large relative to the positive initial shock the system returns to a funda-
mental equilibrium. Thus, bubble equilibria arise when the fundamental shock
and the noise have the same sign, and when the noise is relatively large relative
to the fundamental shock. With sufficiently large fundamental shocks (relative
to the noise) the equilibrium exchange rate is forced back to its fundamental
value. In appendix 1 we show some additional simulations for smaller and larger
initial conditions. These simulations confirm that as the noise increases relative
to the fundamental shocks, the range of bubble equilibria increases and vice
versa.
The previous results allow us to understand not only why bubbles can arise.
They also shed light on why bubbles tend to crash. The noise that triggers
a bubble is temporary. Fundamental shocks, however, typically have a large
permanent component9 . Thus, in a stochastic environment small fundamental
shocks accumulate to large cumulative fundamental changes. These cumulative
changes in the fundamental exchange rate at some point become overwhelming
leading to a crash. We will return to this result when we present the stochastic
7 Note that the intermediate points, i.e. when chartists’ weight is less than 1 the solution

has not converged yet to fixed points. Fundamentalists hold a very small share in the market
which exerts some mean reverting force. However their influence is offset by the chartists
pressure. In figure xxx the simulation results are for T=100000.
8 Our results, however, are not affected qualitatively by the choice of this number.
9 In the simulations reported here a fundamental shock is permanent.

9
 Figure 3:

simulations of the model.

4 The anatomy of bubbles and crashes

In the previous section we identified the existence of two different types of fixed
point solutions, i.e. a fundamental solution characterised by the fact that the
exchange rate converges to its fundamental value while chartists and fundamen-
talists ”co-habitate”, and a bubble solution in which the exchange rate deviates
from its fundamental value and in which chartists dominate the market. In this
section we show that in combination with stochastic shocks in the fundamental
exchange rate these features of the model lead to the emergence of bubbles and
crashes.
The way we proceed is to calibrate the model in such a way that it replicates
the statistical properties of observed exchange rate movements. We describe this
procedure in section ??. Here we present the results of simulations in the time
domain using this calibrated model. We start by presenting a case study of a
typical bubble and crash scenario as produced by the stochastic version of the
model. In the next section we will analyse more systematically the factors that
determine the frequency with which such bubbles and crashes occur. Figure 4
top panel shows the exchange rate and its fundamental value in the time domain;
the bottom panel shows the weight of the chartists in the same time domain.
These two pictures allow us to analyse a number of common features of a typical
endogenously generated bubble and crashes in a stochastic environment.
First, once a bubble emerges, it sets in motion bandwagon effects. As the ex-
change rate moves steadily in one direction, the use of extrapolative forecasting
rules becomes more profitable, thereby attracting more chartists in the market.

10
 Figure 4:

This is clearly visible from a comparison of the bottom panel with the top panel
of figure 4. We observe that the upward movement in the exchange rate coin-
cides with an increase in the weight of chartists in the market. We have checked
this feature in many bubbles produced by the model. In appendix 2 we show
another example of a bubble, and we present the results of a causality test which
shows that the exchange rate leads the weight of chartists during a bubble and
the subsequent crash. Thus, typically a bubble starts after the exchange rate
has moved in one direction, thereby attracting extrapolating chartists which in
turn reinforces the exchange rate movement.
Second, a sustained upward (downward) movement of the exchange rate will
not develop into a full scale bubble if at some point the market does not get
sufficiently dominated by the chartists. As can be seen figure 4 at the height

11
of the bubble the chartists have almost 100% of the market. Put differently,
an essential characteristic of a bubble is that at some point almost nobody is
willing to take a contrarian fundamentalist view. The market is then dominated
by agents who extrapolate the bubble into the future. This raises the question
of why fundamentalists do not take an opposite position thereby preventing the
bubble from developing. After all, the larger the deviation of the exchange rate
from the fundamental the more the fundamentalists expect to make profit from
selling the foreign currency. Yet they do not, and massively leave the market-
place to the chartists. The reason why they do so, is that during the bubble phase
the profitability of chartism increases dramatically precisely because so many
chartists enter the market thereby pushing the exchange rate up and making
chartism more profitable. In addition, during the bubbles phase fundamentalists
make large forecasting errors, reducing their ”appetite” for using fundamental-
ists forecasting rules. As a result, investors who are continously acting against
the trend will make losses. There is therefore a self-fulfiling dynamics in the
profitability of chartism and losses for the fundamentalists.
The limit of this dynamics is reached when chartists have crowded out the
fundamentalists. We arrive at our next characteristics of the bubble-crash
dynamics. When the chartists’ share is close to 100% the self-reinforcing up-
ward movement in the exchange rate and in profitability slows down, increasing
the expected relative profitability of fundamentalists. This is so because while
the bubble developed, the expected profits from fundamentalism also increased.
However, these were overwhelmed by the self-fulfilling profitability of chartism.
When the latter tends to slow down, fundamentalism becomes attractive again.
A small movement of the exchange rate can then trigger a fast decline in the
share of chartism, back to its normal level of a tranquil market. A crash is set
in motion.. .
The dynamics of bubbles and crashes we obtain in our simulated data is
asymmetric, i.e. bubbles are relatively slow and crashes relatively rapid. An
intuitive explanation of this result is that during a bubble chartists and fun-
damentalists rules push the exchange rate in two different directions, i.e. the
positive feedback from chartists and the negative feedback from fundamental-
ists have the effect of slowing down the build-up of a bubble. In a crash the
fundamentalists’ mean reverting force is reinforced by the chartists’ behaviour.
As a consequence, the speed of a crash is higher than the speed with which a
bubble arises.
This asymmetry between bubbles and crashes is a well-known empirical phe-
nomenon in financial markets (see Sornette(2003)). In figure 5 we present the
DEM-USD for the period 1980-1987, which is a remarkable example of a bubble
in foreign exchange markets. As it can be seen from figure 5 the upward move-
ment in the DEM-USD exchange rate is gradual and builds up momentum until
a sudden and much faster crash occurs which brings the exchange rate back to
its value of tranquil periods.
Our model provides a simple explanation for this empirical phenomenon.
Note the contrast with RE-models of bubbles and crashes. These predict that

12
 DEM-USD 1980-87

3.3

2.8

2.3

1.8

1.3
1980

1981

1982

1983

1984

1985

1986

1987
Figure 5:

bubbles and crashes are symmetric (Blanchard(1979) and Blanchard&Watson(1982))10 .

5 The frequency of bubbles

In the previous sections we showed that a very simple model is capable of gen-
erating bubbles and crashes that have the basic features of bubbles and crashes
observed in financial markets. All we need is the existence of agents who max-
imize the utility of their portfolio, make forecasts based on the use of different
forecasting rules and switch to the more profitable of these rules. An impor-
tant issue here concerns the frequency with which bubbles occur in our model.
We analyse this issue by simulating the stochastic version of the model and by
counting the number of periods the exchange rate is involved in a bubble. We
define a bubble here to be a deviation of the exchange rate from its fundamen-
tal value by more than three times the standard deviation of the fundamental
variable for a significant interval of time. We have set this interval equal to 20
periods.We show the result of such an exercise in figure 6 for different values
of the chartists’ extrapolation parameter β. It shows the percentage of time
1 0 Moreover, the symmetry of bubbles and crashes neglects the time scale dynamics in which

a long term change is an accumulation of short term changes. Thus, the symmetry property
in foreign exchange markets is an approximation which holds only in the (very) short-run (see
Johansen and Sornette (1999)).

