Content uploaded by Dongchul Cho
Author content
All content in this area was uploaded by Dongchul Cho on Aug 04, 2015
Content may be subject to copyright.
INTEREST RATES AND EXCHANGE RATES
IN THE KOREAN, PHILIPPINE AND THAI EXCHANGE RATE CRISES
Dongchul Cho
Korea Development Institute
Kenneth D. West
University of Wisconsin
February 2001
Revised April 2001
ABSTRACT
We consider the effect on exchange rates of an exogenous change in interest rates that is
induced by a surprise shift in monetary policy. Our two equation model, which is applicable
during periods of exchange rate crisis, consists of a monetary policy reaction function and an
interest parity relationship. We estimate a special case of the model using weekly data from
1997 and 1998 for Korea, the Philippines and Thailand. Point estimates indicate that exogenous
increases in interest rates led to exchange rate appreciation in Korea and the Philippines,
depreciation in Thailand. Confidence intervals around point estimates are huge, however.
Prepared for the National Bureau of Economic Research conference on Management of Currency
Crises, March 2001. We thank Akito Matsumoto, Mukunda Sharma and Sungchul Hong for
research assistance, and Robert Dekle, Gabriel Di Bella and conference participants for helpful
comments. West thanks the National Science Foundation for financial support.
1. Introduction
A standard policy prescription in exchange rate crises is to tighten monetary policy, at
least until the exchange rate has stabilized. Indeed, in the East Asian countries whose currencies
collapsed in 1997, interest rates were raised, usually quite dramatically. For example, short-term
rates rose from 12 to 30 percent in the space of a month in December 1997 in South Korea. The
successful recovery from the crisis may seem to vindicate this policy.
But that is not clear. High interest rates weaken the financial position of debtors, perhaps
inducing bankruptcies in firms that are debt constrained only because of informational
imperfections. The countries might have recovered, perhaps with less transitional difficulty, had
an alternative, less restrictive, policy been followed. This has been argued forcefully by, for
example, Furman and Stiglitz (1998) and Radelet and Sachs (1998).
There is mixed empirical evidence on the relationship between interest and exchange
rates, even for developed countries (Eichenbaum and Evans (1995), Grilli and Roubini (1997)).
For countries that have undergone currency crises, Goldfajn and Gupta (1999) found that, on
average, dramatic increases in interest rates have been associated with currency appreciations.
But there was no clear association for a subsample of countries that have undergone a banking
crisis along with a currency crisis. This subsample includes the East Asian countries.
Papers that focus on the 1997 currency crises in East Asia also produce mixed results.
Representative results from papers using weekly or daily data are as follows. Goldfajn and Baig
(1998) decided that the evidence is mixed but on balance favor the view that higher interest rates
were associated with appreciations in Indonesia, Korea, Malaysia, the Philippines and Thailand.
Cho and West (1999) concluded that interest rate increases led to exchange rate appreciation in
Korea during the crisis. Dekle et al. (1999) found sharp evidence that interest rate changes are
2
reduced form predictors of subsequent exchange rate appreciations in Korea, Malaysia and
Thailand, though with long and variable lags. Finally, Gould and Kamin (2000) were unable to
find a reliable relationship between interest rates and exchange rates in the five countries.
This paper conducts an empirical study of the relationship between exchange rates and
interest rates during the 1997-98 exchange rate crises in Korea, the Philippines and Thailand.
Our central question is: in these economies, did exogenous monetary-policy-induced increases in
the interest rate cause exchange rate depreciation or appreciation? Our central contribution is to
propose a model that identifies a monetary policy rule, in a framework general enough to allow
either answer to our central question. Our starting point is the observation that the sign of the
correlation between exchange and interest rates–used in many previous studies to decide whether
an increase in interest rates causes an exchange rate appreciation–will be sufficient to answer our
question only if monetary policy shocks are the dominant source of movements in exchange and
interest rates. Since shocks to perceived exchange rate risk are also arguably an important source
of variability during an exchange rate crisis, one must specify a model that allows one to
distinguish the effects of the two types of shocks.
We do so with a model that has two equations and is linear. One equation is interest
parity, with a time varying risk premium. Importantly, we allow the risk premium to depend on
the level of the interest rate. The second equation is a monetary policy rule, with the interest rate
as the instrument. The two variables in the model are the exchange rate and domestic interest
rate. These two variables are driven by two exogenous shocks, a monetary policy shock and a
shock to the component of the exchange rate risk premium not dependent on the level of the
interest rate. The model has two key parameters. One parameter (“a”) indexes how strongly the
3
monetary authority leans against incipient exchange rate movements. The other parameter (“d”)
indexes the sensitivity of exchange rate risk premia to the level of interest rates.
Whether interest rates should be increased or decreased to stabilize a depreciating
exchange rate depends on how sensitive risk premia are to interest rates. Interest rates should be
increased unless risk premia are strongly increasing with the level of the interest rate. This is the
orthodox policy. Interest rates should be lowered if risk premia are strongly positively related to
the interest rate. This is the view of Furman and Stiglitz (1998). Our model precisely defines
“strongly positive” as meaning that the parameter d referenced in the previous paragraph is
greater than one.
According to our model, the sign of the correlation between exchange and interest rates
suffices to reveal whether exogenous increases in interest rates led to exchange rate appreciation
only if shocks to monetary policy dominate the movement of exchange and interest rates.
Suppose instead that shocks to the exchange rate risk premium are the primary source of
movements in exchange and interest rates. Then in our model, the correlation between the two
variables may be positive even if, in the absence of risk premium shocks, increases in interest
rates would have stabilized a depreciating exchange rate (i.e., d<1). (We measure exchange
rates so that a larger value means depreciation. Thus a positive correlation means that high
interest rates are associated with a depreciated exchange rate.) And the correlation between the
two may be negative even if interest rate increases would have destabilized exchange rates (i.e.,
d>1) in the absence of risk premium shocks.
