Foreign Reserve Accumulation and the Mercantilist
Motive Hypothesis: A Latent Factor Approach∗
Patrick Carvalho†‡ and Renée A. Fry-McKibbin†
†
ANU Crawford School of Public Policy –
Centre for Applied Macroeconomic Analysis
‡
ANU Research School of Economics
May 2013
Abstract
A fivefold increase in central bank foreign reserves across the globe over the
past fifteen years has prompted questions whether it constitutes a new form of
mercantilism. According to this view, countries accumulate foreign reserves in
order to support export promotion by influencing exchange rates and/or to signal
economic strenght as a modern version of bullionism. Using a unique dataset on
daily foreign exchange intervention, this paper investigates the mercantilist motive hypothesis in the case of Brazil. The findings support reserve accumulation
as a by-product of successful central bank intervention in the Brazilian foreign
exchange market. Results also indicate regional currency intervention spillovers
on Brazil’s neighbouring countries, opening room for further research.
Keywords: Foreign exchange intervention, currency intervention, exchange rate
volatility, reserve accumulation, factor model, emerging markets
JEL Classification: F31, F36, F41
∗
Corresponding author is Renée McKibbin: [email protected]. The authors express gratitute
to the Central Bank of Brazil for the data on the currency intervention in the Brazilian foreign
exchange market.
Part of the literature understands the fivefold increase in global central bank foreign
reserves (hereafter called reserves) over the past fifteen years as a by-product of a new
mercantilist approach. In a series of papers from the US National Bureau of Economic
Research, Dooley et al. (2003, 2004b,a,c, 2005a,c,b, 2007, 2008, 2009) have described a
current ‘Revived Bretton Woods’ system (BWII). The BWII would consist of some developing countries using currency intervention, and consequent reserves accumulation,
as a methodical way of affecting national currency levels to support export promotion.
Aizenman and Lee (2007), on the other hand, have disputed the so-called BWII framework, minimising the relationship between reserves hoarding and the goal to depreciate
(or retard appreciation of) national currencies. Indeed, in accordance to this critique,
the main drive for recent reserve build-up would be to prevent or mitigate currency
crisis – branded as the insurance motive hypothesis (Durdu et al., 2007; Obstfeld et al.,
2010; Jeanne and Ranciere, 2011; Calvo et al., 2012).
This paper investigates the empirical evidence on the mercantilist motive hypothesis, broadly defined as reserve accumulation with the intent to favour export promotion
by influencing exchange rates and/or to signal economic strenght as a modern version
of bullionism. Using a unique daily dataset on Brazil’s foreign exchange intervention,
the paper tests the link between reserve accumulation and exchange rate volatility as
well as the consequent intervention spillover effects into neighbouring countries.
The paper is organised in two sections. Section 1 explores the link between reserve
accumulation and exchange rate volatility in Brazil. Using Brazil’s central bank currency intervention data, it deploys a latent factor model to account for the volatility
decomposition in the currency market and in the reserve changes of Brazil’s major
trade partners. The model tests for the efficacy of Brazil’s central bank intervention
and assesses if Brazil’s reserve changes could be used as a good proxy for currency intervention. The results confirm both the efficacy of Brazil’s currency intervention and
the use of reserve changes as a proxy for intervention. This provides further evidence
of the mercantilist motive hypothesis for reserve accumulation. Moreover, results also
point to regional spillover effects of Brazil’s currency intervention into neighbouring
countries.
Section 2 investigates further the regional intervention spillover effects on reserves
accumulation as a result of mercantilist motives. A latent factor model explores the
1
empirical evidence of reserve stock co-movements between neighbouring countries due
to deliberate central bank intervention on foreign exchange market. In particular,
the model investigates the impact of Brazil’s central bank intervention on the volatility decomposition of reserve changes in Brazil, Argentina, Chile and Peru. The results confirm the impact of Brazil’s central bank intervention, and consequent reserve
changes, into the volatility decomposition of the reserve stock movements in neighbouring countries. Moreover, parameter estimates do not support the hypothesis of reserve
changes as a result of a modern bullionism practice; leaving co-movements in currency
intervention across neighbouring countries to be the driving force behind such regional
intervention spillovers.
1
The link between foreign reserves and exchange
rate volatility
This section addresses the empirical relationship between reserves and exchange rate
volatility in Brazil. Brazil is an interesting case given its size (the 6th largest economy
in the world), importance (leading emerging country, part of the BRIC group) and its
claim to hold a floating exchange rate. If evidence is found to support the link between
reserves accumulation and currency intervention in an emerging country outside the
fixed-exchange-rate realm, the case for the mercantilist motive would be made stronger.
Lately, foreign exchange interventions in emerging markets have gained much attention in the literature, due to the relatively bigger size of their central banks in the
domestic economy and a more prominent role of developing nations in the global trade
(Menkhoff, 2012). Yet, monetary authorities in emerging economies usually downplay
the use and intent of exchange rate interventions, making it difficult to classify the
level of currency floating (Calvo and Reinhart, 2002). In Brazil, for instance, a flexible
exchange rate regime has been officially in place since 1999; despite a prevailing de
facto unofficial dirty floating management.
Since 2004, Brazil has received vast amounts of foreign currency through trade surpluses, foreign direct investments and international portfolio inflows. The advocates
for the mercantilist motive point to the fact that the accompanying large build-up of
reserves in the same period has been used to contain the volatility and the appreciated level of the Brazilian effective exchange rate. If this argument holds, evidence of
the currency intervention effectiveness through reserve accumulation should be found.
2
This section, therefore, evaluates whether central bank interventions have a greater
impact on currency returns in Brazil, and regarding the mercantilist motive hypothesis, if Brazil’s reserve changes could be used as a proxy to its central bank currency
intervention in determining the variations in the Brazilian exchange rate.
A latent factor model is estimated to decompose the contribution of Brazil’s central
bank intervention to the overall volatility in the currency market. Accordingly, results
support the effectiveness of Brazil’s currency intervention during the sample period
(May, 2009 to June, 2012). Besides, the mercantilist motive hypothesis is also validated:
Brazil’s reserve changes contitutes a good proxy for currency intervention. Benchmark
results show that currency intervention, or reserve changes as a proxy, accounts for 6–
7% of the volatility in the Brazilian currency. Lastly, it is noted that Brazil’s currency
intervention has spillover effects in Argentina and other Latin American countries,
which invites further research.
This section is organised as follows: Subsection 1.1 presents a preliminary analysis
of the data; Subsection 1.2 introduces the model specification; Subsection 1.3 describes
the estimation method; and Subsection 1.4 examines the results.
1.1
Data
This section presents a preliminary analysis of the data used to test the effectiveness
of Brazil’s currency intervention and the mercantilist motive hypothesis. The data
includes exchange rate returns, first-difference of reserve stocks (reserve changes) and
central bank currency intervention in the Brazilian foreign exchange market. The
sample is daily, extending from May 4, 2009 to June 29, 2012. This chosen period is
due to data availability, in particular, the data sample is constrained by the daily time
series for Brazil’s central bank currency intervention, kindly disclosed by the Central
Bank of Brazil.
The selection of exchange rates corresponds to Brazil’s major trade partners, accounting for almost two-thirds of Brazilian external trade, namely, the European Union
(21.7% of the total sum of Brazil’s exports and imports), China (14.9%), United States
(12.5%) and Argentina (8.7%).1 In this regard, exchange rate variables used in the
model comprise the euro, the British pound, the Argentine peso and the Brazilian real,
all expressed against the US dollar. The Chinese yuan was excluded due to China’s
fixed exchange rate, which defeats the purpose to use it in a factor model of variance
1
Source: IMF (DoTS), 2010.
3
decomposition. This combination of foreign exchange rates also captures potential differences between developed and developing economies as well as common geopolitical
factors in South America – and more specifically to the common market of MERCOSUR, which Brazil and Argentina are key members. Figure 1 shows the daily exchange
rates and the respective percentage returns against the US dollar between May 2009
and June 2012. An increase (decrease) in the value of the exchange rates indicates a
depreciation (appreciation) of the local currency against the US dollar. All exchange
rate returns are computed by taking the first difference of the natural logarithm of the
exchange rates, and multiplied by 100.
In order to capture common factors related to reserve changes in emerging economies,
Brazil’s and Argentina’s foreign reserves variables are included in the sample. Figure
2 shows the daily reserves stocks for Argentina and Brazil in US$ billion and their
respective reserve stock changes (calculated as first-differences of stock levels). In May
2009, Brazil and Argentina had the highest reserve stock levels in South America, respectively, US$190 billion and US$46 billion. While Brazil finishes the sample period
with US$373 billion, Argentina after much fluctuations went back to its starting value
of US$46 billion.
The currency intervention data represents the net open-market operations done by
the Central Bank of Brazil in the domestic foreign exchange market. Positive values
indicate a net purchase of US dollars; whereas negative values relate to a net sale
of US dollars. In theory, a central bank can influence the volatility and level of the
exchange rate by exercising such open-market operations. Ceteris paribus, a sizable
net purchase (sale) of US dollars should lead to a depreciation (appreciation) of the
domestic currency against the US dollar. Moreover, there is a possibility – which is
tested in this paper – that central bank foreign exchange operations in one country can
influence exchange rates in other satelite economies such as in the case of Brazil and
Argentina. Figure 3 plots the Brazilian currency intervention data and the percentage
exchange returns against the US dollar in Brazil and Argentina.
Finally, there is a close accounting relationship between central bank currency intervention and reserves. Other things equal, every dollar acquired by a central bank
currency intervention will increase the reserve stock by one dollar. Notwithstading, reserve stocks can also vary due to other operations, such as external debt amortization
and interest-rate services. In Brazil, most of reserve changes are indeed due to central
bank currency intervention, displaying a correlation coefficient of 0.69. Figure 4 shows
4
Euro
Euro Returns
5
0.9
0.85
2.5
0.8
0
0.75
0.7
−2.5
0.65
0.6
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−5
May−09
Jun−12
Dec−09
British Pound
Aug−10
Mar−11
Nov−11
Jun−12
Nov−11
Jun−12
Nov−11
Jun−12
Nov−11
Jun−12
British Pound Returns
5
0.75
2.5
0.7
0.65
0
0.6
−2.5
0.55
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−5
May−09
Jun−12
Dec−09
Argentine Peso
2
4.5
1
4
0
3.5
−1
Dec−09
Aug−10
Mar−11
Mar−11
Argentine Peso Returns
5
3
May−09
Aug−10
Nov−11
−2
May−09
Jun−12
Dec−09
Brazilian Real
Aug−10
Mar−11
Brazilian Real Returns
5
2.2
2.5
2
0
1.8
−2.5
1.6
1.4
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−5
May−09
Jun−12
Dec−09
Aug−10
Mar−11
Figure 1: Daily Exchange Rates and Percentage Exchange Rate Returns against the
US dollar, May, 2009 – Jun, 2012. Source: Datastream 5.0, Thomson Reuters – Codes:
ECUNIT$, UKDOLLR, ARGPE$, BRACRU$.
5
Brazil’s Reserve Stocks
Brazil’s Reserve Stock Changes
380
6
360
5
340
4
320
3
300
280
2
260
1
240
0
220
−1
200
180
May−09 Dec−09 Aug−10 Mar−11 Nov−11 Jun−12
−2
May−09 Dec−09 Aug−10 Mar−11 Nov−11 Jun−12
Argentina’s Reserve Stocks
Argentina’s Reserve Stock Changes
54
1
53
0.5
52
51
0
50
49
−0.5
48
−1
47
46
−1.5
45
44
May−09 Dec−09 Aug−10 Mar−11 Nov−11 Jun−12
−2
May−09 Dec−09 Aug−10 Mar−11 Nov−11 Jun−12
Figure 2: Daily Reserve Stocks and Reserve Stock Changes, in US$ billion, May,
2009 – Jun, 2012. Source: Datastream 5.0, Thomson Reuters – Codes: AGRESST,
BRRESCS.
6
5
5
2.5
2.5
0
0
−2.5
−2.5
5
2
2.5
1
0
0
−2.5
Brazil’s Currency Intervention, US$ bn (left)
Brazil’s Currency Returns, in percent (right)
−5
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−1
Brazil’s Currency Intervention, US$ bn (left)
Argentina’s Currency Returns, in percent (right)
−5
Jun−12
−5
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−2
Jun−12
Figure 3: Daily Brazil’s Central Bank Currency Intervention in US$ billion vs. Percentage Exchange Rate Returns against the US dollar in Brazil and Argentina, May,
2009 – Jun, 2012. Sources: Central Bank of Brazil; Datastream 5.0, Thomson Reuters
– Codes: ARGPE$, BRACRU$.
6
6
4
4
2
2
0
0
Brazil’s Currency Intervention, US$ bn (left)
Brazil’s Reserve Stock Changes, US$ bn (right)
−2
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−2
Jun−12
Figure 4: Daily Brazil’s Central Bank Currency Intervention vs. Daily Brazil’s Reserve
Stock Changes, in US$ billion, May, 2009 – Jun, 2012. Source: Central Bank of Brazil.
7
the daily reserve changes and the currency intervention data for Brazil. In the sample
period, Brazil has intervened in 585 out of 825 days. Apart from net sales of US$924
million on May 7 and US$624 million on June 5 in 2009, all the other intervention days
consisted of net purchases of US dollars, amounting US$145.5 billion purchases in the
period. Accordingly, Brazil’s reserve stock increased US$183.4 billion during the same
period, from US$190.5 billion to US$373.9 billion.
1.2
Model Specification
A latent factor model with iid and unit variance assumptions is used to capture the
co-movements in the volatility of exchange rates and currency intervention, where
volatility decompositions are the main vehicle for analysis. In order to investigate the
mercantilist motive hypothesis, changes in foreign reserves are also introduced and
used as proxy for currency interventions. There are several advantages of using the
latent factor framework: it provides a parsimonious representation of the data without
the need to identify nor to model observable variables; and, the imposed iid and unit
variance assumptions allow for the variance decomposition of the variables in the model,
exactly accounting for the contribution of each factor to the overall volatitity.
The model consists of two subsets Yt of six standardised (zero mean, unit variance)
variables in the 825 daily observations. In Model A, the set Yt comprises exchange rate
returns, reserve changes and Brazil’s central bank currency intervention. In Model B,
Brazil’s central bank intervention data is replaced by Brazil’s reserve changes – the effectiveness of Brazil’s reserve changes as a proxy for currency intervention is a key factor
to test on the mercantilist motive hypothesis. Following Fry-McKibbin and Wanaguru
(2012), a distinction in the sample is imposed to differentiate non-intervention days
(j = 0) from intervention days (j = 1). This allows to isolate the contribution of
Brazil’s central bank currency intervention factor into the exchange rate markets of
Brazil and Argentina.
MODEL A
0
Ytj = rEU Rtj , rU KPtj , rARcjt , dARrtj , rBRcjt , BRintjt
j = 0, 1.
MODEL B
0
Ytj = rEU Rtj , rU KPtj , rARcjt , dARrtj , rBRcjt , dBRrtj
j = 0, 1.
8
rEU Rt ≡ standardised Euro returns
rU KPt ≡ standardised British pound returns
rARct ≡ standardised Argentine currency returns
dARrt ≡ standardised Argentine reserve changes
rBRct ≡ standardised Brazilian currency returns
BRintt ≡ standardised Brazilian currency intervention
dBRrt ≡ standardised Brazilian reserve changes
The dynamics of each standardised variable refers to the set of orthogonal latent
factors, which comprises: a global factor (ωt ), common to all variables; a currency
factor (κt ), common only to the exchange rate returns; a Brazilian factor (brt ) and
an Argentine factor (art ), which respectively capture the combined forces behind the
domestic exchange rate returns and reserve changes in each country; and, lastly, a
residual factor (ut ) catching the idiosyncracies of each separate market.
The model assumes the dynamics of the intervention days differs from the dynamics
of the non-intervention days, which probably prompted the intervention action in the
first place. In this regard, to capture this structural break on the parameters, the
non-intervention day set is nested in the intervention day set. The notation for the
loading factor parameters is λji,f , where: j = {0, 1} accounts for the possible structural
break according to non-intervention and intervention days; i = {1, 2, ..., 6} corresponds
respectively to the standardised variables rEU Rtj , rU KPtj , rARcjt , dARrtj , rBRcjt ,
BRintjt in Model A and rEU Rtj , rU KPtj , rARcjt , dARrtj , rBRcjt , dBRrtj in Model B;
and, f = {ω, κ, br, ar, u} denotes to which corresponding orthogonal latent factor the
parameter is assigned.
Furthermore, there are two extra loading parameters, ιbr and ιar , only present on
the intervention day set (j = 1), which are designed to test for the impact of Brazil’s
central bank currency intervention into the exchange rate returns of Brazil and Argentina. In line with the literature, the parameter ιbr is the raison-d’être of Brazil’s
currency intervention, where central banks operate in the foreign exchange market to
intervene in its own exchange rate path (Menkhoff, 2012). With respect to the capacity
of Brazil’s central bank intervention to affect the Argentine currency returns (parameter ιar ), Brazil is by far Argentina’s biggest trade partner, accounting for one-quarter
of the Argentine total external trade. Therefore, it is reasonable to suspect that, for
its relative economic size, proximity and trade links, Brazil’s currency intervention
9
should indirectly impact Argentina’s exchange rate market. On the same grounds, but
with opposite implications, it is assumed that Brazil’s currency interventions have an
insignificant impact on the British pound and euro currency markets: Brazil is geographically far from Europe; Brazil’s currency is not globally traded; and Brazil plays
a relatively small part in world trade, accounting for less than 3% of Europe’s external
trade.
In matrix form, the model is expressed as
Ytj = Λj Ft
where
j
Y6,t
= BRintjt , for Model A
and
j
Y6,t
= dBRrtj , for Model B.
Non-Intervention Days (j = 0)




