ArticlePDF Available

Tinker, Taper, QE, Bye? The Effect of Quantitative Easing on Financial Flows to Developing Countries

Authors:

Abstract and Figures

This paper examines the effects of quantitative easing (QE) policies in the United States on gross financial inflows to developing countries. Our results support the notion that QE may have been transmitted through liquidity, portfolio balancing, and confidence channels. Moreover, we find that QE had an additional effect over and above these observable channels , which we cannot attribute to either market expectations or changes in the structural relationships between inflows and the observable fundamentals. Our baseline estimates place the lower bound of the effect of QE at around 3 percent of gross inflows, for the average developing economy. We also find evidence of heterogeneity among different types of flows; portfolio (especially bond) flows tend to be more sensitive than FDI to our measured QE effects. Finally, we perform a simulations to explore the potential effects of QE withdrawal on financial flows to developing countries.
Content may be subject to copyright.
Tinker, Taper, QE, Bye? The Effect of
Quantitative Easing on Financial Flows
to Developing Countries
Jamus Jerome Lim, Sanket Mohapatra, and Marc Stocker
January 15, 2014
Abstract
This paper examines the effects of quantitative easing (QE) policies in
the United States on gross financial inflows to developing countries. Our
results support the notion that QE may have been transmitted through
liquidity, portfolio balancing, and confidence channels. Moreover, we find
that QE had an additional effect over and above these observable chan-
nels, which we cannot attribute to either market expectations or changes
in the structural relationships between inflows and the observable fun-
damentals. Our baseline estimates place the lower bound of the effect
of QE at around 3 percent of gross inflows, for the average developing
economy. We also find evidence of heterogeneity among different types of
flows; portfolio (especially bond) flows tend to be more sensitive than FDI
to our measured QE effects. Finally, we perform a simulations to explore
the potential effects of QE withdrawal on financial flows to developing
countries.
Keywords: International financial flows, quantitative easing, developing
countries
JEL Classification: F21, F32, O19
The World Bank. Respective emails: jlim@worldbank.org, smohapatra2@worldbank.org,
and mstocker1@worldbank.org. This paper served as a background paper for the World Bank’s
Global Economic Prespects, 1st half 2014, report. We thank, without implicating, Andrew
Burns, Sergio Schmukler, and Luis Serv´en for valuable comments that substantially improved
the quality of the work. All errors remain firmly our own. The findings, interpretations, and
conclusions expressed in this article are entirely those of the authors. They do not necessarily
represent the views of the World Bank, its Executive Directors, or the countries they represent.
“We’ve had enough.” He took back the report and jammed it under
his arm. “We’ve had a bellyful, in fact.” “And like everyone who’s
had enough,” said Control as Alleline noisily left the room, “he wants
more.”
John le Carr´e,
Tinker, Tailor, Soldier, Spy (1974, p. 97)
There is no fixed calendar schedule [for tapering]. I really have to
emphasize that. We could move later this year, but even if we do
that subsequent steps are tied to the economic [data].
Ben S. Bernanke,
Press conference announcing delay of tapering (Sep 18, 2013)
1 Introduction
In late November 2008, the United States Federal Reserve announced an hitherto
unprecedented policy of unconventional monetary intervention, involving a $600
billion purchase of mortgage-backed securities, in addition to its traditional
purchases of U.S. Treasury notes. This policy—which has come to be known as
quantitative easing (QE)—was designed with the explicit purpose of bolstering
weak asset markets, as well as stimulating real activity, given the perceived
constraints of the zero lower bound on short-term interest rates (which became
a binding constraint that very month). Over the course of the succeeding five
years, the Fed engaged in a total of three separate QE episodes, so that by the
beginning of the Fed’s gradual unwinding in January 2014, its balance sheet had
more than doubled, to $4 trillion.
Although QE was meant to be an expansionary monetary policy for the
U.S. economy, the program had profound implications for developing countries.
Faced with near-zero returns in the U.S. and other high-income countries—
many of which were implementing unconventional monetary policies of their
own—financial capital began to seek alternative sources of yield. Emerging
economies, which had enjoyed heady growth rates and stable political-economic
environments over the past decade, appeared to be an ideal investment alter-
native. The data appear to corroborate this possibility: between 2000 and
mid-2013, annual gross capital inflows to developing countries grew by an order
of magnitude, to $1.8 trillion (annualized).
However, cross-border financial flows respond to a host of standard “funda-
mental” factors—such as interest rates—and circumstantial evidence pointing to
the coincidence if QE and developing country inflows are insufficient to ascertain
whether there are any genuine effects that accrue to the Fed’s unconventional
monetary policy. Furthermore, since QE policies are also designed to influence
these fundamental variables associated with capital flows, the potential effects
of QE may be clouded by changes in these conventional measures that have
nothing to do with unconventional monetary policy.
2
In this paper, we bring data to bear on the question of whether the QE
period may have seen changes in gross financial inflows to developing countries,
and if so, under what channels QE may have been transmitted, and whether
these mechanisms may have differed for different types of financial flows. We go
on to ask the equally-important question of what would happen to such flows
were the Federal Reserve to withdraw QE policies.
Our approach to answering this question is modest. Rather than ascribe a
specific, quantitative estimate to the total effect of QE—which would require
us to first ascertain the impact of QE on a host of secondary variables—our
strategy is to begin by accounting for potential QE spillover effects through
standard transmission channels—namely via liquidity, portfolio balancing, and
confidence—and then seek to establish whether QE episodes saw any additional
effects on financial inflows attributable to unobservables. Although such an
estimate is likely to underestimate the total effects of QE (since the total effect
of QE has to account for its effect via standard channels as well), it sidesteps
the challenging problem of disentangling the relative contributions of QE versus
other factors (such as overall economic activity) on fundamental measures, such
as the interest rate. Nevertheless, our approach is biased toward a negative
outcome, so if indeed we do find an independent effect during QE episodes, we
can be fairly certain that QE played some role in influencing the movement of
cross-border financial flows.
We rely on data for gross capital inflows across as many as 60 developing
countries, using quarterly data for the period between 2000Q1 and 2013Q2. Our
focus on gross flows is motivated, in part, by their greater volatility in develop-
ing countries; this is especially pertinent in the context of seeking to understand
how financial flows move in a post-crisis environment (Broner, Didier, Erce &
Schmukler 2013; Janus & Riera-Crichton 2013). We pair these gross flows data
with global and country-specific variables that are selected to capture transmis-
sion along standard channels, and complement these observable measures with
an indicator corresponding to the different episodes of QE. Our baseline esti-
mation also takes into account (time-invariant) country-specific unobservables,
a time trend, and the discrete fall in capital flows following the financial crisis
of 2008.