Using a special case of our model, we find that exchange rate risk premia in Korea were
inversely related to the level of interest rates. In the Philippines, risk premia were increasing in
4
interest rates, though modestly so. In both these countries, stabilization required raising interest
rates. In Thailand, on the other hand, risk premia were strongly increasing, in the precise sense
that the parameter referenced in the preceding paragraph was estimated to be greater than 1.
Accordingly, ceteris paribus, an exogenous increase in the interest rate led to exchange rate
appreciation in Korea and the Philippines, exchange rate depreciation in Thailand.
Unfortunately, confidence intervals for model parameters are huge. They do not rule out
the possibility that interest rate increases led to depreciation in Korea and the Philippines, to
appreciation in Thailand. To a certain extent this seems to follow unavoidably from the fact that
our sample sizes are small, as is suggested by the similarly weak evidence found in most of the
papers cited above. A second reason our results are tentative is that for tractability and ease of
interpretation we base our inference particularly simple assumptions about the behavior of
unobservable shocks. These assumptions are roughly consistent with the data, but alternative,
more complex models no doubt would fit better. As well, we use an inefficient estimation
technique. A final reason our results are tentative is that we do not allow for the possibility of
destabilizing monetary policy, i.e., a period during which a monetary authority moved interest
rates in a destabilizing direction, perhaps before adopting a policy that ultimately led to exchange
rate stabilization. We leave all such tasks to future research.
We also leave to future research the larger, and more important issue, about what
constitutes good policy in an exchange rate crisis. High interest rates may be bad policy even if
they stabilize exchange rates, and may be good policy even if not. We believe that our paper
contributes to our understanding the larger issue, since any policy analysis must take a stand on
the interest rate - exchange rate relationship. In our own work, brief discussions of policy during
5
the Korean crisis may be found in Cho and Hong (2000) and Cho and West (1999).
Section 2 describes our model, section 3 our data, section 4 our results. Section 5
concludes. An Appendix contains some technical details.
2. Model
Our simple linear model has three equations and two observable variables. The three
equations are interest parity, a relationship between exchange rate risk and interest rates, and an
interest rate reaction function (monetary policy rule). The two variables are the domestic interest
rate and the exchange rate.
We write interest parity as:
(2-1) it = i*
t + Etst+1 - st + dt.
In (2-1), it and i*
t are (net) domestic (i.e., Asian) and foreign nominal interest rates; st is 100 ×
log of the nominal spot exchange rate, with higher values indicating depreciation; Et denotes
expectations; dt is a risk premium. If dt/0, (2-1) is uncovered interest parity. The variable dt,
which may be serially correlated, captures default risk as well as the familiar premium due to risk
aversion.
It presumably is safe to view i*
t as substantially unaffected by domestic (Asian) monetary
policy. The same cannot be assumed for Etst+1, st and dt, all of which are determined
simultaneously with it. But for the moment we follow some previous literature (e.g., Furman and
Stiglitz (1998)) and perform comparative statics using (2-1) alone. Evidently, if it is increased,
but Etst+1 and dt are unchanged, then st must fall (appreciate): the orthodox relationship. If, as
6
well, increases in interest rates today cause confidence that the exchange rate will stay strong
(i.e., that st+1 will be lower than it would have been in the absence of an interest rate hike), then st
must fall even farther for (2-1) to hold.
However, this channel will be offset insofar as increases in it are associated with increases
in dt. Such a rise may come about because higher interest rates are associated with higher default
rates, or because higher interest rates raise risk premia. This, in turn, may lead to expectations of
depreciation (increase) in st+1. Furman and Stiglitz (1998) argue on this basis that equation (2-1)
alone does not tell us even whether increases in it will be associated with increases or decreases
in st, let alone the magnitude of the change.
We agree with this argument. Our aim is to specify a model that allows for the
possibility of either a positive or negative response of st to an exogenous monetary policy
induced increase in it, and then to estimate the model to quantify the sign and size of the effect.
To that end, we supplement the interest parity condition (2-1) with two additional equations. The
first is a simple monetary policy rule. We assume that the nominal interest rate is the instrument
of monetary policy. During a period of exchange rate crisis, the focus of monetary policy
arguably is on stabilizing the exchange rate. We therefore assume:
(2-2) it = a(Et-1st - -
st)+ ~
umt.
In (2-2), a is a parameter, and -
st is the target exchange rate. Conventional interpretation
of IMF policy is that the IMF argues for a>0. This means that the monetary authority leans
against expected exchange rate depreciations. Of course, a<0 means that the monetary authority
lowers the interest rate in anticipation of depreciation. For simplicity, we impound the target
7
level into the unobservable disturbance ~
umt. Upon defining umt=~
umt-a-
st, equation (2-2) becomes
(2-2)' it = aEt-1st + umt.
The variable umt, which may be serially correlated, captures not only changes in the target level
of the exchange rate, but all other variables that affect monetary policy. Ultimately it would be
of interest to model umt's dependence on observable variables such as i*
t and the level of foreign
reserves; once again, we suppose that in the crisis period it is reasonable to focus on the
exchange rate as the dominant determinant of interest rate policy. The “exogenous monetary
policy induced increase in it” referenced in the previous paragraph is captured by a surprise
increase in umt.
Note the dating of expectations: period t expectations appear in (2-1), period t-1
expectations in (2-2)'. This reflects the view that asset market participants, whom we presume to
be setting exchange rates, react more quickly than does the monetary authority to news about
exchange rate risk premia (i.e., to shocks to the variable that we call udt below). Capturing this
view by using t-1 expectations in the monetary rule is most appealing when data frequency is
high. Accordingly, we assume daily decision making, and allow for the effects of time
aggregation when we estimate our model using weekly data. Of course, we do not literally
believe that in setting the interest rate each day the monetary authority is ignorant of intra-day
developments. Rather, we take this as a tractable approximation.