13
the exchange rate is involved in a bubble dynamics. We observe that when β is
smaller than 0.9 the frequency of the occurrence of bubbles is small. For values
of β larger than 0.9 this frequency increases exponentially. Thus the extrapola-
tion by chartists is an important parameter affecting the frequency with which
bubbles occur.The results obtained in figure 6 are determined by the existence

Figure 6:

of bubble equilibria in the deterministic version of the model. Therefore, it is

useful to connect figure 6 with a figure that plots the exchange rate solutions ob-
tained in the deterministic version of the model. We show this in figure 7 where
we set out the equilibrium exchange rate on the vertical axis as a function of β
(horizontal axis). We see that for values of β < 0.85, the exchange rate converges
to its fundamental value (normalized to 0). When β > 0.85 we obtain bubble
equilibria that increasingly deviate from the fundamental value. Note that when
0.88 < β < 0.9 we have a complex structure. The equilibrium jumps back and
forth between the fundamental and a bubble. Thus, in a way figure 7 predicts
what should happen in a stochastic environment. When β < 0.85 the equilib-
rium exchange rate converges to its fundamenal value. Around this fundamental
value a basin of attraction exists which pulls the exchange rate. Only when the
noise is sufficiently high will the exchange rate be attracted to a bubble equilib-
rium (see figure 1 where we showed that with β = 0.8 a sufficiently high initial
shock will pull the exchange rate towards a bubble equilibium). Thus, when
β < 0.85 bubbles will be relatively infrequent events. When β increases above
0.85, however, bubble equilibria appear, increasing the probability of bubbles in
a stochastic environment. Note however that even when β is large enough (e.g.

14
0.9) to produce only bubble equilibria in the deterministic version of the model,
the probability of a bubble is not 1 in the stochastic version. The reason is
that the noise can lead the exchange rate within the basin of attraction around
the fundamental or, more importantly, that the shocks in the fundamentals dis-
place the basin of attraction leading to a crash in the bubble.The frequency of

Figure 7:

the occurrence of bubbles also depends on the parameter γ which measures the
rate with which chartists and fundamentalists revise their forecasting rules. We
have called this parameter rate of revision. In a way, γ also measures the speed
with which agents learn about the profitability of the other rule and revise their
forecasts. The lower is this parameter the less frequently agents will revise their
forecasting rules. In the limit when γ = 0 the agents never revise their forecasts
which could be interpreted as a world which agents perceive to be stationary.
In order to illustrate the importance of this parameter, we first show the
results of the deterministic simulations in figures 8. We observe that for values
of γ lower than (approximately) 1.2 the exchange rate converges to its funda-
mental value. For higher values we obtain bubble equilibria11 . Note also a zone
of complexity where the location of the bubble equilibria is very sensitive to
small changes in the parameter γ. In figure 9 we show the results of the sto-
chastic simulation under the same parameter configuration. We observe that
for low values of γ the occurence of bubbles is very infrequent. As γ increases
the frequency of bubbles increases significantly.
The previous results allow us to shed some additional light on the nature
of bubbles and crashes. As we have seen before, bubbles arise because agents
are attracted by the profitability of the extrapolating (chartist) rule, and this
1 1 In appendix 3 we show a similar figure where we have set β = 0.9. In that case the critical

value of γ which produces bubble equilibria is lowered.

15
attraction in turn makes this forecasting rule more profitable, leading to a self-
fulfilling increase in profitability. For this dynamics to work, agents’ decision to
switch must be sufficiently sensitive to the relative profitabilities of the rules. If
it is not, no bubble equilibria can arise, as is the case when γ does not exceed
1. The larger is γ the more likely it is that these self-fulfilling bubble equilibria
arise. The interesting aspect of this result is that in a world where agents quickly
react to changing profit opportunities, bubbles become more likely than in a
world where agents do not react quickly to these new profit opportunities.
The policy implication of this result is that by increasing the inertia in the
system so that agents react less quickly to changes in relative profitabilities of
forecasting rules, the authorities could reduce the probability of the occurrence
of bubbles. How this can be done and whether some form of taxation of exchange
transactions can do this, is a question we want to analyse in future research.

Figure 8:

6 The model with an endogenous current ac-

count
In the previous sections we used a model in which the exchange rate is often
disconnected from its fundamentals. This happens despite the fact that the
fundamentalists know the fundamental value of the exchange rate and use that
information to forecast the future exchange rate. The selfulfilling expectations of
the chartists regularly crowds out the fundamentalists from the market thereby
weakening the mean reversion forces in the model.
One drawback of the model is that when the fundamentalists leave the mar-
ket, there is no mean reverting force present anymore because the goods market

16
 Figure 9:

is kept outside the model. Incorporating the goods market provides an inde-
pendent channel through which the exchange rate can be forced back to its
equilibrium even if the fundamentalists are temporarily absent. In this section
we introduce the goods market into the model and allow for an interaction
between the goods market and the foreign exchange market.
We use the same model as in the previous sections, except that we now
endogenize the current account. It will be remembered that the current account
determines the net supply of foreign assets in the model. Thus we define the
current account as
∆Zt = Xt .
The current account consists of the trade balance and the net income from
net foreign assets. At this stage of the analysis we will concentrate on the role
of the trade balance and we disregard the role of the net income from foreign
assets. Thus, we will set the current account equal to the trade balance.
There is a large literature on the determinants of the trade balance. Here we
focus on the role of the exchange rate. We postulate that an increase (decline) of
the exchange rate leads to an improvement (deterioration) of the trade balance,
and thus of the current account, ceteris paribus. The reason is that an increase in
the exchange rate (a depreciation) stimulates exports and discourages imports.
The opposite holds for a decline in the exchange rate (an appreciation). This
relationship between the trade balance and the exchange rate holds as a ceteris
paribus proposition. In particular, it holds for a given domestic and foreign price
level. Put differently, it is the real exchange rate that matters for exporters
and importers. In the context of our model it is the difference between the
nominal exchange rate and the fundamental exchange rate that matters for the
decisions of exporters and importers. The fundamental exchange rate can then
be considered to be the difference between the domestic and the foreign price

17
levels. This leads us to postulate the following relationship between the trade
account (the current account) and the exchange rate:

Xt = ρx Xt−1 + (1 − ρx )²(st−1 − s∗t−1 ) (12)

that is, when the exchange rate, st−1 ,exceeds its fundamental value, s∗t−1 ,
the current account improves and vice versa. The sensitivity of the current
account with respect to the exchange rate is given by the parameter ² ≥ 0. This
parameter synthesises the reactions of exporters and importers to changes in the
exchange rate. It can easily be derived from a model of the export and import
markets. We will call this parameter, the trade elasticity, or ”elasticity” for
short. We also assume that the reactions of the exporters and importers is not
instantenous. The speed with which the trade account adjusts to the exchange
rate is given by ρx . Note that 0 ≤ ρx ≤ 1. When ρx is close to 1, there is a
lot of inertia in the trade account and the adjustment to the exchange rate is
very slow The opposite holds when ρx is close to 0. Equation 12 can also be
rewritten as follows:
∞
X
Xt = ²(1 − ρx ) ρi−1 ∗
x (st−i − st−i ) (13)
i=1

i.e., the trade account reacts to all past exchange rates with geometrically
declining weights. From 13 it can be seen that when st = s∗t for all t , Xt = 0.
Thus, when the exchange rate equals its fundamental value, the current ac-
count is in equilibrium. Put differently, we have an equivalence between current
account equilibrium and fundamental equilibrium of the exchange rate.
We will now assume that the fundamentalists are aware of this. It will be
remembered that the forecasting rule of the fundamentalists was assumed to be
¡ ¢
Etf (∆st+1 ) = −ψ st−1 − s∗t−1

Taking the equivalence between current account and fundamental equilbrium,

the fundamentalists set
ψ = −²(1 − ρx ) (14)
i.e., when the exchange rate deviates from its fundamental value (e.g. its PPP-
value), the fundamentalists know that the current account dynamics will ensure
that the speed of adjustment of the exchange rate towards equilibrium is given
by equation 14.
We now have all the elements to solve the model. We proceed in the same
way as before, i.e. we first analyse the deterministic solution, and then we add
stochastics to the model.

6.1 Solution of the model

We first set out the solutions for the exchange rate (attactors) as a function of
the initial conditions for a particular parameter configuration. (We will perform
a sensitivity analysis later). We show the results in figure 10. The contrast with

18
 Figure 10:

the results obtained when the current account was exogenous is important. We
observe that in the model with exogeneous current account (see equation ??)
we obtain bubble equilibria for sufficiently large initial shocks. This is no more
the case when the current account is endogenous. Whatever the initial shock
the exchange rate converges to the fundamental exchange rate (which as will be
remembered was normalised to 0). Thus the endogeneity of the current account
adds an important mean reverting process preventing bubble equilibria from
arising. Note that this result hinges on a particular value of the elasticity that
was selected here. We will return to this when we perform a sensitivity analy-
sis.Other types of equilibria are possible when the current account is endogenous.
We show a simulation in which we increase the sensitivity of the chartists and
fundamentalists to the profitability of the forecasting rules (the parameter γ).
We present the results both for the cases of an endogenous and an exogenous
current account (see figures 11 and 12). The most striking feature is that we
now obtain chaotic attractors when the current account is endogenous. Note,
however, that these chaotic attractors are centered around the fundamental ex-
change rate. The intuition of this result is that in the absence of endogeneity of
the current account, the bubble equilibria are far removed from the fundamental
equilibrium when γ is high. Making the current account endogenous introduces
a strong mean reverting process ensuring that the exchange rate stays close
to the fundamental one, but without preventing erratic movements around the
fundamental exchange rate.
We have also analysed the behaviour of the attractors when shocks occur in
the fundamental exchange rate, s∗t . We show the results in figure 13 and 14 for
the case with and without endogeneity in the current account. We observe a
similar contrast between the two regimes. When the current account is exoge-
nous (figure14), we obtain bubble equilibria. These disappear when the current

19
Figure 11:

Figure 12:

20
 Figure 13:

account is endogenous. Instead bubble equilibria are transformed into chaotic

attractors (at least if γ is sufficiently high).

6.2 Sensitivity analysis: the importance of the elasticity

In this section we present the result of applying an extensive sensitivity analysis.
We start with an analysis of the importance of the sensitivity of the current
account to the exchange rate (parameter ε). As will be remembered, we call this
parameter the ”elasticity”. We show the results in figure 15. As before this figure
is comparable to a ”bifucation” diagram. It plots the exchange rate solutions
(attractors), obtained after simulating the model over 1000 periods as a function
of the elasticity. When the elasticity is zero, the current account does not react
to the exchange rate changes. We observe that in that case we obtain a bubble
attractor. Note that the exact value of this attractor depends on the initial
conditions, as was shown above. When the elasticity exceeds zero the current
account reacts endogenously to exchange rate changes. For small values of the
elasticity we continue to obtain bubble attractors. These, however, come closer
to the fundamental value (which, as before is normalized to 0) as the elasticity
increases. For sufficiently high values of the elasticity, we obtain a fundamental
equilibrium. There is a range of elasticities for which the exchange rate moves
within the confines of a chaotic attractor centered around the fundamental value
of the exchange rate. It is also noteworthy that zones of chaotic attractors
and fixed point attractors alternate. This phenomenon is sometimes called
”intermittency”.
In order to shed more light on the importance of the elasticity parameter for

21
Figure 14:

Figure 15:

22
 Figure 16:

the occurrence of bubbles we simulated the model in a stochastic environment.

We then computed the frequency of the occurrence of bubbles. As before, we
define a bubble as a situation in which the exchange rate deviates from its
fundamental value by three times the standard deviation of the fundamental
during at least 20 consecutive periods. We show the results in figure 16. Each
point represents the average frequency of the occurrence of such a bubble for
a particular value of the elasticity. These frequencies were obtained by making
100 simulations of 1000 periods for each elasticity. We observe that for low
elasticities we obtain a significant frequency of bubbles. For elasticities above 1
the frequency of bubbles drops close to zero.

6.3 Sensitivity analysis: the importance of other parame-

ters
We performed similar sensitivity analysis with respect to other parameters of
the model. Figure 17 shows such a sensitivity analysis of the importance of the
chartists’ extrapolation parameter, β. It can be seen that with an increasing
β, we move from fixed point attractors into a zone of chaotic attractors. This
contrasts with the case of an exogenous current account. In that case, increases
in β lead to bubble equilibria. A similar phenomenon is observed when we allow
the parameter γ to increase. As will be remembered this parameter measures
the sensitivity of the chartists and fundamentalists to changes in profitability of
the forecasting rules. With an exogenous current account increases in γ lead to
bubble equilibria, while they lead to chaotic attactors when the current account
is endogenous( see figure18).

23
Figure 17:

Figure 18:

24
 We conclude from this sensitivity analysis that the endogeneity of the current
account changes the dynamics of the exchange rate profoundly. In general, we
find that in a model withouth a feedback from the current account, bubbles and
crashes are an endemic feature of the exchange rate dynamics. When, however,
the current account reacts to exchange rate changes, this dynamics is changed.
The bubbles and crashes dynamics is weakened. Instead a complex, sometimes,
chaotic dynamics takes the place of the bubbles dynamics.