 0 
  0 
λ1,ω
λ1,κ
0
0
rEU Rt0
 0 
 0 
λ02,κ 
rU KPt0  λ02,ω 
 0 


 0 
  0 

0
λ3,ar 
 0 
λ3,κ 
 rARct  λ3,ω 







 

 dARrt0  = λ04,ω  ωt +  0  κt +  0  brt + λ04,ar  art










λ05,br 
 0 
λ05,κ 
 rBRc0t  λ05,ω 
0
Y6,t
λ06,br
0
0
λ06,ω
 0
 
0
0
0
0
λ1,u 0
u1,t


 0 λ02,u 0
0
0
0  u2,t 


0
 0
 u3,t 
0
λ
0
0
0
3,u
 
+
 0
 
0
0 λ04,u 0
0 

 u4,t 
 0
0
0
0 λ05,u 0  u5,t 
u6,t
0
0
0
0
0 λ06,u

10
(1)
Intervention Days (j = 1)


 0

  0
(λ1,ω + λ11,ω )
(λ1,κ + λ11,κ )
rEU Rt1
(λ02,κ + λ12,κ )
rU KPt1  (λ02,ω + λ12,ω )


 0

  0
1
1
(λ3,κ + λ13,κ )
 rARct  (λ3,ω + λ3,ω )
 κt



 

 dARrt1  = (λ04,ω + λ14,ω ) ωt + 
0






(λ05,κ + λ15,κ )
 rBRc1t  (λ05,ω + λ15,ω )
1
Y6,t
0
(λ06,ω + λ16,ω )




0
0




0


 0 0 1 




0
 brt + (λ03,ar + λ13,ar ) art
+




 0 0 1 
(λ4,ar + λ4,ar )
(λ5,br + λ5,br )