Our analysis indicates that financial inflows to developing economies operate
along all three potential channels of transmission for QE. More importantly, we
find evidence that episodes of QE were also accompanied by increases in inflows
over and above these observable channels. This effect is economically significant:
an average developing country experienced, during episodes of QE, increases
in inflows of 3 percent, a magnitude comparable to a one standard deviation
change along the traditional channels, such as changes in yield curve spreads
(the portfolio channel) or the short-term interest rate (the liquidity channel).
This unmeasured QE effect cannot be ascribed to the unprecedented nature
of economic conditions during and after the financial crisis. We find relatively
little evidence that the sensitivity of our observable measures for the different
transmission channels changed during the QE period. Our results also point
to heterogeneity in the response of different types of inflows during the period.
3
When we decompose aggregate gross flows into their constituent components, we
find that foreign direct investment (FDI) does not vary along either observable
transmission channels or our QE indicator, whereas portfolio (and in particular
bond) flows do, especially along global “push” factors associated with economic
conditions in high-income countries.
We then use our results to perform a simulation that explores the effect of
QE withdrawal on inflows to emerging markets, relative to a no change status
quo. This is a central concern for developing economies, which have struggled
to cope with the surge in financial inflows that they have experienced over
the past several years, and are fearful that the renormalization of high-income
monetary policies will be accompanied by a disorderly sudden stop in capital
inflows. Their fear is far from unfounded: we find that when the Federal Reserve
issued forward guidance that it would begin “tapering” its asset purchases, the
expectation effects along may have accounted for a reduction in inflows in the
order of a tenth of a percent. Across the three tapering scenarios we consider,
we find that inflows to developing countries decline by a cumulative amount of
between 10 to 12 percent (approximately 0.6 percent of GDP), with the main
differences between the scenarios captured by the speed of adjustment in the
earlier versus later years.
Our work speaks to several literatures. The general nature of the questions
we pose have been examined by a vast theoretical (Betts & Devereux 2000, 2004;
Fukuda 1993; Obstfeld & Rogoff 1995; Turnovsky 1986) and empirical (Ammer,
Vega & Wongswan 2010; Ehrmann & Fratzscher 2009; Jannsen & Klein 2011;
Kazi, Wagan & Akbar 2013; Kim 2001; Xiao 2011) literature on cross-border
spillover effects of monetary policies. The vast majority of the papers in this
vein are concerned with interest rates; more precisely, the effect of changes
in the interest rate (or monetary base) on other macroeconomic and financial
market variables. In contrast to these papers, we are concerned primarily with
unconventional monetary policies in the form of quantitative easing, which is
particularly relevant in the zero-lower bound environment of the past few years.
There has also been a small literature that emerged following the imple-
mentation of QE, which has sought to quantify the effects of QE on a range
of macroeconomic phenomena, especially interest rates (Christensen & Rude-
busch 2012; Gagnon, Raskin, Remache & Sack 2011; Krishnamurthy & Vissing-
Jørgensen 2011; Pesaran & Smith 2012) and term premia in financial markets
more generally (Bauer & Rudebusch 2013; Breedon, Chadha & Waters 2012;
Joyce, Lasaosa, Stevens & Tong 2011), but also output and inflation (Chen,
C´urdia & Ferrero 2012; Kapetanios, Mumtaz, Stevens & Theodoridis 2012). By
and large, this family of papers find somewhat modest but nontrivial effects,
consistent with a 75–100 basis point reduction in the policy rate (although evi-
dence on real effects are more ambiguous). However, these papers are uniformly
concerned with the effects of QE on the home economy (in particular the United
States), while we are interested more in the spillover effects as they play out in
other countries, especially those in the developing world.
Of course, there is a massive literature that examines the determinants of
financial inflows, including those that explore the conditions where capital flows
4
may experience surges (Agosin & Huaita 2012; Forbes & Warnock 2012a,b;
Reinhart & Reinhart 2008) and where developing countries may be dispropor-
tionately affected (Brana & Lahet 2010; Chuhan, Claessens & Mamingi 1998;
Hutchison & Noy 2006; Sarno & Taylor 1999). Our modeling approach takes af-
ter these papers, but our explicit goal is to capture the effects of QE on financial
flows, rather than exploring the role of a broader set of determinants. Moreover,
given the relatively brief windows in which QE operations have been in effect,
we emphasize slightly higher-frequency (quarterly) movements in capital flows.
Probably the papers closest in spirit to our own are those by Ahmed & Zlate
(2013), International Monetary Fund (2013), Bauer & Neely (2013), Eichengreen
& Gupta (2013), and Fratzscher, Lo Duca & Straub (2013). While these papers
all consider the international dimensions of QE, the last three are not primarily
concerned with gross financial inflows, nor do they seek to break down hetero-
geneous effects among the distinct constituents of gross flows. As explained
above, our attention to gross inflows is motivated by their greater responsive-
ness to monetary phenomena, which we regard as especially important in un-
derstanding potential cross-border spillover effects, and especially important for
developing economies. International Monetary Fund (2013) does consider gross
flows in some detail, but they limit their analysis to inflows into equity and
bond funds only (as opposed to our broader mix of flows). Finally, while the
paper by Ahmed & Zlate (2013) does deal substantially with a broad range of
gross capital flows and QE, the focus is not on the different transmission chan-
nels involved. Furthermore, none of these papers engage with our substantive
question regarding the implications of QE tapering.
The rest of this paper is organized as follows. In the following section, we
provide some background on the unconventional monetary policies pursued by
the Federal Reserve following the financial crisis of 2008. Section 3 reviews the
theoretical literature on transmission channels for monetary policy in general,
and quantitative easing in particular. We describe our empirical methodology,
data sources, and key variable definitions in Section 4. Sections 5 and 6 report
our baseline results and an array of robustness tests, including the addition of ex-
planatory variables, alternative measures of our main variables, and alternative
estimation techniques. The latter section also considers, in detail, differences
between the behavior of different types of financial flows. Our penultimate sec-
tion performs a number simulations for the withdrawal of QE, before a final
section concludes with some policy reflections.
2 Unconventional monetary policies in the af-
termath of the 2008 financial crisis
In the months leading up to and following the Lehman crisis in the United
States in August 2008, the US Federal Reserve—along with central banks in a
number of other high income economies (the Bank of England, the European
Central Bank, and the Bank of Japan)—sharply cut policy interest rates, in
5
an effort to support demand in the face of weakening output and employment.