7 The model with transactions costs

In this section we add an additional complication, i.e. the existence of transac-
tion costs. There is an increasing body of theoretical literature stressing the im-
portance of transactions costs in the goods markets as a source of non-linearity
in the determination of the exchange rate (Dumas(1992), Sercu, Uppal and
Van Hulle(1995), Obstfeld and Rogoff(2000)). The importance of transaction
costs in the goods markets has also been confirmed empirically (Taylor, Peel,
and Sarno(2001), Kilian and Taylor(2001)). In addition, several recent empir-
ical studies report the continued existence of large price differentials for the
same traded goods across borders (see Haskel and Wolf(2001) and Engel and
Rogers(1995)). These authors document price differentials of up to 40% for
identical products in different countries. This indicates that producers apply
‘pricing to market’. Such pricing strategies, however, can only be applied suc-
cessfully if transaction costs prevent arbitrage. Thus, the large observed price
differentials suggest that transactions costs for traded goods are large. In addi-
tion, for many services, which are non-traded goods, transactions costs are even
higher. (See Obstfeld and Rogoff(2001) who argue that transactions costs are
key to understanding the major puzzles in international economics).
We assume that the existence of transactions costs creates a band of inaction,
i.e. when the deviation of the exchange rate from its fundamental is lower than
the transactions costs, goods arbitrage is not profitable. Put differently, when
the exchange rate is within this band, export and import decisions are not
sensitive to changes in the exchange rates (as long as these changes do not move
the exchange rate outside the band). We, therefore, write that

Xt = ρx Xt−1 + (1 − ρx )²(st−1 − s∗t−1 ) (15)

¯ ¯
¯ ∗ ¯
holds if st−1 − st−1 ≥ C, where C is the transaction cost in the goods
market assumed to be of the iceberg type. Otherwise Xt = 0.(In the stochastic
version of the model we will then set Xt = εt , where εt is white noise).
In addition, we assume that fundamentalists are aware of the existence of
this band of inaction. As a result, when the exchange rate deviation from its fun-
damental value is larger than the transaction costs C, then the fundamentalists
follow the forecasting rule as in equation 7. More formally,
when | st−1 − s∗t−1 |≥ C holds, then equation 7 applies12 .
1 2 Note that since ψ < ∞ market inefficiencies other than transaction costs continue to play

25
 However when the exchange rate deviations from the fundamental value are
smaller than the transaction costs in the goods markets, there is no mechanism
that drives the exchange rate towards its equilibrium value. As a result, fun-
damentalists expect the changes in the exchange rate to follow a white noise
process and the best they can do is to forecast no change. More formally,
when | st −s∗t |< C, then Etf (∆st+1 ) = 0. or Etf (∆st+1 )
= ηt and η t is white noise in the stochastic version of the model

7.1 Solution of the model

We proceed as before, i.e. we solve the deterministic part of the model, and
present the attractors as a function of the initial conditions. We have selected a
value of the elasticity of the current account such that bubbles equilibria do not
occur. We show the solution for the same parameter combinations as in the basic
model (see figure 1). The results are now affected in a fundamental way. First, all
the fixed point attractors are located within the transaction cost band. Second
the attractors are arranged in a complex manner. Some attractors are located
on a continuous line, others are spread around in a discontinuous way. As a
result, small changes in the initial conditions can lead to large displacements
in the attractors. As will be shown, this feature of the attractors creates a
complex dynamics in a stochastic environment characterised by sensitivity to
initial conditions. Note that in the simulations reported here there is no chaotic
dynamics, as each attractor is a fixed point. These fixed point attractors are
separated by basins of attraction with complex features. These are responsible
for the sensitivity to initial conditions.
Before analysing the stochastic solutions it is useful to contrast the attractors
obtained in figure 19 with those that one obtains when the current account is
exogenous (elasticity =0). We show these in the following figure 20. We find, as
before, that in the absence of a feedback from the current account, the model
produces bubble equilibria. Thus the existence of transactions costs does not
affect our previous result.

7.2 The model in a stochastic environment

We now proceed in solving the model in a stochastic environment. We show
the results in the time domain for a particular combinations of parameters
that do not produce bubbles and crashes, i.e. with an elasticity of the current
account that is sufficiently large (² = 0.5). We show the results in figure 21.
Figure 21 shows the simulated exchange rate together with the fundamental
exchange rate. We observe that the exchange rate is disconnected most of
the time from its underlying fundamental. This feature is present despite the
absence of bubble equilibria. Note that sometimes the exchange rate dynamics
a role when the exchange rate moves outside the transaction costs band. As a result, these
inefficiencies prevent the exchange rate from adjusting instantaneously.

26
Figure 19:

Figure 20:

27
 Figure 21:

has the appearance of bubbles and crashes although there is no such underlying
dyamics. We will analyse the nature of this ”disconnect puzzle” in greater detail
when we discuss the empirical relevance of the model. We also observe that the
short-term volatility of the simulated exchange rate appears to be significantly
higher than the volatility of the underlying fundamental. This feature is related
to the sensitivity to initial conditions dynamics: small changes in the underlying
stochastics can lead to large displacements of the attractor. This sensitivity to
initial conditions can also be illustrated in another way. We simulated the model
with two different initial conditions in the exchange rate. In one simulation we
set the initial exchange rate equal to 4, in the other to 4.01. All the rest is kept
identical in the two simulations (the parameters, the underlying stochastics
driving the exogenous variables). The results of figure 22 illustrate the power of
the sensitivity to initial conditions. A small disturbance in the initial exchange
rate leads to sustained deviations between the two exchange rates, despite the
fact that the underlying fundamentals are identical. It appears that the two
exchange rates follow a different ”history”.
One can conclude this section as follows. Without a feedback from the current
account the model produces bubbles and crashes. These bubbles and crashes
tend to disappear when we allow the current account to react to exchange rate
changes. The introduction of transactions costs does not change this result.
However, transactions costs create a new, and complex dynamics whereby the
exchange rate appears to be disconnected most of the time from the underlying
fundamental creating a resemblance with bubbles and crashes. In addition, by

28
 Figure 22:

creating a band of inaction transactions costs are also responsible for the appear-
ance of sensitivity to initial conditions that make it possible for small changes
in the initial conditions to have profound effects on the future movements of the
exchange rate. Thus, in such a world, history seems to matter.

8 Empirical relevance of the model

In this section we analyse how well our model mimics the empirical anomalies
and puzzles that have been uncovered by the flourishing empirical literature. We
calibrate the model such that it replicates the observed statistical properties of
exchange rate movements. In order to do so we selected a parameter config-
uration that mimics these properties most closely. We discuss these different
statistical properties in the following sections.