The final equation is one that relates the risk premium dt to the interest rate it:
(2-3) dt = dit + ~
udt.
8
Equation (2-3) is an equilibrium relationship between risk premia and interest rates. In the
conventional view, d<0, in which case higher interest rates are associated with lower risk, or
perhaps d=0, in which case there is no link between interest rates and risk. The d<0
interpretation seems consistent with Fischer (1998,p4), who argues that temporarily raising
interest rates restores confidence. In an alternative view, such as that of Furman and Stiglitz
(1998), d>>0, and higher interest rates are associated with higher risk. We suppose that d is
structural, in the sense that one can think of d as remaining fixed while one varies the monetary
policy reaction parameter a. Obviously this cannot hold for arbitrarily wide variation in a, but
perhaps is a tolerable assumption for empirically plausible variation in a.
The variable ~
udt captures all other factors that determine the risk premium. Ultimately it
would be of interest to partially proxy ~
udt with observable variables. Candidate variables include
the level of reserves and debt denominated in foreign currency (see Cho and West (1999) for the
role such variables played in Korea). But since such data are not available at high frequencies,
for simplicity we treat ~
udt as unobservable and exogenous.
To simplify notation, and for consistency with our empirical work, we impound i*
t in the
unobservable disturbance to interest parity, defining udt = i*
t + ~
udt. We then combine (2-3) and
(2-1) to obtain
(2-4) (1-d)it = Etst+1 - st + udt.
Equations (2-2)' and (2-4) are a two equation system for the two variables it and st. Upon
substituting (2-2)' into (2-4) and rearranging we obtain
9
(2-5) st + a(1-d)Et-1st = Etst+1 + udt - (1-d)umt.
Equation (2-5) is a first order stochastic difference equation in st. To solve it, we assume
homogeneous and model-consistent expectations. That is, we assume that private sector and
government expectations are consistent with one another, in that the variables used in forming
Et-1 in (2-2) and (2-2)' are the period t-1 values of the period t variables used in forming Et in
(2-1) and (2-4). Moreover, these expectations are consistent with the time series properties of udt
and umt. To make these assumptions operational, we assume as well that Et denotes expectations
conditional on current and lagged values of udt and umt (equivalently, current and lagged values
of st and it).
Define b=[1+a(1-d)]-1. We make the stability assumption 0<b<1 and the “no bubbles”
assumption lim j64 b jEt-1st+j = 0. The stability assumption requires
(2-6) a<0, d>1 OR a>0, d<1.
The algebraic condition (2-6) captures the following common-sense stability condition. Suppose
risk premia are so sensitive to interest rates that d>1. Stability then requires that the monetary
authority lower interest rates (a<0) in response to anticipated depreciations. For if it instead
raised interest rates, we'd have the following never ending spiral: a positive shock to the risk
premium causes exchange rates to depreciate, which with a>0 causes the monetary authority to
raise interest rates, which with d>1 causes a further depreciation, and a further raising of interest
rates, .... Similarly, if d<1, stability requires increasing interest rates in the face of anticipated
depreciation. Note that one can have a stable system when a>0 even if d>0, as long as d<1: in
10
our model, a policy of leaning against exchange rate depreciations (a>0) is stable even if
increases in interest rates are associated with increased risk (d>0), as long as the increase in risk
is not too large (d<1).
To solve the model, project both sides of (2-5) onto period t-1 information, and then solve
recursively forward. The result is
(2-7) Et-1st = b3j
4
=0{b jEt-1[udt+j
-(1-d)umt+j]}.
For given processes of udt and umt, we can solve for Et-1st using (2-7). Putting this solution into
(2-2)' yields it, which in turn may be used in (2-4) to solve for st.
The data we use are to a certain extent consistent with a random walk for both umt and udt,
say
(2-8) umt=umt-1+emt, udt=udt-1+edt.
Such shocks make for quick, one period, movements from one steady state to another in response
to a shock. They are special in other ways as well, as noted below. Under the assumption that
emt and edt are uncorrelated with one another, Figures 1 to 4 plot responses of it and st to 1%
increases in emt and edt, for each of four parameter sets: a=.2, d=-9; a=.7, d=-9; a=.7, d=.6; a=-.5,
d=1.2.
Figure 1 plots the response of it to a 1% increase in emt. Only one line is plotted because
for all four parameter sets, response is identical. As is obvious from equation (2-2)', the impact
response is a 1% increase. The interest rate then returns to initial value. That is, a permanent
11
increase in umt leads to a transitory change in it. Evidently, from (2-2)', in steady state st must
fall by (1/a) (rise by -1/a when a<0). This depicted in Figure 2. Consider first the case where
a>0. Then an exogenous increase in the interest rate causes an exchange rate appreciation: with
a>0, exogenous increases in interest rates stabilize a depreciating exchange rate. In the three
specifications with a>0, the impact elasticity ranges from about -2 to -15. For given d, the
impact effect is smaller for a=.7 than for a=.2: larger a means a harsher monetary policy
response and greater exchange rate stability. On the other hand, when a<0, an exogenous
increase in the interest rate causes the exchange rate to depreciate.
These long responses are of course consistent with long run neutrality of monetary policy.
An increase in emt means a commitment to raise the interest rate for any given expected level of
exchange rates, now and forever. Because the level of the exchange rate adjusts in the long run,
there is no long run effect on the rate of exchange rate depreciation, and therefore no long run
effect on the level of the interest rate.
Figure 3 depicts the response of st to a 1% increase in the risk premium. In all
specifications, the exchange rate increases in the short and long run. The impact effect is greater
than the long run effect because according to (2-2)' it takes a period before interest rates respond
to the increased risk. For given d, the response is less for larger a; for given a, the response is
greater for larger d.