0
0
1
(λ6,br + λ6,br )
0




+



0
0
0
0
0
(λ01,u + λ11,u )
1
0
0
0
0
0
0
(λ2,u + λ2,u )
0
1
0
0
ιar
0
0
(λ3,u + λ3,u )
0
0
0
0
0
(λ04,u + λ14,u )
0
1
0
0
0
0
(λ5,u + λ5,u )
ιbr
0
0
0
0
0
(λ06,u + λ16,u )
The use of the model revolves around the volatility decomposition of the standardised variables in Ytj . Using the assumptions that the latent factors in Ft are iid (0,1)
random variables, the variance of each element in Yt1 on the intervention days (j = 1) is
V ar(rEU Rt1 ) = (λ01,ω + λ11,ω )2 + (λ01,κ + λ11,κ )2 + (λ01,u + λ11,u )2
V ar(rU KPt1 ) = (λ02,ω + λ12,ω )2 + (λ02,κ + λ12,κ )2 + (λ02,u + λ12,u )2
V ar(rARc1t ) = (λ03,ω + λ13,ω )2 + (λ03,κ + λ13,κ )2 + (λ03,ar + λ13,ar )2 + ι2ar + (λ03,u + λ13,u )2
V ar(dARrt1 ) = (λ04,ω + λ14,ω )2 + (λ04,ar + λ14,ar )2 + (λ04,u + λ14,u )2
V ar(rBRc1t ) = (λ05,ω + λ15,ω )2 + (λ05,κ + λ15,κ )2 + (λ05,br + λ15,br )2 + ι2br + (λ05,u + λ15,u )2
1
V ar(Y6,t
) = (λ06,ω + λ16,ω )2 + (λ06,br + λ16,br )2 + (λ06,u + λ16,u )2
11
 
u1,t
 u2,t 
 
 u3,t 
 
 u4,t 
 
 u5,t 
u6,t
As a result, the corresponding proportion of the volatility to each factor is displayed
in Table 1.
Table 1: Volatility Decomposition on Intervention Days
Notes: Regarding the variance of each element in Yt0 for the non-intervention days (j = 0),
the strutural break parameters (λ1i,f , ιbr , ιar ) are dropped.
Global
Currency
Brazil
rEU Rt1
(λ01,ω +λ11,ω )2
V ar(rEU Rt1 )
(λ01,κ +λ11,κ )2
V ar(rEU Rt1 )
—
rU KPt1
(λ02,ω +λ12,ω )2
V ar(rU KPt1 )
(λ02,κ +λ12,κ )2
V ar(rU KPt1 )
rARc1t
(λ03,ω +λ13,ω )2
V ar(rARc1t )
dARrt1
Factors
Argentina
Intervention
Residual
—
—
(λ01,u +λ11,u )2
V ar(rEU Rt1 )
—
—
—
(λ02,u +λ12,u )2
V ar(rU KPt1 )
(λ03,κ +λ13,κ )2
V ar(rARc1t )
—
(λ03,ar +λ13,ar )2
V ar(rARc1t )
ι2ar
V ar(rARc1t )
(λ03,u +λ13,u )2
V ar(rARc1t )
(λ04,ω +λ14,ω )2
V ar(dARrt1 )
—
—
(λ04,ar +λ14,ar )2
V ar(dARrt1 )
—
(λ04,u +λ14,u )2
V ar(dARrt1 )
rBRc1t
(λ05,ω +λ15,ω )2
V ar(rBRc1t )
(λ05,κ +λ15,κ )2
V ar(rBRc1t )
(λ05,br +λ15,br )2
V ar(rBRc1t )
—
ι2br
V ar(rBRc1t )
(λ05,u +λ15,u )2
V ar(rBRc1t )
1
Y6,t
(λ06,ω +λ16,ω )2
1 )
V ar(Y6,t
—
(λ06,br +λ16,br )2
1 )
V ar(Y6,t
—
—
(λ06,u +λ16,u )2
1 )
V ar(Y6,t
1.3
Estimation Method
The factor model described in Subsection 1.2 uses a Generalised Method of Moments
(GMM) estimator, which produces consistent, asymptotically normal and efficient estimates (Hansen, 1982). GMM estimation focuses on the information contained by
the moments of the data. The goal is compute the unknown parameters by matching
the theoretical moments of the model to the empirical moments of the data in both
intervention-day and non-intervention-day sets.
The identification and estimation of the model make use of the precise known days
of currency intervention in the Brazilian foreign exchange market.2 Both Model A
and Model B, in this sense, are exactly identified. Each dataset of intervention and
non-intervention days provides 21 empirical moments. Thus, 42 empirical moments
in total, which are used to identify and estimate the 42 unknown parameters of the
model.
Let H j be a T j -by-21 empirical matrix (T j daily observations in each dataset j =
{0, 1}, 21 contemporaneous cross-products between the 6 standardised variables in Ytj ),
then
2
For similar strategy, see Fry-McKibbin and Wanaguru (2012) and Dungey et al. (2010).
12



H =

j
j
j
Y1,1
Y1,1
j
j
Y1,2
Y1,2
..
.
...
...
..
.
j
j
Y6,1
Y6,1
j
j
Y6,2
Y6,2
..
.
j
j
Y1,1
Y2,1
j
j
Y1,2
Y2,2
..
.
...
...
..
.
j
j
Y1,1
Y6,1
j
j
Y1,2
Y6,2
..
.
j
j
Y2,1
Y3,1
j
j
Y2,2
Y3,2
..
.
...
...
..
.
j
j
Y5,1
Y6,1
j
j
Y5,2
Y6,2
..
.
j
j
j
j
j
j
j
j
j
j
j
j
Y1,T
. . . Y6,T
Y1,T
. . . Y1,T
Y2,T
. . . Y5,T
j Y1,T j
j Y6,T j
j Y2,T j
j Y6,T j
j Y3,T j
j Y6,T j
for j = {0, 1}.
It is straightforward to see that, by the law of large numbers, the average value
of each column of H j asymptotically converges to the respective true second-order
j
expectation E(Yi,tj , Yj,t
) for i, j = {1, 2, 3, 4, 5, 6}.3 In this case, an optmisation problem
is solved by guessing possible parameter values in Λj of eq.(1) such that minimises the
difference between the empirical moments extracted from the columns of H j and the
0
theoretical moments derived from the lower diagonal entries of Λj Λj .




0

Λj Λj = 



j
j
E(Y1,t
Y1,t
)
j
j
E(Y1,t Y2,t )
j
j
E(Y1,t
Y3,t
)
j
j
E(Y1,t
Y4,t
)
j
j
E(Y1,t Y5,t )
j
j
E(Y1,t
Y6,t
)

j
j
E(Y2,t
Y2,t
)
j
j
E(Y2,t Y3,t )
j
j
E(Y2,t
Y4,t
)
j
j
E(Y2,t Y5,t )
j
j
E(Y2,t
Y6,t
)
j
j
E(Y3,t
Y3,t
)
j
j
j
j
E(Y3,t
Y4,t
) E(Y4,t
Y4,t
)
j
j
j
j
j
j
E(Y3,t Y5,t ) E(Y4,t Y5,t ) E(Y5,t
Y5,t
)
j
j
j
j
j
j
j
j
E(Y3,t Y6,t ) E(Y4,t Y6,t ) E(Y5,t Y6,t ) E(Y6,t
Y6,t
)








Lastly, in order to calculate the standard errors of the estimated parameters, a
bootstrap procedure (Efron and Tibshirani, 1994) separetly resamples both datasets
(j = {0, 1}) 1000 times.
1.4
Analysis of the Results
This section examines the effects of Brazil’s central bank currency intervention on the
returns of the Brazilian and Argentine currencies in accordance with eq.(1). Moreover,
it investigates the mercantilist motive hypothesis by using reserve changes as a proxy for
currency intervention. In Section 1.4.1, results for Model A (using Brazil’s central bank
currency intervention data) are presented and discussed. Section 1.4.2 analyses the
results for Model B (using Brazil’s reserve changes as a proxy for currency intervention
j
j
j
j
≡ rU KPtj , Y3,t
≡ rARcjt , Y4,t
≡ dARrtj ,
For clarity, please mind the notation: Y1,t
≡ rEU Rtj , Y2,t
j
j
Y5,t
≡ rBRcjt and Y6,t
≡ BRintjt (Model A) or dBRrtj (Model B).
3
13