8.1 The disconnect puzzle

In the previous section we showed how the model was capable of replicating a
widely observed phenomenon, i.e. that the exchange rate is often disconnected
from its underlying fundamentals. In that analysis we stressed one dimension of
the disconnect puzzle, i.e. the causality (or the lack of it) from the fundamental
to the exchange rate. There is, however, another dimension to the disconnect
puzzle, which relates to the causality running from the exchange rate to the

29
 Figure 23:

current account. From the empirical evidence it appears that the exchange rate
has a weak and unpredictable influence on the fundamentals, in particular on
the current account. (see Obsfeld and Rogoff for a formulation of this puzzle).
The version of the model in which the current account reacts endogenously to
the exchange rate can be used to shed light on this puzzle. In order to do so we
simulated the model assuming an elasticity of 0.5. Thus, we assume that there is
a causality running from the exchange rate to the current account. We show two
examples of such a simulation for different parameter configurations in figures
23 and 24. The interesting aspect of this simulation is that there are periods
in which the current account is influenced by the exchange rate movements.
These periods, however, alternate with other periods during which this influence
is almost totally absent, making the effect of exchange rate changes on the
current account very unpredictable. The intuition of this result is that during
periods of turbulence, the effect of exchange rate changes on the current account
is weakened. Since turbulent and tranquil periods alternate in unpredictable
fashion, the effects of exchange rate changes on the current account also alternate
from being predictable to becoming unpredictable.
In order to be more precise about the nature of the disconnect puzzle we
analyzed the simulated exchange rate and current account econometrically. We
first tested for cointegration of the two series, and found that in general the
two series are cointegrated. We then specified a vector error correction (VEC)
model in the following way:

30
 Figure 24:

n
X n
X
∆st = µ (st−1 − γXt−1 ) + λi ∆st−i + φi ∆Xt−i (16)
i=1 i=1
Xn Xn
∆Xt = µ0 (st−1 − γ 0 Xt−1 ) + λ0i ∆st−i + φ0i ∆Xt−i (17)
i=1 i=1

The first term on the right hand side in both equations are the error cor-
rection terms. We have estimated this model for a broad range of parameter
values. The result of estimating equation 16 for selected parameter values is
presented in table 1 where we have set C = 5, β = 0.9, γ = 1, ² = 0.5 and
ρx = 0.6 and number of lags n = 5.
We find that the error correction coefficients (µ and µ0 ) are low in both equa-
tions. This suggests that the mean reversion towards the equilibrium exchange
rate and current account takes a long time. In particular, only 0.7% and 0.6%
of the adjustments take place each period. It should be noted that in the simula-
tions we have assumed a speed of adjustment in the current account equation of
0.2 (ρ = 0.6 and ² = 0.5), implying that structurally the 20% of the disequilib-
rium in the current account should be corrected. Instead, our model generates
a much slower adjustment in the current account. This slow adjustment in the
current account is due to the chartists’ extrapolation behaviour which adds a
lot of noise in the exchange rate movements. Thus, our model helps to explain
the disconnect puzzle. As stressed by Obstfeld and Rogoff(2000) the disconnect
phenomenon runs in both directions: the exchange rate is disconnected from
the fundamentals (in this case the current account) and the fundamental (the

31
 Table 1: error correction model
Error correction ∆st−i ∆Xt−i
µ γ λ1 λ2 λ3 λ4 λ5 ϕ1 ϕ2 ϕ3 ϕ4 ϕ5
0.007 -3.5 0.38 0.17 0.13 0.03 -0.03 0.02 0.06 0.04 0.05 0.08
2.2 -3.2 8.4 3.5 2.7 0.5 -0.7 0.4 1.2 0.7 1.1 1.6

µ0 γ0 λ01 λ02 λ03 λ04 λ05 ϕ01 ϕ02 ϕ03 ϕ04 ϕ05
0.006 -3.5 0.24 0.04 0.08 0.09 0.13 -0.10 -0.07 -0.07 -0.02 0.01
2.5 -3.2 6.2 0.9 1.9 2.0 3.0 -2.3 -1.5 -1.6 -0.4 0.2

current account) is disconnected from the exchange rate most of the time.
From table 1, we also note an asymmetry in the disconnect puzzle. We ob-
serve that the past changes in the current account have a weak and insignificant
effect on the current exchange rate changes. The converse is not true. Past
changes in the exchange rate have a significant effect on the current changes in
the current account.
The picture that emerges from this analysis can be summarized as follows.
There is a long run cointegration relationship between the exchange rate and the
current account .However, the speed of adjustment of both the exchange rate
and the current account towards this long run equilibrium relationship is very
weak despite the fact that we have built into the structure of the model relatively
strong speeds of adjustment. It is in this sense that there is a disconnect puzzle
that runs in both directions. There is an asymmetry though. Past exchange
rate changes have a significant effect on today’s changes in the current account.
The reverse does not seem to be the case.
We have estimated similar error correction models on simulated exchange
rates and current accounts for other parameter values of the model. We show
the estimated µ0 s and γ 0 s in table 5. We find similarly low speeds of adjustment,
and an asymmetry in the coefficients of the past changes in the exchange rate
and the current account (not shown).

8.2 Fat tails and excess kurtosis

It is well known that the exchange rate changes do not follow a normal dis-
tribution. Instead it has been observed that the distribution of exchange rate
changes has more density around the mean than the normal and exhibits fatter
tails than the normal (see de Vries(2001)). This phenomenon was first discov-
ered by Mandelbrot (1963), in commodity markets. Since then, fat tails and
excess kurtosis have been discovered in many other asset markets including the
exchange market. In particular, in the latter the returns have a kurtosis typ-
ically exceeding 313 and a measure of fat tails (Hill index) ranging between 2
and 5 (see Koedijk, Stork and de Vries (1992), Huisman, et al.(2002)). It im-
plies that most of the time the exchange rate movements are relatively small
1 3 The normal distibution has a kurtosis index equal to 3.

32
 Table 2: ecm for different parameter values
EC-coefficient Error correction
parameter values µ µ0 γ
C = 5, β = 0.8, γ = 1, ² = 0.5 -0.003 0.01 -1.66
-0.5 1.6 -3.8
C = 5, β = 0.8, γ = 10, ² = 0.5 -0.01 0.01 -1.45
-1.7 1.9 -4.7
C = 5, β = 0.8, γ = 10, ² = 1 -0.01 0.01 -0.64
-1.6 2.5 -4.7
C = 5, β = 0.9, γ = 10, ² = 1 -0.02 0.1 -0.65
-2.8 0.9 -5.7
C = 5, β = 0.8, γ = 1, ² = 0.5 -0.02 0.01 -0.05
-4.2 1.4 -0.04

Table 3: Kurtosis and Hill index USD-DEM and JPY-DEM 1975-98

kurtosis median Hill index
Exchange rate 2.5% tail 5% tail 10% tail
USD-DEM 12.1 4.0 3.6 3.1
JPY-DEM 19.6 3.7 3.6 2.9

but that occasionally periods of turbulence occur with relatively large exchange
rate changes.
In table 3 we show the kurtosis and the Hill.index of the USD-DEM and
the JPY-DEM exchange rate returns for the period 1975-1998 . We computed
the Hill index for different cut-off points of the tails (2.5%, 5%, 10%) and for
4 different subsamples of the original series. We find that these exchange rates
exhibited excess kurtosis and fat tails during the sample period.
Another empirical finding that has been observed is that the kurtosis is
reduced under time aggregation (see Lux(1998), Calvet and Fisher(2002)). We
checked this finding for the same exchange rates. In table 4 we show the results
for USD-DEM and JPY-DEM exchange rates, and we confirm that the kurtosis
declines under time aggregation.
The next step in the analysis was to check whether these empirical features
are also shared by the simulated exchange rate changes in our model.
The model was simulated using normally distributed random disturbances
(with mean = 0 and standard deviation = 1). We computed the kurtosis and

Table 4: Kurtosis and time aggregation USD-DEM and JPY-DEM 1975-98

5 period 10 period 50 period
Parameter values returns returns returns
USD-DEM 7.4 5.3 3.4
JPY-DEM 14.9 5.7 2.7

33
the Hill index of the simulated exchange rate returns. We computed the Hill
index for 4 different samples of 2000 observations. In addition, as before, we
considered three different cut-off points of the tails (2.5%, 5%, 10%). We show
the results of the kurtosis and of the Hill index in table 5. We find that for
a broad range of parameter values the kurtosis exceeds 3 and the Hill index
indicates the presence of fat tails.