Figure 4 plots the response of it to a 1% increase in the risk premium. By assumption,
there is no contemporaneous response. When a>0, the interest rate is increased; when a<0, it is
decreased. When a is larger in absolute value, there is a larger increase. In accordance with
(2-4), the long run response of it is 1/(1-d), and thus is governed only by d but not a; in the
simple random walk specification, the long run is achieved in one period and so the responses for
12
a=.2/d=-9 and a=.7/d=-9 are identical.
Some implications of the above are worth noting. First, upon comparing the figures, we
see that when a>0, risk premium shocks cause interest and exchange rates to move in the same
direction, while monetary shocks cause them to move in opposite directions. For a<0, risk
premium shocks cause interest and exchange rates to move in the opposite direction, while
monetary shocks cause them to move in the same direction. This result holds not only for
random walk shocks but also for arbitrary stationary AR(1) shocks.
The implication is that the sign of the correlation between interest and exchange rates is
not sufficient to tell us that interest rate hikes stabilized a depreciating currency. A negative
correlation may result when a<0 because the data are dominated by risk premium shocks. A
positive correlation may result when a>0 because the data are dominated by risk premium
shocks.
Second, suppose one takes a as a choice parameter for a monetary authority that aims to
stabilize a rapidly depreciating exchange rate. If exchange rate risk does not rapidly increase
with the level of interest rates (d<1), then the monetary authority should raise interest rates (set
a>0) when further depreciation is expected. But if exchange rate risk does rapidly increase with
the level of interest rates (d>1), then the monetary authority should lower interest rates (set a<0)
when further depreciation is expected. In either case, stabilization smooths exchange rates.
A third point is that with random walk shocks–an assumption we maintain in our
empirical work–this stabilization of exchange rates will induce a negative first order
autocorrelation in )st. That is, smoothing in the face of random walk shocks causes exchange
rates to exhibit some mean reversion relative to a random walk benchmark. (Our model is
capable of generating positive autocorrelation in )st, but only if the shocks exhibit dynamics
13
beyond that of a random walk.)
Finally, with random walk shocks, one can read the sign of 1-d, and hence whether d is
above or below the critical value of 1, directly from the sign of the correlation between )it and
)st-1. When d is less than 1, this correlation is positive; when d is greater than 1, this correlation
is negative: if stabilization involves increasing (decreasing) interest rates in response to incipient
exchange rate depreciations, then, naturally, )it will be positively (negatively) correlated with
)st-1. Again, this simple result applies because we assume random walk shocks and need not
hold for richer shock processes.
3. Data and Estimation Technique
We obtained daily data for Korea, the Philippines and Thailand, either directly from
Bloomberg or indirectly from others who reported Bloomberg as the ultimate source. The
mnemonics for exchange rates: KRW (Korea ), PHP (Philippines) and THB (Thailand). The
mnemonics for interest rates: KWCR1T (Korea), PPCALL (Philippines: Philippine Peso
Interbank Call Rate) and BITBCALL (Thailand: Thai STD Chartered Bank Call Rate). Because
many days were missing, we constructed weekly data by sampling Wednesday of each week. If
Wednesday was not available we used Thursday; if Thursday was not available we used
Tuesday. Interest rates are expressed at annual rates; exchange rates are versus the U.S. dollar.
We start our samples so that we are two weeks into what arguably can be considered the
post-crisis exchange rate targeting regime. Two weeks allows both the current and lagged value
of interest and exchange rate differences to fall inside the new regime. For Thailand and the
Philippines this means a start date of Wednesday, July 23, 1997. (As noted above, our weekly
data is for Wednesday.) For Korea, the date is Wednesday, December 17, 1997. We ended our
14
samples one year later (sample size of 53 weeks), since the simple monetary rule (equation 2-2)
probably did not well describe policy once the countries had stabilized. We also tried 27 week
samples, with little change in results. Figures 5 to 7 plot our data, in levels. The dashed lines
delimit our one year samples.
Formal unit root tests failed to reject the null of a unit root. Hence we examine interest
and (log) exchange rates in first differences. We failed to find cointegration between it and st.
(Using similar weekly data, Gould and Kamin (2000) and Dekle et al. (1999) also failed to find
cointegration.) Hence in our regression work mentioned briefly below we estimated a VAR in
)it and )st without including an error correction term. And we note in passing that the lack of
cointegration meant that we could not turn to estimation of a cointegrating vector to identify the
monetary policy parameter a.
To identify a and d, we assume that umt and udt follow random walks. In this case, our
model implies a vector MA(1) process for ()it,)st)', which, as explained in the next section, is
more or less consistent with our data. We allow the innovations in umt and udt to be
contemporaneously correlated. Such a correlation might result, for example, if the level of
foreign reserves importantly affected both monetary policy and exchange rate risk. Because we
allow this correlation, it will not be meaningful to decompose the variation of exchange or
interest rates into monetary and risk components. (We do not, however, model or exploit
cross-country correlations in udt or umt, deferring to future work the attractive possibility of using
information in such correlations.) We allow for decisions to be made daily rather than weekly.
That is, we assume that the model described in section 2 generates the data with a time period
corresponding to a day. But we only sample the data once every five observations.
We use five moments to compute the five parameters a and d and the three elements of
15
the variance-covariance matrix of (emt, edt)'. The moments we used included three chosen
because they were estimated relatively precisely: var()it), var()st) and corr()it, )st). The final
two moments used, corr()it,)st-1) and corr()st,)st-1), were largely chosen for clarity and
convenience. As explained at the end of section 2 above, our model has simple and direct
implications for the signs of these correlations. And as a technical matter, with this choice of
moments, the parameters could be solved for analytically, though the equations are nonlinear.