However, with interest rates already fairly low, the perceived constraints from
a zero lower bound on nominal interest rates prompted the US Federal Reserve
and other central banks to subsequently implement unconventional monetary
policies in the form of quantitative easing.
QE involved large-scale purchases of financial assets (LSAPs), such as long-
dated government bonds and mortgage-backed securities. These unorthodox
measures—which were eventually realized over three episodes between 2008 and
2013 (see Table A.1 in the appendix)—were initially undertaken to repair finan-
cial market functioning and intermediation during the crisis, but subsequently
evolved to support weak post-crisis recovery in growth and employment.
This extended period of highly accommodative monetary policies in high in-
come countries has been a source of significant concern among many developing
countries, who fear potential policy spillovers, primarily through an uncontrolled
increase in cross-border financial flows.
1
These fears were not unfounded. Over the four-year period between mid-
2009 and the first quarter of 2013, cumulative gross financial inflows into the
developing world rose from $192 to $598 billion, more than twice the pace com-
pared to the far more modest increase of $185 billion between mid-2002 and the
first quarter of 2006 (Figure 1(a)).
2
When expressed as a share of developing
country GDP, they more than doubled (Figure 1(b)). Developing world equity
markets also experienced substantial gains, and the many emerging economies
that received substantial volumes of inflows relative to their GDP also saw sig-
nificant appreciation of their real effective exchange rates.
The initial concerns of developing countries over unmanageable financial in-
flows have since been compounded by the possibility of disorderly capital flow
reversals. In the middle of 2013, the Fed’s anticipated exit from QE sparked sub-
stantial outflows from a number of emerging market economies. The specter of
further tapering of asset market purchases by advanced-economy central banks
could could mean increases in borrowing costs, as well as other financial market
disruptions due to the unwinding of speculative positions.
3
In January 2014,
the Fed began the long-awaited taper of asset market purchases, which marked
the beginning of the end to the period of unconventional monetary policy.
1
Emerging markets have, since the beginning, voiced their concern over unmitigated finan-
cial flows due to QE. Foreign exchange intervention by developing countries to arrest exchange
rate appreciation due to capital inflows sparked talk of “currency wars” (Mackintosh 2010)
and what many regarded as a forced buildup of foreign exchange reserves (Beckner 2013).
2
This comparison deliberately excludes the pre-crisis boom in inflows between 2006 and
2008. An alternative way to express the same sentiment is that cumulative inflows since the
beginning of QE has paralleled the increases in the immediate pre-crisis “bubble” period.
3
This anticipated tapering of QE has resulted in depreciations in a large number of
developing-country currencies (Geddes 2013), and developing economies have also expressed
concern over additional spillover effects due to the QE exit (Wigglesworth 2013).
6
-400
-300
-200
-100
0
100
200
300
400
500
600
700
2000Q1 2001Q3 2003Q1 2004Q3 2006Q1 2007Q3 2009Q1 2010Q3 2012Q1
Gross
inflows
($ billion)
Portfolio
Loans
FDI
QE2
QE3
QE1
(a) Gross inflows to developing countries, cumulative value
-5
0
5
10
15
20
2000Q1 2001Q3 2003Q1 2004Q3 2006Q1 2007Q3 2009Q1 2010Q3 2012Q1
Gross
inflows/
GDP (%)
QE2
QE3
QE1
(b) Gross inflows to developing countries, as share of GDP
Figure 1: Gross financial inflows to developing countries, by different constituent
flows, in cumulative U.S. dollars (left panel) and as share of developing country
GDP (right panel). The three quantitative easing episodes are shaded. The
sharp contraction in flows following the financial crisis in 2008 is evident. In-
creases in portfolio and loan flows over the period appear to be equiproportional,
but faster than the increase in FDI (left panel). The variations in gross flows
as a share of GDP mimic those in absolute terms, although the post-crisis re-
covery in flows remains below pre-crisis peaks when measured in this manner
(right panel).
7
3 Channels of transmission for unconventional
monetary policy
Traditional monetary policy operates along the interest rate and other asset
price channels—including exchange rates and equity prices—as well as the credit
channel, which includes bank lending and balance sheet mechanisms (Mishkin
1996). In contrast, the premise for unconventional monetary policies is that
these traditional channels are either ineffective, unavailable, or weak, which
justifies large-scale asset market interventions by the central bank.
4
A central transmission channel by which such asset purchases affect cross-
border capital flows is via the portfolio balance channel (Gagnon et al. 2011;
Hamilton & Wu 2012). QE involves the substitution of longer-duration assets
for safe long-term government bonds, which in turn reduces the available stock
of privately-held risky assets. Unmet risk appetite will thus be met by increasing
demand for other risky investments. Thus, we would expect the portfolio bal-
ance channel to be expressed both in terms of heightened demand for temporal
(longer duration) and spatial (developing country) assets, which comes about
as investors rebalance their portfolios.
Another key transmission channel for QE is the liquidity channel (Gagnon
et al. 2011; Joyce et al. 2011; Krishnamurthy & Vissing-Jørgensen 2011). The
long-term assets purchased through QE operations are credited as increased
reserves on the balance sheets of private banks. Since such reserves are more
easily traded in secondary markets than long-term securities, there is a decline
in the liquidity premium, which in turn enables previously liquidity-constrained
banks to extend credit to investors. This results in an decline in borrowing costs
and increases overall bank lending, including lending to developing countries.
Finally, QE can also play a signaling role. Large-scale asset purchases serve
as as a credible commitment to keep interest rates low even after the recovery of
the economy, since a premature increase in interest rates would imply a loss on
assets held by the central bank (Bauer & Rudebusch 2013; Clouse, Henderson,
Orphanides, Small & Tinsley 2003). Moreover, such signaling can also improve
household and business sentiment by diminishing concerns about deflation risk
(Hendrickson & Beckworth 2013); steady central bank asset market interven-
tions can also reduce volatility and hence economic uncertainty. The sum total
of these confidence channel effects is to bolster investment activity, and is the
final channel of QE transmission to developing country financial flows that we
consider.
5
4
Figure A.1 in the appendix summarizes the manner by which we conceive these linkages.