Table 5: Kurtosis and Hill index

kurtosis median Hill index
Parameter values 2.5% tail 5% tail 10% tail
C=5, β=0.9, γ=1;²=0 21.5 6.1 5.3 4.3
C=5, β=0.9, γ=0.5;²=0.5 7.9 3.5 3.4 3.2
C=5, β=0.9, γ=1;²=0.5 20.5 3.1 2.8 3.0
C=5, β=0.9, γ=5;²=0.5 35.9 3.4 3.6 3.5
C=5, β=0.8, γ=1;²=0 15.5 5.6 5.0 4.0
C=5, β=0.8, γ=0.5;²=0.5 6.0 3.3 3.8 3.5
C=5, β=0.8, γ=1;²=0.5 7..3 3.5 3.7 3.6
C=5, β=0.8, γ=0.5,²=0.8 11.9 2.9 3.3 3.1

In figure 25 we show the probability density of the USD-DEM exchange

rate and of our simulated exchange rates, up-left and down-left panel respec-
tively. For the sake of comparison, we plot the probability density of normally
distributed returns on the right panel. We observe that the empirical distrib-
ution differs from the normal distribution and that it strikingly resembles the
distribution of our simulated exchange rate returns.
Finally we check if the kurtosis of our simulated exchange rate returns
declines under time aggregation. In order to do so, we chose different time
aggregation periods and we computed the kurtosis of the time-aggregated ex-
change rate returns. We found that the kurtosis declines under time aggregation.
In table 6 we show the results for some sets of parameter values14 .
The previous results suggest that the speculative dynamics of the model
transforms normally distributed noise in the exchange rate into exchange rate
movements with tails that are significantly fatter than the normal distribution
and with more density around the mean. Thus, our model mimics an important
empirical regularity, i.e. that exchange rate movements are characterised by
tranquil periods (occurring most of the time) and turbulent periods (occurring
infrequently).This phenomenon has been also called intermittency phenomenon
(see Lux(1998)).

8.3 The ” excess volatility” puzzle

In this section we analyse another important empirical regularity, which has
been called the ”excess volatility” puzzle, i.e. the volatility of the exchange rate
1 4 Another empirical regularity of the distribution of exchange returns is its symmetry. We

computed the skewness, and we could not reject that the distribution is symmetric.

34
 Figure 25:

by far exceeds the volatility of the underlying economic variables. Baxter and
Stockman (1989) and Flood and Rose (1995) found that while the movements
from fixed to flexible exchange rates led to a dramatic increase in the volatility
of the exchange rate no such increase could be detected in the volatility of
the underlying economic variables. This contradicted the ’news’ models that
predicted that the volatility of the exchange rate can only increase when the
variability of the underlying fundamental variables increases ( see Obstfeld and
Rogoff (1996) for a recent formulation of this model)15 .
In order to deal with this puzzle we compute the noise to signal ratio in the
simulated exchange rate. We derive this noise to signal ratio as follows:

var(s) = var(f ) + var(n) (18)

where var(s) is the variance of the simulated exchange rate, var(f) is the
variance of the fundamental and var(n) is the residual variance (noise) pro-
1 5 In addition, Goodhart (1989) and Goodhart and Figlioli (1991) found that most of the

changes in the exchange rates occur when there is no observable news in the fundamental
economic variables. This finding contradicted the theoretical models (based on the efficient
market hypothesis), which imply that the exchange rate can only move when there is news in
the fundamentals.

35
 Table 6: Kurtosis and time aggregation
5 period 10 period 25 period 50 period
Parameter values returns returns returns returns
C=5, β=0.9, γ=1,²=0 44.9 32.3 4.9 3.8
C=5, β=0.9, γ=0.5,²=0.5 9.7 10.7 3.2 3.0
C=5, β=0.9, γ=1,²=0.5 10.8 10.9 3.4 2.4
C=5, β=0.9, γ=5,²=0.5 17.3 6.4 3.4 3.6
C=5, β=0.8, γ=1,²=0 85.3 30.3 3.9 2.5
C=5, β=0.8, γ=0.5,²=0.5 5.4 3.4 3.8 2.9
C=5, β=0.8, γ=1,²=0.5 6.2 3.9 3.8 2.9
C=5, β=0.8, γ=0.5,²=0.8 5.3 4.0 4.5 3.0

duced by the non-linear speculative dynamics which is uncorrelated with var(f).

Rewriting (18) we obtain

var(n) var(s)
= −1 (19)
var(f ) var(f )
The ratio var(n)/var(f ) can be interpreted as the noise to signal ratio. It
gives a measure of how large the noise produced by the speculative dynamics
is with respect to the exogenous volatility of the fundamental exchange rate.
We simulate this noise to signal ratio for different values of the extrapolation
parameter β (see figure ??). In addition, since this ratio is sensitive to the time
interval over which it is computed we checked how it changes depending on the
length of the time interval. In particular, we expect that the noise-to-signal
ratio is larger when it is computed on a short than on a long time horizon. We
show the results in figure ?? which assumes the same parameter configuration
as ??.
First, we find that with increasing β the noise to signal ratio increases. This
implies that when the chartists increase the degree with which they extrapo-
late the past exchange rate movements, the noise in the exchange rate, which
is unrelated to fundamentals, increases. Thus, the signal about the fundamen-
tals that we can extract from the exchange rate becomes more clouded when
the chartists extrapolate more. Second, we find that when the time horizon
increases the noise-to-signal ratio declines. This is so because over long time
horizons most of the volatility of the exchange rate is due to the fundamentals’
volatility and very little to the endogenous noise. In contrast, over short time
horizons the endogenous volatility is predominant and the signal that comes
from the fundamentals is weak. This is consistent with the empirical find-
ings following Meese and Rogoff(1983) celebrated studies. This literature tells
us that when the forecasting horizon increases the performance of forecasting
based on fundamentals tends to improve relative to random walk forecasting
(see Mark(1995), Faust, et al. (2002)).