data) and compare it with the ones in Model A. Lastly, Section 1.4.3 reestimates and
discusses the results for Models A and B using other Latin American countries as
controls.
1.4.1
Results for Model A
The volatility decomposition of the factor Model A of central bank intervention is
presented in Table 2. Recapping, the contributing factors for the currency returns
and reserve changes are: ‘Global’, common to all variables; ‘Currency’, common only
to the exchange rate returns; ‘Brazil’ and ‘Argentina’, which respectively capture the
combined forces behind the exchange rate returns and reserve changes in each country;
‘Intervention’, which grasps the impact of Brazil’s central bank intervention on the
Brazilian and Argentine currency markets;4 and, lastly, ‘Residual’, which catches the
idiosyncracies of each separate market.
The top panel of Table 2 provides the percentage contribution of the orthogonal
latent factors for the days with no central bank currency intervention in Brazil. The
currency factor dominates the volatility of the euro and British pound market on
non-intervention days, accounting for 67.01% and 57.17% respectively. On its turn,
Brazil’s currency factor share is 36.71%, whereas Argentina’s is barely present at 0.53%.
Indeed, Argentina’s currency volatility is mostly due to internal factors, reflecting its
isolacionist and idiosyncratic public policy in the past years.
On the bottom panel, the volatility decomposition on the days of Brazil’s central
bank intervention is presented. The influence of the global factor changes substantially,
impacting all markets apart from the insulated Argentina. This is evidence of major
world events behind the decision of Brazil’s central bank to intervene. In particular,
the global factor accounts for 53.95% and 51.50% of the volatility in Brazil’s exchange
rate returns and central bank currency intervention data, respectively.
Interestingly, the Brazil factor (brt ) also plays an important role on intervention days
to both its exchange rate and intervention data, explaining 39.04% and 42.06% of the
volatility in each Brazilian market respectively. This shows the co-movements between
its currency and intervention data. The same pattern is not followed in Argentina.
The Argentina factor (art ) does not influence simultaneously Argentina’s exchange
rate returns and reserve changes, favouring the former at 18.92% as opposed to the
4
The intervention factor is derived from the impact of the residual factor of Brazil’s central bank intervention (u6,t ) into the currency markets of Brazil and Argentina through the respective parameters
ιbr and ιar – see eq.(1).
14
Table 2: Volatility Decomposition for Model A
Notes: Contribution of each factor to total volatility, in percent. The model is estimated over
the period May 4, 2009–June 29, 2012 (see eq.(1) and Table 1).
Global Currency
Non-intervention days (j = 0)
rEU Rt0
9.17
67.01
rU KPt0
rARc0t
dARrt0
rBRc0t
BRint0t
Factors
Brazil Argentina
Intervention
Residual
Total
—
—
—
23.81
100.0
8.11
57.17
—
—
—
34.72
100.0
0.03
0.53
—
82.43
—
17.00
100.0
42.02
—
—
1.31
—
56.67
100.0
9.20
36.71
45.38
—
—
8.71
100.0
1.06
—
0.00
—
—
98.94
100.0
Intervention days (j = 1)
rEU Rt1
54.72
45.28
—
—
—
0.00
100.0
40.07
6.30
—
—
—
53.64
100.0
0.71
3.64
—
18.92
38.83
37.90
100.0
0.83
—
—
0.49
—
98.68
100.0
53.95
0.06
39.04
—
6.95
0.00
100.0
51.50
—
42.06
—
—
6.44
100.0
rU KPt1
rARc1t
dARrt1
rBRc1t
BRint1t
0.49% on the later (the same lack of domestic forces in Argentina is observed in the nonintervention days with 82.43% and 1.31%, respectively). This leads to the conclusion
that the Argentina factor acts mainly as a de facto second residual in the volatility
decomposition method – more evidence of Argentina’s idiosyncratic markets.
After controlling for global, currency and national factors, the impact of Brazil’s
central bank intervention data is clear in both Brazil’s and Argentina’s exchange rate
markets. The intervention factor accounts for 6.95% of the overall volatility in Brazil’s
exchange rate return, sweeping any further residual contribution to this market. Indeed, not only the residual contribution to the Brazilian exchange rate returns is 0.00%,
but even its corresponding estimates (λ05,u , λ15,u ) are statistically zero.5 It is also worth
noting that the Brazil’s central bank intervention residual factor (u6,t ) on intervention
days had basically the same impact in the overall volatility of both Brazil’s central bank
intervention data and currency returns – 6.44% and 6.95%, respectively. This indicates
the full transfer of Brazil’s central bank intervention impact into its currency market.
In Argentina, the impact of Brazil’s central bank intervention factor was even higher, at
38.83% – a fivefold increase from Brazil’s case. This highlights the cross-border effects
5
For full list of parameter estimates and p-values, see Table 13 in the Appendix.
15
Table 3: Wald Tests on Intervention and Structural Breaks in Factor Model A
Notes: A bootstrap procedure (resampling 1000 times) was used to calculate the variancecovariance matrix of the parameters. The model is estimated over the period May 4, 2009–
June 29, 2012 (see eq.(1) and Table 1). DOF stands for degrees of freedom.
Intervention Parameters
ιbr
ιar
Wald Test Hypothesis
Joint Intervention parameters
H0 : ιbr = ιar = 0
H0 : ιbr = ιar =
λ16,u
Estimates
-0.2473
Standard Deviation
0.1802
p-value
0.085
-0.6343
0.2850
0.013
DOF
Test Statistic
p-value
2
11.38
0.003
3
12.92
0.005
22
304.03
0.000
=0
Joint structural break parameters
H0 : ιbr = ιar = λ1i,f = 0
of Brazil’s central bank intervention on the neighbour country and satelite economy.
In this regard, further research would be welcome to better understand the dynamics
behind this contagion.
Lastly, Table 3 reports the results on the statistical significance of intervention parameters and joint structural breaks. Both intervention parameters, ιbr and ιar , are
individually significant at 10% level. Moreover, further Wald tests on joint parameters also confirm the validity of the model, including the structural break imposed on
intervention days.
1.4.2
Results for Model B
Model B, which replaces Brazil’s central bank intervention data for its reserve changes,
provides further evidence of the mercantilist motive hypothesis. That is, its results
present robust evidence of the strong relationship between foreign reserves build-up
and foreign exchange rate in Brazil.
On non-intervention days (top part of Table 4), a different structure in the volatility
decomposition of Model B in comparison with Model A corresponds to the broader
nature of reserve changes. Reserve changes encompass central bank interventions in
the foreign exchange market, but also may be due to other factors such as external
debt amortization and interest-rate servicing.6 While intervention data is zero in non6
See Subsection 1.1 for more information on the links between central bank currency intervention
and reserve build-up.
16
Table 4: Volatility Decomposition for Model B
Notes: Contribution of each factor to total volatility, in percent. The model is estimated over
the period May 4, 2009–June 29, 2012 (see eq.(1) and Table 1).
Global Currency
Non-intervention days (j = 0)
rEU Rt0
42.98
36.16
rU KPt0
rARc0t
dARrt0
rBRc0t
dBRrt0
Factors
Brazil Argentina
Intervention
Residual
Total
—
—
—
20.86
100.0
40.58
22.82
—
—
—
36.60
100.0
0.06
0.47
—
1.66
—
91.20
100.0
8.76
—
—
91.20
—
0.04
100.0
43.55
6.73
4.27
—
—
45.45
100.0
68.58
—
12.65
—
—
18.77
100.0
Intervention days (j = 1)
rEU Rt1
54.76
42.25
—
—
—
3.00
100.0
40.10
6.72
—
—
—
53.18
100.0
0.89
4.35
—
19.36
47.20
28.20
100.0
0.83
—
—
0.51
—
98.67
100.0
54.05
0.07
39.58
—
6.30
0.00
100.0
51.46
—
42.76
—
—
5.78
100.0
rU KPt1
rARc1t
dARrt1
rBRc1t
dBRrt1
intervention days by definition, the same does not apply to reserve change data. This
explains the different weights on non-intervention day factors, which creates a distinct
orthogonal state-space set to decompose the data when j = 0.
Notwithstanding, the volatility decomposition in Model B on intervention days
(bottom part of Table 4) mimics exacly the same trends and weights of Model A. The
global factor predominates in all markets apart from Argentina; the currency factor is
stronger in the euro zone; the national factor brt shows strong co-movements behind
the Brazilian exchange rate and its reserve changes; the national factor art favours the
Argentine exchange rate as opposed to its reserve changes, hence acting as a de facto
second residual in the volatility decomposition model for the Argentine markets.
The impact of central bank intervention using Brazil’s reserve change data is unequivocal. Both intervention parameters ιbr and ιar are statistically significant at 5%
(top part of Table 5) and very close7 to the ones estimated in Model A. After controlling
for global, currency and national factors, the intervention factor accounts for 6.30%
(little less than the 6.95% estimated in Model A) of the overall volatility in Brazil’s
7
The estimated intervention parameters ιbr and ιar in Model B lie inside the 99% confidence interval
of the estimated intervention parameters in Model A, and vice-versa.
17
Table 5: Wald Tests on Intervention and Structural Breaks in Factor Model B
Notes: A bootstrap procedure (resampling 1000 times) was used to calculate the variancecovariance matrix of the parameters. The model is estimated over the period May 4, 2009–
June 29, 2012 (see eq.(1) and Table 1). DOF stands for degrees of freedom.
Intervention Parameters
ιbr
ιar
Wald Test Hypothesis
Joint Intervention parameters
H0 : ιbr = ιar = 0
H0 : ιbr = ιar =
λ16,u
Estimates
-0.2354
Standard Deviation
0.1359
p-value
0.042
-0.6993
0.2562
0.003
DOF
Test Statistic
p-value
2
19.63
0.000
3
12.92
0.000
22
504.18
0.000
=0
Joint structural break parameters
H0 : ιbr = ιar = λ1i,f = 0
exchange rate return, sweeping any further residual contribution to this market (as
happened in Model A).
With regards to Argentina, the impact of Brazil’s reserve changes on intervention days account for 47.20% of the overall volatility in Argentina’s exchange rate
return. Interestingly, this percentage is even higher than the one estimated in Model
A (38.83%).
Lastly, Table 5 reports the results on the statistical significance of intervention
parameters and joint structural breaks. As noted above, both intervention parameters,
ιbr and ιar , are individually significant at 5% level.8 Moreover, Wald tests on joint
parameters also confirm the validity of Model B, including the structural break imposed
on intervention days.
1.4.3
Results using other Latin American countries as controls
Limited by data availability on daily reserves, further estimation is carried replacing the
Argentine reserves and currency markets for other Latin American countries, namely
the Chilean and Peruvian reserves and currency markets. Brazil is the 5th largest trade
partner of Chile and Peru, accounting for 7.2% and 5.2% of their total external trade
respectively. Conversely, Chile and Peru rank 7th and 24th as Brazil’s major trade
partners, accounting for 2.2% and 0.8% of Brazil’s external trade, respectively.9
8
9
For full list of parameter estimates and p-values, see Table 13 in the Appendix.
Source: IMF (DoTS), 2010.
18
The results presented in this subsection are produced with the same estimation
methodology in Subsection 1.3, including the same number of standardised variables
in Ytj . In Case 1, models A and B are estimated replacing Argentina’s currency return
(rARc) and reserves change (dARr) variables with the Chilean counterpart (rCHc and
dCHr, respectively). In Case 2, models A and B are re-estimated, but now replacing
with the Peruvian currency return (rP Ec) and reserves change (dP Er) variables instead.
Case 1: Estimation with Chile’s data as Latin America control
0
Model A: Ytj = rEU Rtj , rU KPtj , rCHcjt , dCHrtj , rBRcjt , BRintjt
j = 0, 1.
0
Model B: Ytj = rEU Rtj , rU KPtj , rCHcjt , dCHrtj , rBRcjt , dBRrtj
j = 0, 1.
Case 2: Estimation with Peru’s data as Latin America control
0
Model A: Ytj = rEU Rtj , rU KPtj , rP Ecjt , dP Ertj , rBRcjt , BRintjt
j = 0, 1.
0
Model B: Ytj = rEU Rtj , rU KPtj , rP Ecjt , dP Ertj , rBRcjt , dBRrtj
j = 0, 1.
Maintaining the same number of standardised variables in Ytj is crucial for the
validity of the results. With 6 standardised variables in Ytj , there is exact identification of the model, with 42 empirical moments to estimate 42 parameters.10 If the
model accommodated the Chilean and Peruvian data and kept the Argentine variables,
the sample set would have increased to 10 standardised variables.11 In this case, 110
empirical moments would be available to estimate 66 parameters: therefore, an overidentification case. In theory, this is not an impediment in itself, but after estimating
the augmented model with 10 standardised variables, two problems arose from the
overidentification issue. First, the high number of degrees of freedom12 led to very
conflicting and unstable results depending on the initial values of the GMM procedure.
Second, the results failed the overidentification test, implying that the assumption on
the ortogonality of the latent factors was not valid.
In general, results using Chile’s (Case 1) and Peru’s (Case 2) data are in line
with results in Subsections 1.4.1 and 1.4.2. Both Cases 1 and 2 show that Brazil’s
central bank currency intervention in Models A and B had a significant impact in
Brazil’s as well as in Chile’s and Peru’s currency markets. This is further evidence
10
See Subsection 1.3 on the Estimation Method.
0
Ytj = rEU Rtj , rU KPtj , rARcjt , dARrtj , rCHcjt , dCHrtj , rP EHcjt , dP Ertj , rBRcjt , BRintjt
12
In this case, Degrees of Freedom = 110 - 66 = 44.