5
While we recognize that these channels that we describe do not constitute an exhaustive
list, our decision to focus on the liquidity, portfolio balance, and confidence channels is due to
three reasons. First, there is a substantial degree of overlap between some of the more esoteric
channels that have been explored in the literature, and what we identify here. For instance,
Vayanos & Vila (2009) identify a duration risk channel where central bank asset purchases
are able to alter investors’ preferred duration risk, and hence compress the yield curve. But
when interpreted more broadly a mechanism that alters term premia and hence the shape of
the yield curve, this channel—along with others, such as the safety channel (Krishnamurthy
& Vissing-Jørgensen 2012)—falls within the broader rubric of portfolio rebalancing. Second
8
4 Measuring and estimating the effects of QE
on financial flows
4.1 Econometric model and estimation methodology
Our baseline regression specification is a lagged-dependent model of the form
GF I
it
= GF I
i,t1
+ λL
it
+ πP B
it
+ χC
it
+ θQE
it
+ β
0
X
it
+ CRISIS
t
+ P OST CRISIS
t
+ α
i
+ τ
t
+
it
,
(1)
where the effects of unconventional monetary policy on gross financial inflows to
country i at time t, GF I
it
, may be transmitted via (observable) liquidity (L
it
),
portfolio balance (P B
it
), and confidence (C
it
) channels, but may also encompass
additional effects due to unobservables, which we proxy with the indicator vari-
able QE
it
. We further include dummies to account for the sharp drop in crisis
(CRISIS
t
) flows and the possibility of a post-crisis (P OST CRISIS
t
) secular
stagnation (the so-called “new normal”) (Figure 1).
6
We also include a vector
X
it
of additional time-varying idiosyncratic controls (such as the countrys GDP,
growth rate, and its risk rating) controls, country-specific fixed effects (α
i
), and
a time trend (τ
t
). N ID
0, σ
2
is the residual. Our coefficients of interest
correspond to the vector [λ π χ θ] of estimated coefficients that correspond to
the different observable and unobservable transmission channels.
Since (1) is a dynamic model with fixed effects, these estimates may be
biased for finite T (Nickell 1981). Since Given the time coverage of the dataset
is relatively long (54 quarters), we suspect that the inconsistency of estimates
should not pose a major problem (since the bias is of O
1
T
). Nevertheless, our
coefficients are estimated using bias-corrected Least Squares Dummy Variables
(LSDV) (Bruno 2005), under the strictest condition for bias approximation (up
to O
1
NT
2
), with bootstrapped standard errors.
7
In our simulations where we examine possible tapering scenarios (Section 7),
we discipline our conditioning variables by implementing a vector autoregressive
(VAR) specification comprising the various transmission channels. We thus
modify our baseline specification 1 to
TC
it
= α +
K
X
k=1
β
k
TC
i,tk
+
it
, (2)
(and relatedly), we have chosen to subsume channels that may be distinct but are likely to
be measured in a similar fashion. For example, given the difficulty of identifying proxies for
sentiment, we have chosen to fold expectational and signaling effects into a single channel,
confidence. Third, while it is clear to us what the cross-border spillover effects are of the
channels that we identify, other possible channels may have more ambiguous cross-border
implications, which justifies our exclusion.
6
These variables take on the value of unity for all quarters between 2008Q3 and 2009Q2
(inclusive), and 2009Q3 and 2013Q2 (inclusive), respectively.
7
The correction is initialized by the Anderson & Hsiao (1982) consistent estimator for β
0
,
and the bootstrapped asymptotic variance-covariance matrix is constructed with 100 replica-
tions.
9
where the matrix TC
it
= [L
it
PB
it
C
it
X
it
GFI
it
] is composed of the (global)
transmission channels, along with controls (X
it
) and aggregate developing-
country gross inflows (GFI
it
), β
k
is a vector of coefficients, and
it
is a vector
of disturbances.
8
Since this application is designed to capture the effect of the
movements in primarily global variables, i in this case applies either to the
developing or high-income aggregate.
To estimate (2), we impose additional assumptions. Our (structural) iden-
tification strategy imposes a Cholesky decomposition of the covariance matrix,
with an ordering such that the first variable cannot respond to contemporane-
ous shocks (within the same quarter) from any other variables, the second one
responds to contemporaneous shocks affecting only the first variable but no oth-
ers, and so on.
9
Since the system given by (2) produces independent estimates
of aggregate gross inflows, we use these as a useful cross-check for our scenario
projections which are based on (1).
4.2 Data sources and definitions of key variables
Our analysis draws primarily on balance of payments data from the International
Monetary Fund’s International Financial Statistics (IFS) for gross portfolio and
FDI inflows. We supplement these two flows with bank lending data from the
Bank of International Settlements’ Locational Banking Statistics (LBS).
10
We
define our main dependent variable of interest, aggregate gross financial inflows,
as the sum of changes in foreign holdings of these three categories of assets (port-
folio, FDI, and loans) in the developing economy, net of their own disinvestment
in each of these three flows. In our robustness checks, we also draw on EPFR
Global’s Global Fund Flows and Allocations Data—which compiles secondary
market transactions of bond and equity purchases in emerging market mutual
funds—to obtain a complementary gross inflow measure; we define this alter-
native measure as gross fund inflows.
11
Other additional control variables were
obtained from the IFS and the World Bank’s World Development Indicators,
supplemented by data from Haver Analytics and Datastream where gaps exist.
Our main independent variables of interest is a suite of variables designed to
8
We rule against a panel VAR specification, which would allow for country-specific controls,
because the channel variables we include in TC are not country specific; introducing country-
level controls would also further reduce the degrees of freedom in an already small sample.
9
The specificities of our VAR identification assumptions are discussed in Subection 7.1.
10
The IFS data do include a residual category, “other investments,” that includes loans as a
subcomponent. However, this category also includes other forms of cross-border finance (such
as trade credit and cash) that are of a fundamentally different nature from bank loans, which
make it harder to draw inferences when we disaggregate by flow type. The main advantage of
using the “other investments” data is that they aggregate consistently with outflows to yield
the balance of payments. Since ensuring this consistency is not important for our application,
we use the more clearly-delineated LBS data instead.
11
Since only relatively few countries report this breakdown to the IMF, the IFS data provide
relatively scant country coverage. While our alternative measure of gross inflows represents
only a relatively small segment of the total market for financial assets, it tends to closely track
actual balance of payments flows remarkably well (Miao & Pant 2012), and serves as a useful
robustness check for our main dependent variable.
10
capture the effects that occurred during QE episodes accruing to unobservables.
We consider three alternative primary measures, all of which are global in na-
ture: an indicator variable that corresponds to any of the three distinct periods
for which a QE program was implemented, separate indicator variables for each
of the three episodes, and a continuous measure of QE interventions based on
expansions in the size of the central bank’s balance sheet.