36
Figure 26:

Figure 27:

37
9 Conclusion
In this paper we provide a framework for analysing the dynamics of exchange
rate movements. The special feature of our model is that individual agents
recognize that they are not capable of understanding and processing the complex
information structure of the underlying model. As a result, they use simple rules
to forecast the exchange rates. None of these rules is rational in the technical
sense. Yet we claim that these agents act rationally within the context of the
uncertainty they face. That is, agents check the ’fitness’ (profitability) of the
forecasting rule at each point in time and decide to reject the rule if it is less
profitable (in a risk adjusted sense) than competing rules. Our model is in the
tradition of evolutionary dynamics where agents use trial and error strategies.
We assume that some of the forecasting rules are based on extrapolating past
exchange rate movements (chartism) and others are based on mean reversion
towards the fundamental rate (fundamentalism).
We analysed this model first within a framework where the current account
(the fundamental) is exogenous. The model then generates two types of equilib-
ria. The first one, which we called a fundamental equilibrium, is one in which
the exchange rate converges to its fundamental value. The exchange rate, how-
ever, can also converge to a second type of equilibrium, which we called a bubble
equilibrium, and which is reached in a self-fulfilling manner. An important fea-
ture of the bubble equilibrium is that chartism (extrapolative forecasting) takes
over most of the market. We simulated the model in a stochastic environment
and generated complex scenarios of bubbles and crashes. One interesting aspect
of the model is that it explains both the emergence of the bubble and its subse-
quent crash. The model also predicts that bubbles and crashes are asymmetric,
i.e. the bubble phase is slower than the subsequent crash. This asymmetry has
been widely observed in financial markets. It cannot be explained by RE-models
of bubbles and crashes which predict symmetry (Blanchard and Watson(1982)).
We also analysed under what conditions bubbles and crashes occur. We find
that when agents react quickly to changing relative profitabilities of the different
forecasting rules, the frequency of bubbles increases. In such an environment
chartists will make large profits and will tend to dominate the market, crowding
out fundamentalists who have a poor forecasting record and make losses. It will
then be quite rational to be a chartist.
In a second stage we extended the model to allow for a feedback from the
current account. We found that when the elasticity of the current account
with respect to exchange rate changes is sufficiently high the bubble equilibria
disappear.
In a third stage, we introduced transactions costs in the goods market. These
transactions costs create a band of inaction within which exports and imports
do not react to exchange rate changes. The implications of introducing such
transactions costs for the dynamics of the exchange rate are far-reaching. We
found that this version of the model is capable of generating the disconnect
phenomenon (misalignment), whereby the exchange rate is disconnected most of
the time from the underlying fundamentals. Interestingly, these misalignments

38
resemble the bubbles and crashes dynamics, although the deterministic part of
the model does not produce bubble equilibria.
We also found that the model produces a ”sensitivity to initial conditions”,
i.e.small stochastic changes have permanent effects on the future movements
of the exchange rate. This imples that exchange rates are influenced by trivial
shocks in a permanent way.
Finally we tested our model in the sense that we reproduced the statistical
properties of exchange rate changes observed in reality, i.e. excess volatility,
excess kurtosis and fat tails.

10 References
Anderson, S., de Palma, A., Thisse, J.-F., 1992, Discrete Choice Theory of
Product Differentiation, MIT Press, Cambridge, Mass.
Bacchetta, P. and van Wincoop E., 2003, Can information heterogeneity
explain the exchange rate determination puzzle?, NBER working paper 9498.
Baxter, M., Stockman, 1989, A., ”Business Cycles and the Exchange Rate
Regime. Some International Evidence”, Journal of Monetary Economics, 23,
may 377-400.
Blanchard ,O.J., 1979, ” Speculative bubbles, crashes and rational expecta-
tions”, Economics Letters, 3, 387-389.
Blanchard, O.J., and Watson, M.W., 1982, ”Bubbles, rational expectations
and speculative markets”, in Wachtel, P., eds., Crisis in economic and financial
stucture: bubbles, bursts, and shocks. Lexington books: Lexington.
Brock, W., and Hommes, C., 1997, A Rational Route to Randomness, Econo-
metrica, 65, 1059-1095
Brock, W., and Hommes, C.,1998, Heterogeneous beliefs and routes to chaos
in a simple asset pricing model, Journal of Economic Dynamics and Control,
22, 1235-1274.
Brunnermeier, M., 2001, Asset pricing under asymmetric information: bub-
bles, crashes technical analysis and herding, Oxford University Press.
Cheung Y. and Lai K., 2000. ” On the purchasing power parity puzzle”.
Journal of International Economics, 52 .
Cheung Y., Lai K. and Bergman M., 2001, ” Dissecting the PPP puzzle: the
unconventional roles of nominal exchange rate and price adjustments”. Paper
presented at CES-Ifo Conference Munich 2002.
Cheung, Y., and Chinn, M., (1989), Macroeconomic Implications of the Be-
liefs and Behavior of Foreign Exchange Traders, mimeo, University of California,
Santa Cruz.
Cheung, Y., Chinn, M., and Marsh, (1999), How Do UK-Based Foreign
Exchange Dealers Think Their Markets Operates?, CEPR Discussion Paper,
no. 2230
Chiarella, C., Dieci, R., Gardini, 2002, L., “Speculative behaviour and com-
plex asset price dynamics”, Journal of Economic Behaviour and Organisation.

39
 Copeland, L., 2000, Exchange Rates and International Finance, 3rd ed.,
Prentice Hall.
De Grauwe, P. , Dewachter, H., and Embrechts, 1993, M., Exchange Rate
Theories. Chaotic Models of the Foreign Exchange Markets, Blackwell.
De Grauwe, P., and Grimaldi, M., Exchange Rate Puzzles. 2003, A Tale of
Switching Attractors, paper presented at the EEA Meeting, Stockholm, 2003
de Vries, C., 2000, ”Fat tails and the history of the guilder”, Tinbergen
Magazine, 4, Fall, pp. 3-6.
De Long, J., Bradford, B., Schleiffer and Summers, L., 1990, ”Noise Trader
Risk in Financial Markets”, Journal of Political Economy.
Dornbusch R., 1976, ”Expectations and exchange rate dynamics”, Journal
of Political Economy 84.
Engel C., 2000, ” Long run PPP may not hold after all”, Journal of Inter-
national Economics, 57.
Engel C. and Morley J., 2001, ”The adjustment of prices and the adjustment
of the exchange rate”, Discussion paper, Department of Economics, University
of Wisconsin.
Evans, M., and Lyons, R., 1999, ”Order Flow and Exchange Rate Dynam-
ics”, NBER Working Paper, no. 7317.
Evans, G., and Honkapohja, S., 2001, Learning and Expectations in Macro-
economics, Princeton University Press.
Faust, J., Rogers, J., Swanson, E., and Wright, J., 2002, Identifying the
effects of monetary policy shocks on exchange rates using high-frequency data,
Board of Governors of the Federal Reserve System, International Finance Dis-
cussion Paper, no. 739, October.
Flood, R, and Rose, A., 1995, ”Fixing the Exchange Rate Regime: A virtual
Quest for Fundamentals”, Journal of Monetary Economics, 36, August, 3-37.
Frankel, J., and Froot, K., 1986, ”The Dollar as a Speculative Bubble: A
Tale of Fundamentalists and Chartists”, NBER Working Paper, no. 1963.
Garber,P.M., 2000, ”Famous first bubbles”, MIT press.
Goodhart, C., 1989, ”News and the Foreign Exchange Market”, LSE Finan-
cial Markets Group Discussion paper, 71.
Goodhart, C., and Figliuoli, L., 1991, ”Every Minute Counts in the Foreign
Exchange Markets”, Journal of International Money and Finance, 10, 23-52.
Guillaume D., 1996 ”Chaos, randomness and order in the foreign exchange
markets” PhD Thesis K.U.Leuven
Hallwood, P., MacDonald, R., 1994, International Money and Finance, 2nd
ed., Blackwell, Oxford.
Huisman, R., Koedijk, K., Kool, C., and Palm, F., 2002, The tail-fatness of
FX returns reconsidered, in De Economist, 150, no. 3, September, 299-312.
Isard, P., 1995, Exchange Rate Economics, Cambridge University Press.
Johansen,A., Sornette,D., 1999, Modeling the stock market prior to large
crashes, The European Physical Journal B, 9, 167-174.
Johansen, S., 1991, Estimation and Hypothesis Testing of Cointegration
Vectors in Gaussian Vector Autoregressive Models, Econometrica, 55, 1551-80.