11
19
of the effectiveness of Brazil’s central bank intervention and, more importantly, of the
validity of the mercantilist motive hypothesis.
Table 6 summarises the contribution of Brazil’s intervention to the overall volatility
in Cases 1 (Chile) and 2 (Peru),13 plus the counterpart result in the benchmark model
using the Argentine data (Subsections 1.4.1 and 1.4.2). Both Chile’s and Peru’s currency markets also seem to be affected by Brazil’s central bank currency intervention.
The volatility contributions of the intervention factor in their respective markets, although lower than Argentina’s, are considerable: 16.89% and 25.88% in Chile, 26.12%
and 9.06% in Peru, for Models A and B respectively.
With respect to the impact in the Brazilian currency market, both Cases 1 and 2
also show the significant contribution of Brazil’s central bank currency intervention in
the overall volatility of the Brazilian exchange rate (rBRc): in Case 1, 16.04% and
19.39% in Models A and B; whereas 23.42% and 67.51% in Case 2, respectively.
Table 6: Contribution of Brazil’s Intervention to Overall Volatility in Each Market
Notes: Values in percent. The model is estimated over the period May 4, 2009–June 29, 2012
(see eq.(1), Table 1 and Section 1.4.3).
Case 1 : rCHc
Model A
16.89
Model B
25.88
Case 2 : rP Ec
26.12
9.06
Benchmark : rARc
38.83
47.20
Case 1 : rBRc
16.04
19.39
Case 2 : rBRc
23.42
67.51
Benchmark : rBRc
6.95
6.30
Table 7 reports the statistical significance of intervention parameters and joint
structural breaks. Apart from Case 1 using intervention data (Model A), all intervention parameters proved statistical significance at 10% level. Moreover, Wald tests for
both Cases 1 and 2 confirm the structural break on intervention days.
Overall, results using other Latin American countries as controls are in line with
the benchmark model, albeit with different intensity. Again it is emphasised that
further research should follow to better understand the cross-border links on central
bank intervention. Accordingly, the next section investigates the spillover effects of
13
For full results on the volatility decomposition for Case 1 (Tables 14 and 15) and Case 2 (Tables
16 and 17), see Appendix.
20
active reserve build-up through central bank intervention on neighbouring countries.
Table 7: Wald Tests on Intervention and Structural Breaks in Factor Model A/B
Notes: A bootstrap procedure (resampling 1000 times) was used to calculate the variancecovariance matrix of the parameters. The model is estimated over the period May 4, 2009–
June 29, 2012 (see eq.(1), Table 1 and Section 1.4.3).
Intervention Parameters
Estimates
Standard Deviation
p-value
Models A/B
Models A/B
Models A/B
Case 1:
ιbr
-0.3756 / -0.4129
0.2958 / 0.2542
0.102 / 0.052
ιch
0.3856 / 0.4597
0.3856 / 0.3394
0.168 / 0.088
Case 2:
ιbr
-0.4538 / -0.7705
0.1938 / 0.2578
0.010 / 0.001
ιpe
-0.5299 / -0.3121
0.2024 / 0.1919
0.004 / 0.052
Degrees of
Freedom
Test Statistic
p-value
Models A/B
Models A/B
2
4.96 / 4.55
0.084 / 0.103
3
5.09 / 12.75
0.165 / 0.005
Joint structural break parameters
H0 : ιbr = ιch = λ1i,f = 0
22
994.2 / 410.7
0.000 / 0.000
Case 2:
Joint Intervention parameters
H0 : ιbr = ιpe = 0
2
29.26 / 8.99
0.000 / 0.011
3
55.35 / 9.11
0.000 / 0.028
22
488.6 / 76.42
0.000 / 0.000
Wald Test Hypothesis
Case 1:
Joint Intervention parameters
H0 : ιbr = ιch = 0
H0 : ιbr = ιch =
H0 : ιbr = ιpe =
λ16,u
λ16,u
=0
=0
Joint structural break parameters
H0 : ιbr = ιpe = λ1i,f = 0
21
2
Regional Intervention Spillover Effects on Foreign
Reserves
This section investigates the regional spillover effects of reserve accumulation through
central bank intervention. There are some possible theoretical explanations on the
cross-country dynamics of reserve accumulation. First, reserve changes through central
bank intervention in one country might trigger reserve changes through central bank
intervention in another country in order to mantain relative exchange rates pegged
together. Second, part of the reserve hoarding phenomena might be due to signalling in
the credit market: a peacock effect. In an assymetric information environment, massive
reserve accumulations can signal to market players about a country’s economic health,
attracting foreign investments and minimising any potential self-fulfulling speculative
attack on its currency. Yet, such peacock-driven reserve accumulation must take into
account reserve stocks of neighbouring countries, as economic power is always relative.
In this case, game theory might help explain multiple Nash equilibria of reserve stocks
between two or more countries. Such new form of bullionism might be considered
as another manifestation of the mercantilist motive for accumulating reserves. In the
BWII framework, for instance, Dooley et al. (2004c) argue that reserve stocks can be
used as collateral for foreign direct investment. In this case, competing countries might
accumulate peacock-driven reserves in attempt to lure international productive capital.
This section looks to the empirical evidence of reserve stock co-movements between
neighbouring countries due to deliberate central bank intervention on foreign exchange
market. In particular, it investigates the impact of Brazil’s central bank intervention
on the volatility decomposition of reserve changes in Brazil, Argentina, Chile and
Peru. By attesting the empirical evidence of regional intervention spillover effects on
foreign reserves among neighbouring countries, this paper hopes to contribute to a new
research agenda on the mercantilist motive for reserve accumulation, which looks into
the cross-country links of reserve stocks.
On balance, after controlling for global, regional and domestic factors, results confirm the spillover effects of Brazil’s central bank intervention on foreign reserves among
neighbouring countries. Yet, parameter estimates do not support the hypothesis of
reserve changes as a result of a modern bullionism pratice; leaving co-movements in
currency intervention across neighbouring countries to be the driving force behind such
regional intervention spillovers.
22
This section is organised as follows: Subsection 2.1 presents a preliminary analysis
of the data; Subsection 2.2 introduces the model specification; Subsection 2.3 describes
the estimation method; and Subsection 2.4 examines the results.
2.1
Data
This subsection presents a preliminary analysis of the data used to investigate the
regional intervention spillover effects on foreign reserves. The data includes Brazil’s
central bank currency intervention and reserve changes for Brazil, Argentina, Chile
and Peru. Following Section 1, the sample period is constrained by the availability
of Brazil’s central bank currency intervention daily data, extending from May 4, 2009
to June 29, 2012. Once again, we express gratitude to the Central Bank of Brazil for
kindly disclosing this time series.
The selection of the sample reserve series follow the criteria of: data availability, as
few countries publicise their reserve stocks on a daily basis; regional links, the sample
is concentrated in South America with substantial trade links with Brazil; and reserve
volume, Brazil, Argentina, Peru and Chile led, in this order, the South America ranking
of reserve stocks in 2009.
Table 8 presents the descriptive statistics of the variables. Brazil’s reserve stock
is the highest throughout the sample, starting at US$190.5 billion and finishing at
US$373.9 billion. Argentina’s reserve stock begins as the second highest in the sample,
at US$46.4 billion, achieves a maximum of US$52.7 billion in the beginning of 2011
before finishing in June 2012 at the same level it started in 2009. Chile’s reserves are
the lowest of all four country stocks, although doubling during the sample period, from
US$23.6 billion to US$40.3 billion. Peru’s reserve stock started at US$31.1 billion and
steadily increased to US$57.2 billion in June 2012, overtaking Argentina’s second place
in the sample. Figure 5 plots the daily reserve stocks of Brazil, Argentina, Peru and
Chile in US dollars.
Brazil’s central bank intervention data is the same as the one described in Subsection 1.1. In 825 daily observations, the Central Bank of Brazil intervened 585 days,
purchasing a net amount of US$143.9 billion. Apart from net sales of US$924 million
on May 7, and US$624 million on June 5, 2009, all the other intervention days consisted
of net purchases of US dollars; with the highest purchase of US$4.9 billion on October
8, 2009; and a median daily purchase of US$141 million.
Figure 6 plots Brazil’s central bank currency intervention and the daily reserve
23
Table 8: Descriptive Statistics
Notes: Values in US$ billion, May, 2009 – Jun, 2012 (825 daily observations). Brazil’s Central
Bank Intervention statistics on intervention days only. Sources: Datastream 5.0, Thomson
Reuters – Codes: BRRESCS, AGRESST, CLINTRS, PERESIN; Central Bank of Brazil.
Start
190.5
Finish
373.9
Max
374.6
Min
190.0
Median
286.1
Std Dev.
56.5
Brazil’s Reserve Stock Changes
—
—
5.28
-1.40
0.157
0.508
Brazil’s Central Bank Intervention
—
—
4.90
-0.92
0.141
0.386
46.4
46.4
52.7
44.5
48.0
2.4
—
—
0.50
-1.87
0.006
0.142
23.6
40.3
42.0
23.3
26.8
5.9
—
—
2.40
-1.99
0.000
0.355
31.1
57.2
58.5
30.5
44.1
8.2
—
—
2.82
-0.96
0.005
0.220
Brazil’s Reserve Stock
Argentina’s Reserve Stock
Argentina’s Reserve Stock Changes
Chile’s Reserve Stock
Chile’s Reserve Stock Changes
Peru’s Reserve Stock
Peru’s Reserve Stock Changes
changes in Brazil, Argentina, Chile and Peru. As noted before, Brazil’s central bank
intervention account for most of Brazil’s reserve changes, displaying a correlation coefficient of 0.69 in the sample period. The correlation coefficients between Brazil’s central
bank intervention and reserve changes in Argentina, Chile and Peru are, respectively,
0.03, 0.04 and 0.03.
24
Brazil’s Reserve Stocks
Argentina’s Reserve Stocks
400
55
350
50
300
250
45
200
150
May−09
Dec−09
Aug−10
Mar−11
Nov−11
Jun−12
40
May−09
Dec−09
Chile’s Reserve Stocks
Aug−10
Mar−11
Nov−11
Jun−12
Nov−11
Jun−12
Peru’s Reserve Stocks
45
65
60
40
55
50
35
45
30
40
35
25
30
20
May−09
Dec−09
Aug−10
Mar−11
Nov−11
Jun−12
25
May−09
Dec−09
Aug−10
Mar−11
Figure 5: Daily Reserve Stocks in US$ billion, May, 2009 – Jun, 2012. Source: Datastream 5.0, Thomson Reuters – Codes: BRRESCS, AGRESST, CLINTRS, PERESIN.
25
5
5
2.5
2.5
0
0
−2.5
−2.5
Brazil’s Currency Intervention, US$ bn (left)
Brazil’s Reserve Stock Changes, US$ bn (right)
−5
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−5
Jun−12
5
2
2.5
1
0
0
−2.5
−1
Brazil’s Currency Intervention, US$ bn (left)
Argentina’s Reserve Stock Changes, US$ bn (right)
−5
May−09
Dec−09
Aug−10
Mar−11
Nov−11
5
−2
Jun−12
5
2.5
2.5
0
0
−2.5
−2.5
Brazil’s Currency Intervention, US$ bn (left)
Chile’s Reserve Stock Changes, US$ bn (right)
−5
May−09
Dec−09
Aug−10
Mar−11
Nov−11
5
−5
Jun−12
5
2.5
2.5
0
0
−2.5
−2.5
Brazil’s Currency Intervention, US$ bn (left)
Peru’s Reserve Stock Changes, US$ bn (right)
−5
May−09
Dec−09
Aug−10
Mar−11
Nov−11
−5
Jun−12
Figure 6: Daily Reserve Stock Changes vs. Brazil’s Central Bank Currency Intervention, in US$ billion, May, 2009 – Jun, 2012. Sources: Central Bank of Brazil; Datastream 5.0, Thomson Reuters – Codes: BRRESCS, AGRESST, CLINTRS, PERESIN.
26
2.2
Model Specification
As in Subsection 1.2, a latent factor model with iid and unit variance assumptions is
used to analyse the volatility decompositions of the sample. The model consists of a set
Yt of five standardised (zero mean, unit variance) variables in 825 daily observations,
comprising Brazil’s central bank currency intervention and reserve changes in Brazil,
Argentina, Chile and Peru. A distinction in the sample is again imposed to differentiate
non-intervention days (j = 0) from intervention days (j = 1). This enables the model
to isolate the contribution of Brazil’s central bank currency intervention factor into the
reserve changes of the sample countries.
0
Ytj = dARrtj , dCHrtj , dP Ertj , dBRrtj , BRintjt
j = 0, 1.
(2)
dARrt ≡ standardised Argentina’s reserve changes
dCHrt ≡ standardised Chile’s reserve changes
dP Ert ≡ standardised Peru’s reserve changes
dBRrt ≡ standardised Brazil’s reserve changes
BRintt ≡ standardised Brazil’s central bank currency intervention
The dynamics of each standardised variable refers to the set of orthogonal latent
factors, which comprises: a global factor (ωt ), common to all variables; a neighbourcountry factor (ηt ), to identify countries other than the one where the currency intervention is held; and, lastly, a residual factor (ut ), catching the idiosyncracies of each
separate market.
The model assumes the dynamics on intervention days differs from the one on
non-intervention days, which probably prompted the intervention action in the first
place. In this regard, to capture this structural break on the parameters, the nonintervention day set is nested in the intervention day one. The notation for the loading
factor parameters, λji,f , follows the rule: j = {0, 1} accounts for the possible structural
break according to non-intervention and intervention days; i = {1, 2, ..., 5} corresponds
respectively to the standardised variables dARrtj , dCHrtj , dP Ertj , dBRrtj , BRintjt ; and,
f = {ω, η, u} denotes to which corresponding orthogonal latent factor the parameter
is assigned.
Furthermore, there are four extra loading parameters, ιbr , ιar , ιch and ιpe , only
present on the intervention day set (j = 1), which are designed to test for the im27
pact of Brazil’s central bank currency intervention in the reserve changes of Brazil,
Argentina, Chile and Peru, respectively.
In matrix form, the model is expressed as
Ytj = Λj Ft
Non-Intervention Days (j = 0)
 0 
  0 