12
For the indicator
variables, our coding scheme for the start/end quarters defines a quarter as be-
longing to the implementation window if the total number of implementation
days exceeded half the days in any given quarter (e.g. QE1 operations, which
began on December 16, 2008, is coded as starting 2009Q1, while QE2, which
came into effect on November 3, 2010, is coded as beginning 2010Q4) (precise
details of this coding scheme are provided in the appendix). In our baseline,
we consider only QE operations by the U.S. Federal Reserve (which we subse-
quently expand in robustness checks to allow for QE operations in other major
advanced-economy central banks).
We use a number of distinct measures to capture each of the potential trans-
mission channels for QE. For each channel, we include a primary indicator (or
set of indicators), which we use in our more parsimonious specification, and
additional secondary indicators, which are distinct but related alternative mea-
sures that we introduce in an extended specification.
For the liquidity channel, our primary indicator is the 3-month Treasury bill
rate. This measures the effect of changes in short-term rates resulting from QE
operations.
13
Our secondary liquidity measure is the (lagged) money supply
(M2), which serves as a quantity-based measure of available liquidity.
14
While
analogous, these two variables capture slightly different notions of liquidity: the
former is a price signal that may or may not translate into actual changes to the
stock of outstanding money. Note, as well, that our use of M2 as the measure of
the money supply ensures that it overlaps only minimally with changes in the
monetary base that result from QE operations.
15
Our primary measures for the portfolio balance channel are the yield curve
(the difference between the long-term interest rate and short-term policy rates)
12
It is conceivable that these indicators are capturing unobservable effects that are never-
theless directly attributable to the observable channels, and merely reflect a change in the
structural relationship along these channels during the QE period. We explore this possibility
by examining interaction effects of the QE episode indicator with our other channel measures
in Subsection 5.2.
13
Note that reductions in the liquidity premium that result from QE will tend to lower
the price of short-term Treasuries, which is reflected in higher yields. Increased Treasury
yields raise the opportunity cost of alternative investments—including that of developing
world assets—such that, ceteris paribus, inflows can be expected to fall (implying a negative
coefficient).
14
This variable is lagged since quantity signals are generally regarded as slower to dissem-
inate than price signals. Although its interpretation is more indirect, a priori, we expect
this coefficient to be negative: an increase in M2 indicates an increase in available financing,
which lowers the liquidity premium (raises yields on liquid assets), and substitutes away from
financial investments in developing countries.
15
Pairwise correlations are also relatively low: more precisely, ρ
r
US,t
, ms
US,t
= 0.49
and ρ
ms
US,t
, mb
US,t
= 0.X.
11
and the interest rate differential between the developing country vis-`a-vis the
United States. The first is a global variable, and captures the effect that QE
can have on long-term yields, and hence temporal rebalancing toward higher-
risk asset classes, of which developing-country investments are one. The second
is a country-specific variable, and captures the more traditional spatial rebal-
ancing that arbitrages cross-country differences in yields that result from QE.
Since these are sufficiently distinct aspects of portfolio rebalancing due to QE,
we include them both as primary indicators. The secondary measures that we
consider supplement our primary return differentials with their growth ana-
logues: the country-specific lagged growth differential (relative to the United
States), and the global composite purchasing managers’ index (PMI). These
proxy spatial rebalancing toward asset classes that are more sensitive to short-
term growth expectations and longer-term expectations of overall global growth,
respectively.
Finally, our primary confidence channel indicator is a global variable, the
VIX index. This measure is designed to capture market sentiment for investing
in risk assets, in particular, although it has been used in other applications as
a measure for broader financial market uncertainty.
The other controls that we include, such as GDP, are standard and de-
tailed definitions and source information are relegated to the appendix. Here
we highlight the inclusion of the Institutional Investor risk rating in our baseline
specification; although not a channel for QE transmission, this country-specific
measure captures the important aspect of the attractiveness of a given devel-
oping country as an investment destination. The data periodicity is quarterly,
spanning 2000Q1 through 2013Q2 (inclusive), and constitutes an unbalanced
panel comprising as many as 60 developing countries. To maximize coverage,
we impute quarterly observations using a cubic spline for a small number of
low-volatility control variables which were only available at an annual frequency.
Additional details, including country coverage, summary statistics, and cross-
correlations, are in the appendix.
4.3 Identifying the potential effect of QE
It is important to raise two important caveats to the discussion above. First,
while the variables we select are meant to proxy for the observable transmission
channels of QE, these measures may well be relevant for capital flows even in the
absence of unconventional monetary policy. For instance, while the flattening
of the yield curve is one of the primary goals of QE, changes in the long-term
cost of capital will also alter the shape of the yield curve, which can in turn
affect financial flows even in periods of unexceptional monetary policy.
Second, these variables may also vary for reasons unrelated to QE. For ex-
ample, exogenous improvements in productivity can alter the growth differential
between economies, even without monetary stimulus in one country versus an-
other. Indeed, the growth differential is a fairly standard feature of most models
of cross-border financial flows, as is the interest rate differential.
Consequently, we do not make claims that point estimates corresponding to
12
these variables necessarily represent the full effects of QE spillovers on financial
inflows. Nor do we seek to pin down the marginal contribution of QE via the
standard channels, which would require us to first determine the precise impact
of QE on these fundamentals. Rather, as discussed in the introduction, our
goal is to establish whether there are any additional, unobserved effects of QE,
after taking into account changes in the observable channels. This allows us to
sidestep the issue of identifying a causal influence of QE on the fundamentals,
as long as we treat any estimated effect from the unobserved component as
representing a lower bound to the potential effects of QE.
That said, it is useful to note that since the global variables represent changes
in financial and economic conditions in high-income (in particular the U.S.)
countries, changes in these variables are plausibly exogenous from the point of
view of our dependent variable (gross inflows). Of course, one could argue that
endogeneity may still arise from unobserved, common factors that affect both
high-income and developing countries equally. We recognize this possibility—
indeed, our empirical strategy embodied in (1) is to account for the effects of as
many distinct common unobservables as possible—but in our robustness checks,
we also consider using only a single global factor, which aims to fully account
for global unobservables in a systematic fashion.
Finally, note that in our VAR specification in (2), we do not include an
explicit QE measure. Our goal in implementing a VAR is not so much to directly
model the effect of QE in a system, but rather to discipline the coevolution of the
fundamentals in a standard normalization scenario. Consequently, we capture
the role of QE during the normalization process explicitly (and solely) through
our estimates of the unobservable QE effect.