An appendix gives details on how we mapped moments into parameters. Two points
about the mapping are worth noting here. The first is that since the five equations are nonlinear,
in principle they can yield no reasonable solutions. For example, for a given set of moments, the
implied value of the variance of edt might be negative. The second is that our algorithm solves
for a from a root to a quadratic. If the estimated first order autocorrelation of )st is between -0.5
and 0, this quadratic is guaranteed to have two real roots, one implying a positive value of a, the
other a negative value. We chose the root consistent with stability: the root implying a positive
value of ^
a if ^
d<1, a negative value if ^
d<1. (The solution algorithm is in part recursive, with d
estimated prior to a.) We made this choice because an unstable solution implies explosive data,
at least if the unstable policy is expected to be maintained indefinitely; this is inconsistent with
our use of sample moments.
We report 90 percent confidence intervals. These are “percentile method” intervals,
constructed by a nonparametric bootstrap using block resampling with nonoverlapping blocks.
Details are in the Appendix.
4. Empirical Results
Table 1 has variances and auto- and cross-correlations for lags 0, 1 and 2, with the
16
bootstrap confidence intervals in parentheses. A skim of the Table reveals that virtually all the
auto- and cross-correlations are insignificantly different from zero at the ten percent level. The
only exceptions are the correlations between )st and )it-1 in the shorter sample in Korea,
between )it and )it-1 in the longer sample in the Philippines, and between )it and )st in both
samples in Thailand. (We did not report confidence intervals for var()it) and var()st) in Table 1;
all point estimates of these variances were significant at the 90 percent level–indeed, at any
significance level–by construction.)
The insignificance of the point estimates at lag 2 is consistent with a vector MA(1)
process for ()it,)st)', since population auto- and cross-correlations will all be zero for lags 2 and
higher for such a process. This is the main sense in which a random walk for umt and udt implies
a process more or less consistent with the data. As well, the estimates of the first order
autocorrelation of )st is negative in all samples, though barely so for the Philippines and
Thailand in the one year samples (point estimates = -0.07 and -0.02); as noted in section 2 above,
a negative autocorrelation is implied by our model if shocks are random walks.
On the other hand, the insignificance of the point estimates at lag 1, and of the
contemporaneous correlation between )it and )st in Korea and the Philippines is bad news for
our MA(1) model, and, in our view, for any empirical study of these data. Since the data are
noisy, estimates of model parameters–which of course will be drawn from moments such as
those reported in Table 1–will likely be imprecise. That, perhaps, is an inevitable consequence
of our decision to focus on a sample small enough that it a priori seemed likely to have a more or
less stable interest rate rule.
We note in passing that when a second order VAR in ()it, )st) is estimated for one year
samples, F-tests (not reported in the table) yield slightly sharper results. Specifically, the null of
17
no predictability is rejected for lagged interest rates in the )it equation in the Philippines and for
lagged exchange rate changes in both the )it and the )st equations in Korea but not otherwise.
This suggests the importance of allowing for richer dynamics in the shocks, an extension
suggested as well by the fact that the absolute value of the Philippine estimate of corr()it,)it-1) is
greater than 0.5, a magnitude inconsistent with )it following a MA(1) process. We leave that as
a task for future research.
Using the algorithm described in the Appendix and the previous section, we estimated a
and d from some moments reported in Table 1. (The algorithm also automatically produces
estimates of the variance-covariance matrix of (emt, edt)', which we do not discuss because these
are not of economic interest.) Columns 3 and 4 in Table 2 present these estimates, again with 90
percent confidence intervals from a bootstrap given in parentheses. The algebraic values of the
estimates of d are lowest for Korea and highest for Thailand, with
(4-1) ^
d for Korea < 0 < ^
d for Philippines < 1 < ^
d for Thailand.
The implication is that in equilibrium, increases in interest rates were associated with decreases
in exchange rate risk in Korea. The association between interest rates and exchange rate risk was
positive in the Philippines, but sufficiently small that if monetary policy was to be stabilizing,
interest rates must be increased in response to expected exchange rate depreciations (a>0). The
association is also positive in Thailand, with the estimated value of d greater than one. Hence if
monetary policy was to be stabilizing in Thailand, interest rates must be decreased in response to
expected exchange rate depreciations (a<0). As explained above, the signs of ^
d follow from the
signs of the estimates of the correlation between )it and )st-1: negative in Thailand, positive in
18
Korea and the Philippines.
As we feared, the confidence intervals on the estimates of a and d are large; indeed, they
are staggeringly large. Using a two-tailed test, one can reject the null that a=0 in Korea in the
one year sample at the16% level (not reported in the table); all other parameters are even more
imprecisely estimated.
Let us abstract from the confidence intervals and focus on the point estimates. We do not
know of estimates from other studies that can be used to directly gauge the plausibility of the
estimate of d. This ranking does conflict with Barsuto and Ghosh (2000), who concluded that
real interest rate hikes increased the exchange rate risk premium in Korea, decreased it in
Thailand. (Barsuto and Ghosh (2000) did not study the Philippines.) On the other hand, it is our
sense that the ranking in (4-1) accords with the view that fundamentals were best in Korea, worst
in Thailand. And the bottom line conclusion–that interest rate increases caused depreciation in
Thailand, appreciation in Korea and the Philippines–is consistent with Goldfajn and Baig (1998,
Table 3, full sample estimates) and with Di Bella's (2000) findings for Thailand (Di Bella (2000)
does not consider other Asian countries).
For all practical purposes, impulse responses to orthogonal movements in emt and edt are
given in Figures 1 to 4. For Korea, see the lines for a=0.2, d=-9; for the Philippines, see a=0.7,
d=0.6; for Thailand, see a=-0.5, d=1.2. The exact responses of st to a 1% positive value of emt
are given in columns (5) and (6) of Table 2. Once again, the confidence intervals are very large,
as is inevitable since these elasticities are simple transformations of the estimates of a and d.