λ1,η
0
λ1,ω
dARrt
0 

λ2,η


 dCHrt0  λ02,ω 



 0 


 dP Ert0  = λ03,ω  ωt + λ3,η  ηt + 







 0 

 dBRrt0  λ04,ω 
 0 
0
0
BRintt
λ5,ω
0
λ01,u 0
0
0
0
0
0
0 λ02,u 0
0
0 λ03,u 0
0
0
0
0
0 λ4,u 0
0
0
0
0 λ05,u
(3)


u1,t
 u2,t 
 
 u3,t 
 
 u4,t 
u5,t
Intervention Days (j = 1)

 0

  0

(λ1,ω + λ11,ω )
(λ1,η + λ11,η )
dARrt1
(λ02,η + λ12,η )
 dCHrt1  (λ02,ω + λ12,ω )






 dP Ert1  = (λ03,ω + λ13,ω ) ωt + (λ03,η + λ13,η ) ηt



 



 dBRrt1  (λ04,ω + λ14,ω )
0
1
1
0
BRintt
0
(λ5,ω + λ5,ω )



+


(λ01,u + λ11,u )
0
0
0
ιar
0
1
0
(λ2,u + λ2,u )
0
0
ιch
0
ιpe
0
0
(λ03,u + λ13,u )
0
1
ιbr
0
0
0
(λ4,u + λ4,u )
0
0
0
0
0
(λ5,u + λ15,u )


u1,t
 u2,t 
 
 u3,t 
 
 u4,t 
u5,t
The use of the model revolves around the volatility decomposition of the standardised variables in Ytj . Using the assumptions that the latent factors in Ft are iid (0,1)
random variables, the variance of each element in Yt1 on the intervention days (j = 1) is
V ar(dARrt1 ) = (λ01,ω + λ11,ω )2 + (λ01,η + λ11,η )2 + (λ01,u + λ11,u )2 + ι2ar
28
V ar(dCHrt1 ) = (λ02,ω + λ12,ω )2 + (λ02,η + λ12,η )2 + (λ02,u + λ12,u )2 + ι2ch
V ar(dP Ert1 ) = (λ03,ω + λ13,ω )2 + (λ03,η + λ13,η )2 + (λ03,u + λ13,u )2 + ι2pe
V ar(dBRrt1 ) = (λ04,ω + λ14,ω )2 + (λ04,u + λ14,u )2 + ι2br
V ar(BRint1t ) = (λ05,ω + λ15,ω )2 + (λ05,u + λ15,u )2
As a result, the corresponding proportion of the volatility to each factor is displayed
in Table 9.
Table 9: Volatility Decomposition on Intervention Days
Notes: For the variance of each element in Yt0 on the non-intervention days (j = 0), the
strutural break parameters (λ1i,f , ιar , ιch , ιpe , ιbr ) are dropped.
Factors
Global
Neighbour Country
Intervention
Residual
dARrt1
(λ01,ω +λ11,ω )2
V ar(dARrt1 )
(λ01,η +λ11,η )2
V ar(dARrt1 )
ι2ar
V ar(dARrt1 )
(λ01,u +λ11,u )2
V ar(dARrt1 )
dCHrt1
(λ02,ω +λ12,ω )2
V ar(dCHrt1 )
(λ02,η +λ12,η )2
V ar(dCHrt1 )
ι2ch
V ar(dCHrt1 )
(λ02,u +λ12,u )2
V ar(dCHrt1 )
dP Ert1
(λ03,ω +λ13,ω )2
V ar(dP Ert1 )
(λ03,η +λ13,η )2
V ar(dP Ert1 )
ι2pe
V ar(dP Ert1 )
(λ03,u +λ13,u )2
V ar(dP Ert1 )
dBRrt1
(λ04,ω +λ14,ω )2
V ar(dBRrt1 )
—
ι2br
V ar(dBRrt1 )
(λ04,u +λ14,u )2
V ar(dBRrt1 )
BRint1t
(λ05,ω +λ15,ω )2
V ar(BRint1t )
—
—
(λ05,u +λ15,u )2
V ar(BRint1t )
2.3
Estimation Method
As in Subsection 1.3, the factor model described in the eq.(3) uses a Generalised Method
of Moments (GMM) estimator, which produces consistent, asymptotically normal and
efficient estimates (Hansen, 1982). GMM estimation focuses on the information only
contained by the moments of the data. The goal is to compute the unknown parameters
by matching the theoretical moments of the model to the empirical moments of the
data in both intervention-day and non-intervention-day sets.
The identification and estimation of the model make use of the precise known
29
days of currency intervention in the Brazilian foreign exchange market.14 The model
described in eq.(3), in this sense, is exactly identified. Each dataset of intervention and
non-intervention days provides 15 empirical moments. Thus, 30 empirical moments in
total, which are used to identify and estimate the 30 unknown parameters.
Let H j be a T j -by-15 empirical matrix (T j daily observations in each dataset j,
15 contemporaneous cross-products between the 5 standardised variables in Ytj for
j = 0, 1), then



Hj = 

j
j
Y1,1
Y1,1
j
j
Y1,2
Y1,2
..
.
...
...
..
.
j
j
Y5,1
Y5,1
j
j
Y5,2
Y5,2
..
.
j
j
Y2,1
Y1,1
j
j
Y1,2
Y2,2
..
.
...
...
..
.
j
j
Y5,1
Y1,1
j
j
Y1,2
Y5,2
..
.
j
j
Y3,1
Y2,1
j
j
Y2,2
Y3,2
..
.
j
j
Y5,1
Y4,1
j
j
Y4,2
Y5,2
..
.
...
...
..
.
j
j
j
j
j
j
j
j
j
j
j
j
Y1,T
. . . Y5,T
Y1,T
. . . Y1,T
Y2,T
. . . Y4,T
j Y1,T j
j Y5,T j
j Y2,T j
j Y5,T j
j Y3,T j
j Y5,T j
It is straightforward to see that, by the law of large numbers, the average value
of each column of H j asymptotically converges, respectively, to the true second-order
j
) for i, j = {1, 2, 3, 4, 5}.15 In this case, an optmisation problem
expectation E(Yi,tj , Yj,t
is solved by guessing possible parameter values in Λj of eq.(3) such that minimises the
difference between the empirical moments extracted from the columns of H j and the
0
theoretical moments derived from the lower diagonal entries of Λj Λj .




j j0
Λ Λ =


j
j
E(Y1,t
Y1,t
)
j
j
E(Y1,t Y2,t )
j
j
E(Y1,t
Y3,t
)
j
j
E(Y1,t Y4,t )
j
j
E(Y1,t
Y5,t
)

j
j
E(Y2,t
Y2,t
)
j
j
j
j
E(Y2,t Y3,t ) E(Y3,t
Y3,t
)
j
j
j
j
j
j
E(Y2,t Y4,t ) E(Y3,t Y4,t ) E(Y4,t
Y4,t
)
j
j
j
j
j
j
j
j
E(Y2,t
Y5,t
) E(Y3,t
Y5,t
) E(Y4,t
Y5,t
) E(Y5,t
Y5,t
)






Lastly, in order to calculate the standard errors of the estimated parameters, a
bootstrap procedure (Efron and Tibshirani, 1994) separetly resamples both datasets
(j = {0, 1}) 1000 times.
14
For similar strategy, see Fry-McKibbin and Wanaguru (2012) and Dungey et al. (2010).
j
j
j
j
≡ dP Ertj , Y4,t
≡ dBRrtj ,
For clarity, please mind the notation: Y1,t
≡ dARrtj , Y2,t
≡ dCHrtj , Y3,t
j
j
Y5,t ≡ BRintt .
15
30