5 Baseline results
5.1 Regression results and main findings
As alluded to in the previous section, our baseline estimates of (1) include
two alternative specifications: a parsimonious model that includes only our
primary indicators for each channel, and an extended specification that includes
our secondary indicators. These are reported in Table 1, where the first three
columns (B1 )–(B3 ) correspond to the parsimonious specification for each of the
three alternative measures of confidence, and the next three columns (B4 )–(B6 )
to the extended specification, again for each of the three confidence measures.
The lagged dependent variable is highly significant across all the specifica-
tions, suggesting a certain degree of partial adjustment, which is not unexpected
for quarterly data. While this might suggest the need to incorporate a deeper
lag structure, model selection criteria point to retaining just a single lag.
16
Sum-
mary statistics suggest that much of the variation from the sample is, as to be
16
We considered lag depths of up to 2 years (8 lags). Both the Akaike and Bayesian infor-
mation criteria select the model with only one lag. Results for these additional regressions
are available on request.
13
14
Table 1: Baseline regressions for gross financial inflows, unbalanced quarterly panel,
2000Q1–2013Q2
B1 B2 B3 B4 B5 B6
Lagged inflows 0.469 0.477 0.476 0.466 0.473 0.473
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
All QE 0.027 0.026
episodes (0.01)
∗∗∗
(0.01)
∗∗∗
QE1 episode 0.047 0.049
(0.01)
∗∗∗
(0.01)
∗∗∗
QE2 episode 0.033 0.035
(0.01)
∗∗∗
(0.01)
∗∗∗
QE3 episode 0.008 0.006
(0.01) (0.01)
QE-related 0.003 0.002
expansion (0.00)
∗∗∗
(0.00)
∗∗∗
Liquidity channel
3M T-bill -0.013 -0.020 -0.003 -0.016 -0.017 -0.006
rate (0.00)
∗∗∗
(0.00)
∗∗∗
(0.00) (0.01)
(0.01)
∗∗
(0.01)
Money supply -0.106 0.144 -0.098
(M2) (0.22) (0.26) (0.22)
Portfolio balance channel
Yield curve -0.018 -0.027 -0.007 -0.018 -0.024 -0.007
(0.00)
∗∗∗
(0.01)
∗∗∗
(0.00) (0.01)
∗∗
(0.01)
∗∗∗
(0.01)
Interest rate -0.000 -0.000 -0.000 -0.000 -0.000 -0.000
differential (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Global PMI -0.001 -0.001 -0.002
(0.00) (0.00) (0.00)
Growth 0.001 0.001 0.001
differential (0.00)
(0.00)
(0.00)
Confidence channel
VIX -0.001 -0.002 -0.001 -0.002 -0.002 -0.002
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
Basic controls
GDP 0.139 0.134 0.137 0.129 0.125 0.128
(0.03)
∗∗∗
(0.03)
∗∗∗
(0.03)
∗∗∗
(0.03)
∗∗∗
(0.03)
∗∗∗
(0.03)
∗∗∗
Developing 0.003 0.000 0.002 0.004 -0.000 0.004
GDP growth (0.00)
(0.00) (0.00) (0.00)
∗∗
(0.00) (0.00)
∗∗
High-income -0.001 -0.000 -0.001 -0.000 0.001 0.000
GDP growth (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Country 0.002 0.002 0.002 0.002 0.002 0.002
rating (0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
Crisis period -0.019 -0.028 -0.023 -0.021 -0.026 -0.026
(0.01) (0.01)
∗∗
(0.01)
(0.01) (0.01)
(0.01)
Post-crisis 0.002 -0.011 -0.028 0.002 -0.010 -0.026
period (0.01) (0.01) (0.02)
(0.01) (0.01) (0.02)
Adj. R
2
0.368 0.371 0.366 0.368 0.371 0.367
R
2
(within) 0.372 0.376 0.370 0.374 0.377 0.372
R
2
(between) 0.522 0.525 0.525 0.526 0.529 0.528
N (countries) 1,938 (60) 1,938 (60) 1,938 (60) 1,925 (60) 1,925 (60) 1,925 (60)
All level variables are in logarithmic form, but rates, indices, and indicator variables are untrans-
formed. Bootstrapped standard errors (with 100 replications) are reported in parentheses. A time
trend, country fixed effects, and constant term were included in the regressions, but not reported.
indicates significance at 10 percent level,
∗∗
indicates significance at 5 percent level, and
∗∗∗
indicates significance at 1 percent level.
expected, between economies, although within variation (on which our estimates
depend) is reasonably high (explaining slightly more than a third of variation
in the data).
There are several broad conclusions to be drawn from this set of baseline
results.
First, the QE episode indicators enter with statistically and economically
significant coefficients: the combined QE episode indicator, for instance, sug-
gests that the QE period saw an increase in gross financial inflows to developing
countries of around 3 percent, over and above the effects that QE may have
had on observable channels, such as a reduction in the VIX due to improved
confidence, or the flattening of the yield curve as investors rebalanced their port-
folios. This is nontrivial, and comparable to a one standard-deviation change in
these other traditional channels (such as changes in yield curve spreads or the
short-term interest rate).
It is also notable that when we break down the measure into the three sepa-
rate variables, these measures of confidence display a diminishing effect for each
episode: the magnitude of the coefficient decreases from the first and second
QE interventions, and is actually insignificant for QE3 in the extended specifi-
cation (B4 ). This possibility—that LSAPs were more efficacious in the earlier
QE episodes than the most recent one—is in fact consistent with what the liter-
ature has found for the U.S. economy (C´urdia & Ferrero 2013; Krishnamurthy
& Vissing-Jørgensen 2013).
Second, to the extent that QE affected the fundamentals, there is evidence
that its transmission occurred along all three channels. The rate on 3-month
Treasury bills is generally negative and significant; this is consistent with how
reductions in the liquidity premium due to QE increased yields on short-term
bills, which in turn served as a substitute for developing country asset, hence
reducing financial inflows
17
(the money supply, in contrast, tends to be indis-
tinguishable from zero). The coefficient on the U.S. yield curve also typically
enters with a significant, negative coefficient, consistent with temporal portfolio
rebalancing. The evidence on spatial rebalancing, however, is more mixed; while
the coefficient on the (lagged) growth differential is statistically significant, it
is small in magnitude, and the coefficient on the interest rate differential is in-
distinguishable from zero across all specifications.
18
There is also evidence that
confidence effects are relevant: the coefficient on the VIX is highly significant,
and is in fact the most robust covariate among the different transmission channel
proxies.