40
 Kandel,E. and Pearson, N.D., 1995, ”Differential interpretation of public
signals and trade in speculative markets”, Journal of political Economy, 4, 831-
872.
Kilian L. and M. Taylor, 2001, ”Why is it So Difficult to Beat the Random
Walk Forecast of Exchange Rates?” Mimeo, University of Warwick, pp. 29.
Kindleberger, C,. Manias, Panics, and Crashes. A History of Financial
Crises. 1978, John Wiley & Sons, New York, 263 pages.
Kurz, M., 1994, ”On the Structure and Diversity of Rational Beliefs”, Eco-
nomic Theory, 4, 877-900.
Kurz, M., and Motolese, M., 2000, ”Endogenous Uncertainty and Market
Volatility”, mimeo, Stanford University.
LI K., 1999, ”Testing symmetry and proportionality in PPP : A panel data
approach”, Journal of Business and Economic Statistics 17 (4) , 409-418.
Lui, Y., and Mole, D., The Use of Fundamental and Technical Analyses
by Foreign Exchange Dealers: Hong Kong Evidence, Journal of International
Money and Finance, 17, pp. 535-45
Lux T., 1998, ”The socio-economic dynamics of speculative markets: inter-
acting agents, chaos, and fat tails of return distributions”, Journal of Economic
Behaviour and Organisation, vol.33.
Lux T., Marchesi M., 2000, ”Volatility clustering in financial markets: a
microsimulation of interacting agents”, International Journal of Theoretical and
Applied Finance.
Lux T., Sornette D., 2002, "On rational bubbles and fat tails", Journal of
Money, Credit and Banking, 34, No 3, pp 589-610.
Lyons, R., 2001, The Microstructure Approach to Exchange Rates, MIT
Press, Cambridge, Mass.
Mandelbrot, B., 1963, The variation of certain speculative prices, The Jour-
nal of Business, University of Chicago, 36, 394-419.
Mandelbrot, B., 1997, Fractals and Scaling in Finance, Springer Verlag, 551
pages.
Meese, R., and Rogoff, 1983, ”Empirical Exchange Rate Models of the Sev-
enties: Do they Fit Out of Sample?”, Journal of International Economics, 14,
3-24.
Mentkhoff, L., (1997), Examining the Use of Technical Currency Analysis,
International Journal of Finance and Economics, 2, pp. 307-18
Mentkhoff, L., (1998), The Noise Trading Approach - Questionnaire Evi-
dence from Foreign Exchange, Journal of International Money and Finance, 17,
pp. 547-64.
Obstfeld, M. and Rogoff, K., 1996, Foundations of International Macroeco-
nomics, MIT Press, Cambridge, Mass.
Obstfeld, M., and Rogoff, K., 2000, ”The Six Major Puzzles in International
Macroeconomics: Is there a Common Cause?”, NBER Working Paper no. 7777,
July.
Rogoff, K., 1996, ”The purchasing power parity puzzle”, Journal of Economic
Literature, 34, June, 647-668.
Shiller, R., 2000, Irrational Exuberance, Princeton University Press,

41
 Schittenkopf C., Dorffner G., Dockner E.,2001, ”On nonlinear, stochastic
dynamics in economics and financial time series”, Studies in Nonlinear Dynamics
and Econometrics 4(3), pp. 101-121.
Schleiffer, A., 2000, Introduction to Behavioural Finance, Clarendon Press.
Sornette, D., 2003,Why Stock Markets Crash, Princeton University Press.
Taylor, M., and Allen, H., 1992, ”The Use of Technical Analysis in the
Foreign Exchange Market”, Journal of International Money and Finance, 11,
304-14.
Thaler, R., 1994, Quasi Rational Economics, Russell Sage Foundation, New
York.
Wei Shang-Jin and Kim Jungshik 1997. ”The big players in the foreign
exchange market:do they trade on information or noise?”. NBER working paper
6256.
Williamson, J., 1985, ”The Exchange Rate System”, Policy Analyses in In-
ternational Economics, 5, Institute for International Economics, Washington,
D.C.

42
A Fixed attractors and fundamental shocks: ad-
ditional results
In this appendix we present additional simulations of the effect of shocks in the
fundamental on the exchange rate. We assume different values of the initial con-
ditions. The results are shown in figures 28, 29and 30. When the initial condition
(noise) is small (figure A1) no bubble equilibria exist and the exchange rate
always coincides with its fundamental value. When the initial condition is grad-
ually increased (figures A2 and A3) the range of bubble equilibria progressively
increases.

Figure 28:

43
Figure 29:

Figure 30:

44
B Causality tests between exchange rate and
chartist weight
In this appendix we present the results of causality tests between the exchange
rate and the weight of chartists during a bubble and crash episode. We simulated
the model using the standard set of parameters, and we selected an episode
during which a bubble and crash occurred. We show such an episode in figure
A2. A visual inspection of the graph reveals that the exchange rate appears
to lead the chartist weight. at least when the bubble starts and later when the
bubble bursts. Note also that the crash occurs faster than the bubble phase, a
feature we often find in our simulated bubbles and crashes. This has also been
found in empirical data (see Sornette(2003))

Exchange rate and chartist weight during bubble

30 1.0

25
0.8
20

15 0.6

10
0.4
5

0 0.2

exchange rate chartist weight

Next we performed a Granger causality test on the exchange rate and the
chartist weight during the bubble and crash episode represented in figure A216 .
The result of this causality test is presented in table A1. We observe that we
cannot reject the hypothesis that the exchange rate leads the chartists’ weight
during the bubble and crash episode, while we can reject the reverse. We find
this feature in most bubble and crash episodes.
1 6 We checked for stationarity and could not reject that the two series are stationary during

the sample period.

45
 Table 7: Granger causality tests
Null Hypothesis: F-statistic Probability
cw not Granger cause exchange rate 0.377 0.865
exchange rate not Granger cause cw 6.85 6.4E-06
Note: obs=211, lags=5.

46
C Stylised Facts of JPY-DEM Exchange Rate
The up left panel shows the distribution of the JPY-DEM returns. The bottom
left panel represents the distribution of our simulated returns. The right panels
show the distribution of normally distributed exchange rate returns.

47