2.4
Analysis of the Results
The volatility decomposition of the factor model described in eq.(3) is presented in
Table 10. Recapping, the contributing factors for the reserve changes and Brazil’s
central bank intervention are: ‘Global’, common to all variables; ‘Neighbour Country’,
related to countries other than the one where the currency intervention is held; ‘Intervention’, which grasps the impact of Brazil’s central bank intervention on the reserve
changes of Brazil, Argentina, Chile and Peru;16 and, lastly, ‘Residual’, which catches
the idiosyncracies of each separate market.
The top panel of Table 10 provides the percentage contribution of the orthogonal
latent factors for the days with no central bank currency intervention in Brazil. It is
clear the lack of common factors behind the reserve changes in the non-intervention day
sample. For instance, Chile’s and Peru’s volatilities are mainly determined by their
respective residual factors – 99.84% and 95.75%. On the other hand, Argentina’s
volatility is mostly explained by the neighbour-country factor. The fact that the
neighbour-country factor barely impacts Chile’s and Peru’s reserve changes suggests
that it actually serves de facto as a second Argentina’s residual factor. Brazil’s reserve
changes are completely led by the global factor, impacting 99.96% of its volatility. Indeed, in the non-intervention days, the global factor also works as a ‘Brazil’ country
factor, since in addittion to explaining Brazil’s reserve changes, it also accounts for
20.10% of Brazil’s central bank intervention volatility – while hardly impacting other
markets.
On the intervention days, a richer dynamics appears. Apart from Chile’s volatility,
which is mostly due to its residual factor,17 all other variables seem to share common
factors. The global factor accounts for one-fifth to one-fourth of the volatility of the
variables – dARrt1 (18.21%), dP Ert1 (23.51%), dBRrt1 (26.37%) and BRint1t (24.74%).
The neighbour-country factor, on its turn, explains 14.44% and 39.54% of Argentina’s
and Peru’s reserve change volatility, respectively.
Lastly, the impact of Brazil’s central bank intervention (the intervention factor),
as expected, is clear in Brazil’s reserve changes, accounting for 25.97% of its volatility. Additionally, the regional spillover effects of Brazil’s central bank intervention on
16
The intervention factor is derived from the impact of the residual factor of Brazil’s central bank
intervention (u5,t ) into the reserve stock changes of Brazil, Argentina, Chile and Peru through the
respective parameters ιbr , ιar , ιch and ιpe – see eq.(3).
17
Indeed, the parameter estimates for Chile’s global factor (λ02,w , λ12,w ) and neighbour-country factor
0
(λ2,η , λ12,η ) are all statistically insignificant at 10% level, see Table 12.
31
Table 10: Volatility Decomposition for Model 2
Notes: Contribution of each factor to total volatility, in percent. The model is estimated over
the period May 4, 2009–June 29, 2012 (see eq.(3) and Table 9).
Global
Non-intervention days (j = 0)
dARrt0
dCHrt0
dP Ert0
dBRrt0
BRint0t
Intervention days (j = 1)
dARrt1
dCHrt1
dP Ert1
dBRrt1
BRint1t
Factors
Neighbour Country Intervention
Residual
Total
5.58
93.83
—
0.59
100.0
0.14
0.01
—
99.84
100.0
2.38
1.87
—
95.75
100.0
99.96
—
—
0.04
100.0
20.10
—
—
79.90
100.0
18.21
14.44
5.00
62.34
100.0
0.02
0.93
0.65
98.40
100.0
23.51
39.54
5.85
31.10
100.0
26.37
—
25.97
47.66
100.0
24.74
—
—
75.26
100.0
neighbouring countries are present in Argentina’s and Peru’s reserve changes, accounting for 5.00% and 5.85% of their respective volatility, yet at a negligible level in Chile’s
(0.65%).
Table 11 reports on the statistical significance of intervention parameters and joint
structural breaks. Wald tests on joint parameters confirm the validity of the model,
including the structural break imposed on intervention days. Moreover, all intervention
parameters (ιbr , ιar , ιch and ιpe ) are statistically significant at 10% level of confidence.
Table 12 displays the parameter estimates and their respective p-values. Accordingly, the signs of the intervention parameters do not entirely support the peacock
effect – at least in the short term. Whereas Brazil’s central bank intervention leads
to an increase in Brazil’s and Chiles reserve levels (positive values for ιbr and ιch ), the
opposite happens in Argentina and Peru (negative values for ιar and ιpe ). That is,
Brazil’s central bank intervention leads to a decrease in the reserve levels of Argentina
and Peru. Further research should explore the reasons behind this feature and, data
availability permitting, extend the analysis to other countries.
32
Table 11: Wald Tests of Intervention and Structural Breaks in Factor Model 2
Notes: A bootstrap procedure (resampling 1000 times) was used to calculate the variancecovariance matrix of the parameters. The model is estimated over the period May 4, 2009–
June 29, 2012 (see eq.(3) and Table 9). DOF stands for degrees of freedom.
Intervention Parameters
ιbr
Estimates
0.5817
Standard Deviation
0.2279
p-value
0.005
ιar
-0.2410
0.1530
0.058
ιch
0.0732
0.0505
0.074
ιpe
-0.2175
0.1382
0.058
DOF
Test Statistic
p-value
4
31.41
0.000
=0
5
152.08
0.000
Joint structural break parameters
H0 : ιbr = ιar = ιch = ιpe = λ1i,f = 0
17
467.38
0.000
Wald Test Hypothesis
Joint Intervention parameters
H0 : ιbr = ιar = ιch = ιpe = 0
H0 : ιbr = ιar = ιch = ιpe =
λ15,u
Table 12: Parameter Estimates in Factor Model 2
Notes: A bootstrap procedure (resampling 1000 times) was used to calculate the variancecovariance matrix of the parameters. The model is estimated over the period May 4, 2009–
June 29, 2012 (see eq.(3) and Table 9).
Parameters
λ01,w
λ02,w
λ03,w
λ04,w
λ05,w
λ01,η
λ02,η
λ03,η
λ01,u
λ02,u
λ03,u
λ04,u
λ05,u
ιbr
ιar
Estimates
0.2510
-0.0449
0.1552
0.8454
0.2275
1.0293
0.0146
0.1375
-0.0817
-1.1935
-0.9839
0.0159
-0.4536
0.5817
-0.2410
p-value
0.054
0.243
0.003
0.000
0.000
0.001
0.294
0.027
0.250
0.000
0.000
0.267
0.000
0.005
0.058
Parameters
λ11,w
λ12,w
λ13,w
λ14,w
λ15,w
λ11,η
λ12,η
λ13,η
λ11,u
λ12,u
λ13,u
λ14,u
λ15,u
ιch
ιpe
33
Estimates
0.1639
0.0329
0.3281
-0.3028
0.3402
-0.6598
-0.1021
-0.7643
-0.6860
0.2935
1.5399
-0.7113
1.4439
0.0732
-0.2175
p-value
0.152
0.244
0.037
0.031
0.074
0.028
0.103
0.001
0.038
0.035
0.001
0.000
0.000
0.074
0.058
3
Conclusion
The recent uprise in central bank foreign reserve stocks around the globe has sparked
lively debate in the literature. Part of the suggested rationalle for reserve accumulation
lies on the mercantilist motive hypothesis, that is, countries would accumulate foreign
reserves in order to support export promotion by influencing exchange rates and/or to
signal economic strenght as a modern version of bullionism.
Using a unique dataset on daily foreign exchange intervention, this paper investigates the mercantilist motive hypothesis in the case of Brazil. A latent factor model,
using a GMM estimation method, is devised to determine the impact of Brazil’s central
bank currency intervention on the overall volatility of the sample variables.
The paper tackles the issue in two sections. Section 1 explores the link between
foreign reserves and exchange rate volatility. In particular, a latent factor model is
estimated to decompose the contribution of Brazil’s central bank intervention to the
overall volatility in the currency market. Accordingly, results support the effectiveness
of Brazil’s currency intervention during the sample period (May, 2009 to June, 2012).
Besides, the mercantilist motive hypothesis is also validated: Brazil’s reserve changes
contitutes a good proxy for currency intervention. Benchmark results show that currency intervention, or reserve changes as a proxy, accounts for 6–7% of the volatility
in the Brazilian currency. Lastly, it is noted that Brazil’s currency intervention has
spillover effects in Argentina and other Latin American countries, which opens room
for further research.
Section 2 investigates further the regional spillover effects of reserve accumulation
through central bank intervention. Accordingly, a latent factor model looks to the
empirical evidence of reserve stock co-movements between neighbouring countries due
to deliberate central bank intervention on foreign exchange market. In particular, it
investigates the impact of Brazil’s central bank intervention on the volatility decomposition of reserve changes in Brazil, Argentina, Chile and Peru. On balance, after
controling for global, regional and domestic factors, results confirm the spillover effects
of Brazil’s central bank intervention on foreign reserves among neighbouring countries.
Yet, parameter estimates do not support the hypothesis of reserve changes as a result of a modern bullionism pratice; leaving co-movements in currency intervention
across neighbouring countries to be the driving force behind such regional intervention
spillovers.
34
Overall, this paper contributes to the literature of foreign reserves. First, it provides
evidence of mercantilist motives for reserve accumulation in Brazil. During the sample
period, Brazil’s central bank has successfully intervened in its foreign exchange market, with a sizable by-product reserve build-up. Second, significant regional spillover