Third, these significant fundamental variables all appear to operate at a
global level: the measures tend to be global “push” factors of abundant liq-
17
To the extent that the T-bill rate was falling during the early part of QE, this would
suggest that inflows into developing countries would have increased as short-term Treasuries
became less attractive.
18
The insensitivity of financial flows to interest rate differentials, while disappointing, has
been fairly widely replicated in the literature on gross flows; see, for example, Bruno &
Shin (2013) and Forbes & Warnock (2012a). One reason for this may be the countercyclical
relationship of capital flows to the real interest rate (Contessi, De Pace & Francis 2013), which
would obviate any portfolio rebalancing effect due to changes in interest rate differentials.
15
uidity (falling 3-month bill rates), portfolio rebalancing away from long bonds
(a flattening yield curve), and improved confidence in investing in general (the
episode indicators) and in risky assets in particular (the VIX). Our results in
this regard are consistent with the broader literature on financial flows, which
has found that global “push” factors tend to dominate country-specific “pull”
factors (Baker, Wurgler & Yuan 2012; Fratzscher 2012; Fratzscher et al. 2013).
Finally, we note that a number of the controls, in particular GDP and coun-
try risk, consistently enter with significant coefficients that are in accordance
with theory. For example, the coefficient on GDP is generally positive (an in-
crease in GDP is associated with more financial inflows), although the small size
of the coefficient (significantly lesser than one) suggests a diminishing effect.
19
The crisis dummy is also negative and significant, a result consistent with the
substantial reduction in global capital flows following the crisis (recall Figure 1).
What would be the total effect of QE in this case? Given our estimates
in Table 1, the lower bound of QE effects would likely be, for the average
country, around 3 percent of gross financial inflows. Thus, even if one assumes
no transmission via the observable channels, this unobservable component of
the effect of QE still accounts for an increase in inflows in the order of several
percent. If one is willing to make the additional assumption that changes in
the observable fundamentals are entirely attributable to QE, the effects would
be even greater. For example, including the portfolio balance effect from a 129
basis point decrease in yield curve spreads (one standard deviation in our data)
would yield an increase in gross inflows of between 2 and 3 percent. If changes
along all the three channels were assumed to be fully due to QE, the total effects
could be as large as 8 to 12 percent.
20
5.2 Understanding the effects of unobservables due to QE
It is tempting to assign a specific interpretation to the unobservable effect of
QE. In this subsection, we consider two possible candidate explanations that
may potentially explain the significance of the QE episode variable.
The first explanation we probe is whether the unmeasured effects are im-
plicit measures of expectations. Although difficult to precisely measure, market
expectations are, in principle, recoverable from data on futures and forwards.
We draw on two market-based measures in this regard: the yield implied by the
19
The total effect of this coefficient is (assuming a lagged dependent coefficient of 0.47) equal
to 0.13/ (1 0.47) 0.25. Since the model includes country fixed effects, this amounts to a
within estimate of a concave relationship between inflows and GDP. However, this estimate
likely underestimates the total effect (i.e. when between differentials are taken into account).
For example, the standardized coefficient—which implicitly captures between-country vari-
ation in GDP since it draws on the pooled sample to compute standard deviations—is sig-
nificantly larger (in excess of one), which strongly suggests that the inflow/GDP ratio is not
diminishing when examined at the cross-country level (estimates with standardized coefficients
for the baseline are reported in the appendix).
20
These are computed from the minima and maxima among the significant coefficients across
all specifications in Table 1, assuming a one standard deviation change for all measures with
the exception of the QE episode indicator, which is assumed to hold at unity.
16
3-year futures contract for the 3-month T-bill, and an “implied” yield curve, cal-
culated as the difference between the 3-year implied forward rate for the 10-year
Treasury note and 3-year futures of the 3-month bill.
21
Since the VIX already
embodies an expectations component, we do not introduce any additional con-
trols for expectations via the confidence channel.
The most straightforward way to incorporate expectations is to take the
difference between a future/forward-implied rate and the prevailing rate; for
example, the difference between the current 3-month T-bill futures rate and the
3-month T-bill rate. This captures the manner by which differences between
market expectations of future short rates and contemporaneous short rates can
affect financial flows; put another way, these are anticipated rate changes. We
introduce additional “expectation” measures along these lines for the 3-month
rate only, the yield curve only, and both, in columns (E1 )–(E3 ) of Table 2,
respectively.
Another way to think about expectations is to consider how expectational
errors may come into play. Computation such errors amounts to taking the
difference between current realizations of a yield and the 3-year lagged im-
plied yield from futures/forwards; a positive value of the deviation between
the T-bill rate and lagged 3-year forecasts of the same rate would suggest that
market participants systematically underpredicted yields. We include these “er-
ror” measures—which we can treat as unanticipated rate changes—in columns
(E4 )–(E6 ).
On the basis of these results, we see no basis for attributing the QE episode
effect to unmeasured expectations. Both classes of expectational measures enter
with small, and statistically insignificant, signs. In some ways, this should not be
entirely surprising; market-implied forecasts of interest rates generally perform
rather poorly, especially at longer horizons (Campbell & Shiller 1991; Lange,
Sack & Whitesell 2003). We are thus inclined to discount the possibility that
the QE effect due to unobservables is due to either anticipated or unanticipated
expectations of future interest rates changes.
The second explanation that we explore is whether the QE episode indicator
is indirectly capturing structural shifts in the observable factors, due to the
unprecedented nature of QE. Framed another way, the magnitude of monetary
policy intervention in asset markets may have led to a change in the elasticity
of response of gross inflows to the conventional, observable channels.
We operationalize this hypothesis by interacting our measures of the liquid-
ity, portfolio balance, and confidence channels with the QE episode indicator.
For each channel, we consider both the parsimonious (odd-numbered columns)
and the extended specifications (even-numbered specifications). In the first six
columns (I1 )–(I6 ), we interact the measures separately by channel; in the final
two (I7 )–(I8 ) we consider them all in tandem. These results are reported in
21
We use a 3-year time frame to maintain consistency with our forward-looking exercise in
Section 7, which has a 3-year projection window. Since equivalent price data for futures on the
10-year note are not generally available (and even if they were would likely embed a nontrivial
liquidity premium), we instead rely on computed implied forwards to capture expectations of
yields for the 10-year note.