Now, Thailand's agreements with the IMF called for it to maintain interest rates in
indicative ranges that were high relative to pre-crisis levels (for example, 12-17 percent in the
August 1997 agreement (IMF (1997a, Annex B)), 15-20 percent in the December 1997
19
agreement (IMF (1997b, Annex B)). As well, some agreements suggested raising interest rates
when the exchange rate is under pressure (IMF (1997b, 1998)). How can this be reconciled with
our Thai estimates (^
a<0, ^
d>1), which indicate that the stabilization was accomplished by
lowering interest rates in the face of incipient depreciation? One interpretation is that
IMF-increases appear in our data as occasional and very visible large positive values of umt; most
of the day-to-day systematic component of policy implicitly lowered interest rates in the face of
incipient exchange rate depreciation, despite the agreement to raise interest rates when the
exchange rate was pressured. On this interpretation, the appreciation would have occurred
sooner absent the early increases in interest rates. A second interpretation is that policy did raise
interest rates in the face of depreciation, both in the form of one-time increases early in the
sample, and systematically throughout the sample. But sampling error caused the estimate of d
to be greater than 1 and thus the estimate of a to be negative. (We refer to ^
d rather than ^
a
because ^
a is solved from a quadratic with one negative and one positive root, and we choose the
root consistent with stability: the root that yields ^
a<0 when ^
d>1, the root that yields ^
a>0 when
^
d<1. See section 3 and the Appendix.)
We do not have any direct evidence on either of these interpretations. We hoped that
some indirect evidence might be found by rolling the samples forward, recomputing the
estimates of a and d. Table 3 presents results of such an exercise, for one year samples, and for
all three countries. We dropped the initial observation as we added a final observation, keeping
the sample at T=53 weeks. In Korea and Thailand, we stopped the process when the estimated
first order autocorrelation of )st turned positive. That date does not occur until January 1998 for
the Philippines, and so to conserve space we stopped at September 1997.
The estimates for the Philippines and Thailand move little–surprisingly little, in light of
20
the huge confidence intervals in the previous table. In the Philippines, the estimate of d ranges
from about 0.5 to 0.7; in Thailand, the range is about 1.1 to 1.4. As well, the estimate of a does
not fall, which one might expect if Thailand systematically raised interest rates in response to
incipient exchange rate depreciation in the early but not the later parts of the sample. Thus this
exercise is not particularly helpful in interpreting the results for Thailand.
One estimate that is quite sensitive to the sample is that for d, for Korea. The estimated
value rises rapidly, from -8.99 to -0.36. A possible rationalization of this pattern is that as a
country stabilizes, exchange rate risk becomes insensitive to the level of the interest rate.
Perhaps d=0 in developed countries, or at least in countries without credit rationing (see Furman
and Stiglitz (1998)). But clearly this is a speculative interpretation, and the large confidence
intervals in Table 2 make it reasonable to attribute the wide variation to sampling error in
estimation of d.
5. Conclusions
We have formulated and estimated a model that allows for interest rate shocks to either
appreciate or depreciate exchange rates. Using weekly data, we have estimated a special case of
the model using data from Korea, the Philippines and Thailand. We have found that an
exogenous increase in interest rates caused exchange rate appreciation in Korea and the
Philippines, depreciation in Thailand. The estimates are, however, quite noisy.
One set of priorities for future work is to use higher frequency data, allow for richer
shock processes, and use more efficient estimation techniques. A second is to allow for the
possibility that for some period of time, monetary policy was destabilizing, with a switch in the
sign of the interest rate reaction function necessary for stabilization. A third is to bring
21
additional variables, such as the level of foreign reserves, into the model. A final, and broad, aim
of our future work is to use our knowledge of the relationship between interest rates and
exchange rates to analyze the macroeconomic effects of monetary policy in countries undergoing
currency crises.
22
Appendix
Mapping from moments to model parameters: Let udt and umt follow random walks
(A-1) udt = udt-1 + edt, umt = umt-1 + emt,
where edt and emt are vector white noise. Then the solution of the model is
(A-2) it = (1-d)-1udt-1 + emt, st = -(a-1+1-d)umt + (1-d)umt-1 + [1+(1-d)-1a-1]udt - udt-1.
Define ~
udt = (1-d)-1udt, ~
edt = (1-d)-1edt, *=1-d, "=a-1. Rewrite (A-2) as
(A-3) it = ~
udt-1 + emt, st = ("+*)(~
udt-umt) - *(~
udt-1-umt-1)
Suppose we sample data every n periods (n=5 in the computations in the text). Then
(A-4a) it - it-n = ~
edt-1 + ~
edt-2 + ... + ~
edt-n + emt - emt-n,
A-4b) st - st-n = ("+*)(~
edt-emt) + "(~
edt-1-emt-1) + ... + "(~
edt-n+1-emt-n+1) - *(~
edt-n-emt-n).
Define )nit = it-it-n, )nst = st-st-n, ~
Fmd=cov(emt,~
edt), ~
F2
d =var(~
edt), F2
m=var(emt). Then
(A-5) var( )ni) = n~
F2
d + 2F2
m - 2~
Fmd,
(A-6) var( )ns) = (n"2+2"*+2*2)(~
F2
d + F2
m - 2~
Fmd),
(A-7) cov()ni, )ns) = [(n-1)"-*]~
F2
d - ("+2*)F2
m - [(n-1)"-"-3*]~
Fmd,
(A-8) cov()ns, )ns-n) = -*("+*)(~
F2
d + F2
m - 2~
Fmd),
(A-9) cov()ni, )ns-n) = ("+*)(~
F2
d + F2
m - 2~
Fmd).