effects of Brazil’s central bank intervention into neighbouring countries are detected,
impacting the volatility of the exchange rates and reserve changes in the region. These
results invite further research into the cross-country links between reserve accumulations, which might provide additional testimony of foreign reserve mercantilist motives.
References
Aizenman, J. and Lee, J. (2007). International reserves: Precautionary versus mercantilist views, theory and evidence. Open Economies Review, 18(2):191–214.
Calvo, G. A., Izquierdo, A., and Loo-Kung, R. (2012). Optimal holdings of international reserves: Self-insurance against sudden stop. Working Paper 18219, National
Bureau of Economic Research.
Calvo, G. A. and Reinhart, C. M. (2002). Fear of floating. The Quarterly Journal of
Economics, 117:379–408.
Dooley, M., Folkerts-Landau, D., and Garber, P. (2005a). International financial stability: Asia, interest rates and the dollar. Technical report, Deustche Bank Collected
Volume.
Dooley, M. P., Folkerts-Landau, D., and Garber, P. (2003). An essay on the revived
bretton woods system. Working Paper 9971, National Bureau of Economic Research.
Dooley, M. P., Folkerts-Landau, D., and Garber, P. (2004a). Direct investment, rising
real wages and the absorption of excess labor in the periphery. Working Paper 10626,
National Bureau of Economic Research.
Dooley, M. P., Folkerts-Landau, D., and Garber, P. (2004b). The revived bretton woods
system: The effects of periphery intervention and reserve management on interest
rates & exchange rates in center countries. Working Paper 10332, National Bureau
of Economic Research.
35
Dooley, M. P., Folkerts-Landau, D., and Garber, P. M. (2004c). The US current account
deficit and economic development: Collateral for a total return swap. Working Paper
10727, National Bureau of Economic Research.
Dooley, M. P., Folkerts-Landau, D., and Garber, P. M. (2005b). Interest rates, exchange rates and international adjustment. Working Paper 11771, National Bureau
of Economic Research.
Dooley, M. P., Folkerts-Landau, D., and Garber, P. M. (2005c). Savings gluts and
interest rates: The missing link to Europe. Working Paper 11520, National Bureau
of Economic Research.
Dooley, M. P., Folkerts-Landau, D., and Garber, P. M. (2008). Will subprime be a twin
crisis for the United States? Working Paper 13978, National Bureau of Economic
Research.
Dooley, M. P., Folkerts-Landau, D., and Garber, P. M. (2009). Bretton woods II still
defines the international monetary system. Working Paper 14731, National Bureau
of Economic Research.
Dooley, M. P., Garber, P. M., and Folkerts-Landau, D. (2007). The two crises of international economics. Working Paper 13197, National Bureau of Economic Research.
Dungey, M., Fry, R., Gonzalez-Hermosillo, B., Martin, V., and Tang, C. (2010). Are
financial crises alike? Technical Report WP/10/14, IMF Working Paper.
Durdu, C. B., Mendoza, E. G., and Terrones, M. E. (2007). Precautionary demand for
foreign assets in sudden stop economies: An assessment of the new merchantilism.
Working Paper 13123, National Bureau of Economic Research.
Efron, B. and Tibshirani, R. (1994). An Introduction to the Bootstrap. Chapman &
Hall/CRC.
Fry-McKibbin, R. A. and Wanaguru, S. (2012). Currency intervention: A case study
of an emerging market. Technical report, CAMA Working Paper Series.
Hansen, L. (1982). Large sample properties of generalised method of moments estimators. Econometrica, 50:1029–1054.
36
Jeanne, O. and Ranciere, R. (2011). The optimal level of international reserves for
emerging market countries: A new formula and some applications. The Economic
Journal, 121:905–930.
Menkhoff, L. (2012). Foreign exchange intervention in emerging markets: A survey of
empirical studies. Diskussionspapiere der Wirtschaftswissenschaftlichen Fakultt der
Leibniz Universitt Hannover dp-498, Leibniz Universitt Hannover, Wirtschaftswissenschaftliche Fakultt.
Obstfeld, M., Shambaugh, J. C., and Taylor, A. M. (2010). Financial stability, the
trilemma, and international reserves. American Economic Journal: Macroeconomics,
2(2):57–94.
37
Appendix
Table 13: Parameter Estimates in Factor Models A and B
Notes: A bootstrap procedure (resampling 1000 times) was used to calculate the variancecovariance matrix of the parameters. The model is estimated over the period May 4, 2009–
June 29, 2012 (see eq.(1) and Table 1). Results for ιbr and ιar , see Tables 3 and 5.
Intervention Parameters
λ01,w
λ02,w
λ03,w
λ04,w
λ05,w
λ06,w
λ01,κ
λ02,κ
λ03,κ
λ05,κ
λ05,br
λ06,br
λ03,ar
λ04,ar
λ01,u
λ02,u
λ03,u
λ04,u
λ05,u
λ06,u
λ11,w
λ12,w
λ13,w
λ14,w
λ15,w
λ16,w
λ11,κ
λ12,κ
λ13,κ
λ15,κ
λ15,br
λ16,br
λ13,ar
λ14,ar
λ11,u
λ12,u
λ13,u
λ14,u
λ15,u
λ16,u
ιbr
ιar
Model A
Estimates p-values
-0.3010
0.008
-0.2570
0.008
-0.0178
0.216
0.6887
0.024
-0.3445
0.005
0.0521
0.117
0.8137
0.000
0.6822
0.000
-0.0694
0.186
0.6881
0.000
-0.7650
0.000
0.0029
0.212
0.8650
0.000
-0.1218
0.018
-0.4850
0.003
-0.5317
0.000
0.3928
0.087
-0.7999
0.000
-0.3352
0.141
-0.5047
0.000
1.0420
0.000
0.9131
0.000
-0.0682
0.178
-0.7771
0.016
1.0334
0.000
0.6209
0.001
-0.1397
0.176
-0.4220
0.027
0.2635
0.020
-0.7115
0.006
1.3510
0.000
0.6054
0.001
-1.3078
0.002
0.1899
0.153
0.4850
0.094
1.2908
0.000
0.2338
0.186
-0.1660
0.217
0.3352
0.178
0.2667
0.063
-0.2473
0.085
-0.6343
0.013
Model B
Estimates p-values
-0.6516
0.000
-0.5748
0.000
0.0224
0.263
0.3144
0.017
-0.7495
0.000
0.6989
0.000
0.5977
0.000
0.4310
0.000
-0.0651
0.212
0.2947
0.014
-0.2346
0.087
-0.3002
0.011
0.1228
0.012
-0.0147
0.000
-0.4540
0.014
0.5459
0.000
-0.9423
0.000
0.0224
0.231
-0.7656
0.000
-0.3656
0.024
0.0896
0.102
0.0817
0.087
-0.0736
0.151
0.2261
0.081
-0.0600
0.179
1.3717
0.000
-1.2487
0.000
-0.6997
0.000
-0.1473
0.118
-0.2696
0.082
0.3555
0.028
0.3131
0.026
-0.5707
0.041
1.0840
0.005
0.2806
0.158
0.2101
0.013
1.4829
0.001
-0.9882
0.030
0.7656
0.030
0.1401
0.177
-0.2354
0.042
-0.6993
0.003
Table 14: Volatility Decomposition for Model A
Notes: Contribution of each factor to total volatility, in percent. The model is estimated over
the period May 4, 2009–June 29, 2012 (see eq.(1), Table 1 and Section 1.4.3).
Global Currency
Non-intervention days (j = 0)
rEU Rt0
0.59
73.35
rU KPt0
rCHc0t
dCHrt0
rBRc0t
BRint0t
Factors
Brazil Argentina
Intervention
Residual
Total
—
—
—
26.05
100.0
1.98
68.91
—
—
—
32.11
100.0
15.61
39.11
—
20.47
—
25.82
100.0
17.20
—
—
18.41
—
64.39
100.0
5.56
50.24
39.81
—
—
4.39
100.0
0.10
—
0.46
—
—
99.44
100.0
Intervention days (j = 1)
rEU Rt1
34.22
24.96
—
—
—
40.82
100.0
26.09
45.86
—
—
—
28.05
100.0
46.62
0.03
—
7.44
16.89
29.02
100.0
0.03
—
—
1.65
—
98.32
100.0
81.30
0.00
2.65
—
16.04
0.00
100.0
80.89
—
3.37
—
—
15.74
100.0
rU KPt1
rCHc1t
dCHrt1
rBRc1t
BRint1t
Table 15: Volatility Decomposition for Model B
Notes: Contribution of each factor to total volatility, in percent. The model is estimated over
the period May 4, 2009–June 29, 2012 (see eq.(1), Table 1 and Section 1.4.3).
Global Currency
Non-intervention days (j = 0)
rEU Rt0
60.37
19.05
rU KPt0
rCHc0t
dCHrt0
rBRc0t
dBRrt0
Factors
Brazil Argentina
Intervention
Residual
Total
—
—
—
20.58
100.0
50.49
12.75
—
—
—
36.76
100.0
35.56
0.50
—
0.18
—
63.76
100.0
0.04
—
—
98.07
—
1.90
100.0
84.82
8.82
6.36
—
—
0.00
100.0
50.12
—
49.88
—
—
0.00
100.0
Intervention days (j = 1)
rEU Rt1
34.40
35.28
—
—
—
30.31
100.0
26.20
32.17
—
—
—
41.63
100.0
56.58
0.73
—
0.15
25.88
16.67
100.0
0.03
—
—
89.17
—
10.79
100.0
80.54
0.00
0.08
—
19.39
0.00
100.0
80.50
—
0.13
—
—
19.36
100.0
rU KPt1
rCHc1t
dCHrt1
rBRc1t
dBRrt1
Table 16: Volatility Decomposition for Model A
Notes: Contribution of each factor to total volatility, in percent. The model is estimated over
the period May 4, 2009–June 29, 2012 (see eq.(1), Table 1 and Section 1.4.3).
Global Currency
Non-intervention days (j = 0)
rEU Rt0
98.27
1.73
rU KPt0
rP Ec0t
dP Ert0
rBRc0t
BRint0t
Factors
Brazil Argentina
Intervention
Residual
Total
—
—
—
0.00
100.0
55.37
6.27
—
—
—
38.37
100.0
19.48
1.58
—
0.09
—
78.84
100.0
7.45
—
—
82.29
—
10.25
100.0
39.97
8.89
0.02
—
—
51.12
100.0
0.11
—
95.73
—
—
4.16
100.0
Intervention days (j = 1)
rEU Rt1
91.66
8.34
—
—
—
0.00
100.0
69.41
30.59
—
—
—
0.00
100.0
6.21
0.46
—
58.34
26.12
8.88
100.0
1.19
—
—
1.11
—
97.70
100.0
30.07
0.00
46.51
—
23.42
0.00
100.0
30.10
—
46.46
—
—
23.44
100.0
rU KPt1
rP Ec1t
dP Ert1
rBRc1t
BRint1t
Table 17: Volatility Decomposition for Model B
Notes: Contribution of each factor to total volatility, in percent. The model is estimated over
the period May 4, 2009–June 29, 2012 (see eq.(1), Table 1 and Section 1.4.3).
Global Currency
Non-intervention days (j = 0)
rEU Rt0
79.63
0.20
rU KPt0
rP Ec0t
dP Ert0
rBRc0t
dBRrt0
Factors
Brazil Argentina
Intervention
Residual
Total
—
—
—
20.17
100.0
67.31
36.69
—
—
—
0.00
100.0
22.47
0.22
—
0.30
—
77.02
100.0
7.43
—
—
43.68
—
48.90
100.0
43.41
0.01
14.70
—
—
41.88
100.0
38.67
—
2.73
—
—
58.60
100.0
Intervention days (j = 1)
rEU Rt1
91.66
8.34
—
—
—
0.00
100.0
69.41
30.59
—
—
—
0.00
100.0
6.21
0.46
—
1.25
9.06
83.02
100.0
1.19
—
—
51.82
—
46.99
100.0
30.07
0.00
2.42
—
67.51
0.00
100.0
30.10
—
2.35
—
—
67.55
100.0
rU KPt1
rP Ec1t
dP Ert1
rBRc1t
dBRrt1