Equations (A-5) to (A-9) were used to solve for ~
F2
d, F2
m, ~
Fmd, " and *. From these, a and d can be
computed. When cov()ns, )ns-n)<0, a quadratic that is used to solve for " is guaranteed to have
one negative and one positive root. We chose the root consistent with stable monetary policy:
the negative root when the estimate of d was greater than one, the positive root otherwise.
23
Description of bootstrap technique: The bootstrap confidence intervals in Table 2 were based
on 5000 replications of the following procedure. Each replication was based on an artificial
sample constructed by sampling, with replacement, nonoverlapping blocks of size 6, from the
actual data. For the larger sample (T=53 weeks), we sampled the blocks from a sample of 54
weeks. We used 54 rather than 53 weeks so that the sample contained an integral multiple of
blocks; the 54 weeks consisted of the 53 used in the estimation plus an additional week at the end
of the sample (e.g., 12/17/97-12/23/98 for Korea). In the smaller sample (T=27 weeks), we
sampled the blocks from a sample of 30 weeks, adding three weeks to the data used in estimates
reported in the table (e.g., 12/17/97-7/8/98 for Korea).
For each of the 5000 samples, we applied the procedure used to obtain the point
estimates, to samples of size 53 or 27. We sorted the results from lowest to highest. For the
autocorrelations in Table 1, the confidence intervals were obtained by dropping the lowest and
highest 5% of the results (i.e., the 500 lowest and 500 highest). For the point estimates in Table
2, we first dropped all results in which (1)the point estimate of the first order autocorrelation of
)st was positive or less than -.5, or (2)the point estimate of var(edt) or var(emt) was negative. The
confidence intervals were then obtained by dropping the lowest and highest 5%of the remaining
results. The number of observations that remained after dropping those with inadmissable point
estimates were as follows. Korea: 3318 (T=53) and 2746 (T=27); the Philippines 1977 (T=53)
and 2728 (T=27); Thailand 1661 (T=53) and 2285 (T=27). The relative paucity of remaining
observations in the Philippines and Thailand for T=53 results from a relatively large number of
bootstrap samples in which the point estimate of the first order autocorrelation of )st was
positive.
Table 1
------------------------------Correlations----------------------------
Variance )it )it-1 )it-2 )st)st-1 )st-2
A. Korea 12/17/97- )it 2.06 1.00 0.15 -0.28 0.54 0.12 -0.17
12/16/98 (-0.06,0.28) (-0.41,0.17) (-0.01,0.69) (-0.05,0.27) (-0.24,0.20)
)st20.49 0.54 -0.22 0.00 1.00 -0.39 0.21
(-0.01,0.69) (-0.30,-0.05) (-0.56,0.12) (-0.45,0.27) (-0.15,0.27)
12/17/97- )it 3.83 1.00 0.14 -0.33 0.57 0.11 -0.18
6/17/98 (-0.16,0.17) (-0.53,0.04) (-0.14,0.76) (-0.10,0.27) (-0.24,0.38)
)st36.04 0.57 -0.24 0.01 1.00 -0.45 0.24
(-0.14,0.76) (-0.35,-0.06) (-0.68,0.14) (-0.54,0.35) (-0.22,0.33)
B. Philippines 7/23/97- )it144.91 1.00 -0.59 0.30 0.20 0.05 0.10
7/22/98 (-0.75,-0.11) (-0.18,0.42) (-0.11,0.37) (-0.08,0.27) (-0.14,0.33)
)st10.62 0.20 -0.25 0.17 1.00 -0.07 0.01
(-0.11,0.37) (-0.49,0.20) (-0.15,0.33) (-0.11,0.16) (-0.15,0.11)
7/23/97- )it283.87 1.00 -0.61 0.31 0.24 0.06 0.14
1/21/98 (-0.76,-0.14) (-0.21,0.44) (-0.12,0.47) (-0.15,0.33) (-0.19,0.35)
)st15.17 0.24 -0.28 0.22 1.00 -0.18 -0.05
(-0.12,0.47) (-0.64,0.25) (-0.22,0.42) (-0.24,0.25) (-0.27,0.10)
C. Thailand 7/23/97- )it 5.22 1.00 0.11 0.10 0.32 -0.20 -0.00
7/22/98 (-0.30,0.24) (-0.22,0.37) (0.20,0.49) (-0.45,0.08) (-0.19,0.20)
)st16.74 0.32 0.01 0.01 1.00 -0.02 0.09
(0.20,0.49) (-0.10,0.23) (-0.14,0.17 ) (-0.18,0.20) (-0.27,0.16)
7/23/97- )it5.93 1.00 0.31 -0.02 0.36 -0.13 -0.14
1/21/98 (-0.30,0.31) (-0.35,0.19) (0.14,0.56) (-0.60,0.02) (-0.46,0.26)
)st14.58 0.36 -0.08 0.12 1.00 -0.47 0.11
(0.14, 0.56) (-0.29, 0.30) (-0.15,0.23 ) (-0.49,0.22) (-0.33,0.28)
Notes:
1. )it is the weekly change in the overnight interest rate, expressed at annual rates; )st the weekly
percentage change in the exchange rate versus the U.S. dollar. Higher values of st indicate
depreciation.
2. 90% confidence intervals from bootstrap, in parentheses. Point estimates in bold are
significant a the 90% level.
Table 2
Parameter Estimates
(1) (2) (3) (4) (5) (6)
Country Sample ad% response of st to a 1% shock to umt
Impact Long-run
Korea 12/17/97- 0.25 -8.87 -13.9 -4.1
12/16/98 (-0.05,0.35) (-27.7,14.2) (-43.9,36.3) (-15.2,12.1)
12/17/97- 0.36 -11.27 -15.0 -2.8
6/17/98 (-0.12,0.37) (-27.7,39.4) (-46.8,61.7) (-20.6,17.0)
Philippines 7/23/97- 0.76 0.57 -1.8 -1.3
7/22/98 (-1.77,5.76) (-0.72,2.43) (-8.6,7.1) (-6.9,5.9)
7/23/97- 1.12 0.31 -1.6 -0.9
1/21/98 (-2.70,9.04) (-0.92,3.25) (-6.7,7.5) (-3.8,4.9)
Thailand 7/23/97- -0.54 1.16 2.0 1.8
7/22/98 (-1.07,0.19) (-2.25,7.74) (-12.4,21.8) (-9.0,15.1)
7/23/97- -0.96 6.59 6.6 1.0
1/21/98 (-1.33,0.11) (-3.7,14.5) (-13.8,23.4) (-8.9,10.4)
Notes:
1. a is a monetary policy reaction parameter defined in equation (2-2). d measures the sensitivity
of exchange rate risk premia to the interest rate, as defined in equations (2-1) and (2-3).
2. 90% confidence intervals, from bootstrap, in parentheses.
3. The elasticities in columns (5) and (6) are the response to a surprise, permanent 1% increase in
umt.
Table 3
Rolling Sample Estimates of a and d
Korea Philippines Thailand
Start ^
a ^
dStart ^
a ^
dStart ^
a ^
d
12/17/97 0.24 -8.99 7/23/97 0.74 0.57 7/23/97 -0.53 1.16
12/24/97 0.48 -3.37 7/30/97 0.68 0.54 7/30/97 -0.54 1.12
12/31/97 0.41 -1.82 8/6/98 0.68 0.55 8/6/98 -0.54 1.13
1/07/98 0.29 -2.28 8/13/98 0.63 0.55 8/13/98 -0.52 1.14
1/14/98 0.31 -1.93 8/20/98 0.66 0.51 8/20/98 -0.49 1.21
1/21/98 0.35 -1.27 8/27/98 0.70 0.53 8/27/98 -0.53 1.29
1/28/98 0.35 -0.36 9/3/98 0.73 0.55 9/3/98 -0.58 1.39
2/6/98 n.a. n.a. 9/10/98 1.41 0.75 9/10/98 -0.56 1.31
2/13/98 n.a. n.a. 9/17/98 1.41 0.75 9/17/98 -0.59 1.14
2/20/98 n.a. n.a. 9/24/98 1.41 0.73 9/24/98 n.a. n.a.
Notes:
1. The estimates of a and d are computed from 53 week samples with the indicated starting date.
For each country, the estimate in the first line repeats the figures in Table 2.
2. The algorithm used to map data to parameters cannot be used when the estimate of the first
order autocorrelation of )st is positive. The “n.a.” entries flag samples in which the estimate of
this autocorrelation is positive.
References
Basurto, Gabriela and Atish Ghosh, 2000, “The Interest Rate-Exchange Rate Nexus in Currency
Crises,” manuscript, International Monetary Fund.
Cho, Dongchul and Kiseok Hong, 2000, “Currency Crisis of Korea: Internal Weakness or
External Independence?”, 3-54 in I. Shin (ed.) The Korean Crisis: Before and After, Seoul:
Korea Development Institute.
Cho, Dongchul and Kenneth D. West, 1999, “The Effect of Monetary Policy in Exchange Rate
Stabilization in Post-Crisis Korea,” manuscript, Korea Development Institute.
Dekle, Robert, Cheng Hsiao, and Siyan Wang, 1999, “Interest Rate Stabilization of Exchange
Rates and Contagion in the Asian Crisis Countries,” manuscript, University of Southern
California.
Di Bella, Gabriel, 2000, “Exchange Rate Stabilization in Mexico and Thailand: Was the Tight
Monetary Policy Effective,” manuscript, University of Wisconsin.
Eichenbaum, Martin and Charles L. Evans, 1995, “Some Empirical Evidence on the Effects of
Shocks to Monetary Policy on Exchange Rates,” Quarterly Journal of Economics, 110(4),
975-1009.
Fischer, Stanley, 1998, “The Asian Crisis: A View from the IMF,”
www.imf.org/external/np/speeches/1998/012298.htm.
Furman, Jason and Joseph E. Stiglitz, 1998, “Economic Crises: Evidence and Insights from East
Asia,” Brookings Papers on Economic Activity, 1-135, Vol.2.
Goldfajn, Ilan and Taimur Baig, 1998, “Monetary Policy in the Aftermath of Currency Crises:
The Case of Asia”, IMF Working Paper 99-42.
Goldfajn, Ilan and Poonam Gupta, 1999, “Does Monetary Policy Stabilize the Exchange Rate
Following a Currency Crisis?”, IMF Working Paper 99-42.
Gould, David M. and Steven B. Kamin, 2000, “The Impact of Monetary Policy on Exchange
Rates During Financial Crises,” Board of Governors of the Federal Reserve System International
Financial Discussion paper No. 669.
Grilli, Vittorio and Nouriel Roubini, 1996, “Liquidity Models in Open Economies: Theory and
Empirical Evidence,” European Economic Review 40(3-5), 847-59.
International Monetary Fund, 1997a, “Thailand Letter of Intent, August 14, 1997,”
www.imf.org/external/np/loi/081497.htm.
International Monetary Fund, 1997b, “Thailand Letter of Intent, November 25, 1997,”
www.imf.org/external/np/loi/112597.htm.
International Monetary Fund, 1998, “Thailand Letter of Intent and Memorandum on Economic
Policies, February 24, 1998,” www.imf.org/external/np/loi/022498.htm.
Radelet, Steven and Jeffrey D. Sachs, 1998, “The East Asian Financial Crisis: Diagnosis,
Remedies, Prospects,” Brookings Papers on Economic Activity, 1-74, Vol.1.