17
Table 2: Regressions for gross financial inflows with expectational measures, unbal-
anced quarterly panel, 2000Q1–2013Q2
E1 E2 E3 E4 E5 E6
Lagged inflows 0.462 0.462 0.462 0.461 0.463 0.463
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
All QE 0.031 0.031 0.031 0.028 0.029 0.030
episodes (0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
3M T-bill -0.001 0.019
(expectation) (0.02) (0.06)
Yield curve 0.002 0.012
(expectation) (0.01) (0.03)
3M T-bill -0.005 0.003
(error) (0.01) (0.01)
Yield curve 0.013 0.017
(error) (0.01) (0.01)
Channel variables Yes Yes Yes Yes Yes Yes
Basic controls Yes Yes Yes Yes Yes Yes
Adj. R
2
0.367 0.367 0.366 0.367 0.368 0.367
R
2
(within) 0.372 0.372 0.372 0.373 0.373 0.373
R
2
(between) 0.527 0.527 0.527 0.527 0.527 0.527
N (countries) 1,938 (60) 1,938 (60) 1,938 (60) 1,925 (60) 1,925 (60) 1,925 (60)
All level variables are in logarithmic form, but rates, indices, and indicator variables are untransformed.
Bootstrapped standard errors (with 100 replications) are reported in parentheses. A time trend, country
fixed effects, and constant term were included in the regressions, but not reported.
indicates signifi-
cance at 10 percent level,
∗∗
indicates significance at 5 percent level, and
∗∗∗
indicates significance at
1 percent level.
Table 3.
The main message one receives from this set of results is that there is little
evidence that supports the argument that the sensitivity of transmission chan-
nels for unconventional monetary policy changed as a result of QE (with the
exception of the interaction with the money supply). By and large, the coef-
ficients on most of the uninteracted variables in Table 3 remain significant (if
they were before in Table 1), whereas the coefficients on the interaction terms
are generally statistically indistinguishable from zero.
22
The significant, negative coefficient on the interaction of QE and the money
supply deserves some comment. First, we note that it enters with a negative
sign, which is consistent with the relevance of a liquidity channel being operative,
since the rapid expansion of M2 since the financial crisis
23
would have lowered
liquidity premia, in turn raising yields on liquid assets that served as substitutes
22
Note also that the insignificant coefficient on the uninteracted QE episode variable in
most of the specifications need not be a real cause for concern; the total effect the confidence
channel has to be inferred from the sum of both the uninteracted and interaction terms, and
some weighted standard error computed for proper inference.
23
A small but vocal minority of economists have argued that monetary policy following
the crisis was actually contractionary, rather than expansionary (Sumner 2009). Our money
supply data, which (like the rest of our model) is measured in real terms, actually shows an
acceleration of real M2 since mid-2008, although admittedly M2 has lagged its linear trend,
mainly because there was a substantial slowdown in M2 expansion between 2004 and 2008.
18
19
Table 3: Regressions for gross financial inflows with interacted channels, unbalanced quarterly panel,
2000Q1–2013Q2
I1 I2 I3 I4 I5 I6 I7 I8
Lagged inflows 0.474 0.471 0.474 0.474 0.474 0.470 0.471 0.473
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
All QE 0.002 -0.004 -0.005 -0.031 5.086 0.102 -0.003 10.201
episodes (0.01) (0.02) (0.01) (0.02) (2.37)
∗∗
(0.07) (0.02) (5.07)
∗∗
Liquidity channel
3M T-bill -0.009 -0.022 -0.011 -0.018 -0.012 -0.018 -0.011 -0.008
rate (0.00)
∗∗
(0.01)
∗∗∗
(0.00)
∗∗∗
(0.01)
∗∗∗
(0.01) (0.01)
(0.01) (0.01)
3M T-bill 0.065 0.072 0.003 0.059
× QE (0.03)
∗∗
(0.09) (0.04) (0.09)
Money supply 0.268 0.012 -0.006 0.403
(0.27) (0.23) (0.23) (0.29)
Money supply -0.316 -0.630
× QE (0.15)
∗∗
(0.31)
∗∗
Portfolio balance channel
Yield curve -0.014 -0.029 -0.016 -0.026 -0.021 -0.025 -0.015 -0.019
(0.00)
∗∗∗
(0.01)
∗∗∗
(0.00)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗
(0.01)
∗∗
(0.01) (0.01)
Yield curve 0.015 0.017 0.017 -0.011
× QE (0.01) (0.01) (0.01)
(0.02)
Interest rate -0.000 -0.000 -0.000 -0.000 -0.000 -0.000 -0.000 -0.000
differential (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Interest rate -0.001 -0.000 -0.000 -0.000
diff. × QE (0.00) (0.00) (0.00) (0.00)
Global PMI -0.001 -0.000 -0.001 -0.000
(0.00) (0.00) (0.00) (0.00)
Global PMI -0.002 -0.000
× QE (0.00) (0.00)
Growth 0.001 0.001 0.001 0.001
differential (0.00)
(0.00) (0.00)
(0.00)
Growth diff. 0.001 0.001
× QE (0.00) (0.00)
Confidence channel
VIX -0.001 -0.001 -0.001 -0.002 -0.002 -0.001 -0.002 -0.001
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
VIX × QE 0.001 -0.000 0.001 -0.003
(0.00)
∗∗
(0.00) (0.00)
(0.00)
Basic controls Yes Yes Yes Yes Yes Yes Yes Yes
Adj. R
2
0.369 0.369 0.369 0.370 0.371 0.370 0.369 0.371
R
2
(within) 0.374 0.374 0.374 0.375 0.377 0.376 0.375 0.379
R
2
(between) 0.524 0.522 0.524 0.524 0.530 0.532 0.528 0.534
N (countries) 1,938 (60) 1,938 (60) 1,938 (60) 1,938 (60) 1,925 (60) 1,925 (60) 1,925 (60) 1,925 (60)
All level variables are in logarithmic form, but rates, indices, and indicator variables are untransformed. Bootstrapped standard
errors (with 100 replications) are reported in parentheses. A time trend, country fixed effects, constant term, and basic controls
from the extended specification were included in the regressions, but not reported.
indicates significance at 10 percent level,
∗∗
indicates significance at 5 percent level, and
∗∗∗
indicates significance at 1 percent level.
for developing country assets.
That said, we are inclined to somewhat discount this particular result, for a
number of reasons. First, we do not observe a similar significance in the inter-
action effect for the 3-month T-bill rate, which would corroborate the potential
importance of elasticity changes along this channel. Second, this coefficient is
significant in specifications where the coefficients on the 3-month T-bill rate (in-
teracted and uninteracted) fall out of significance, which raises the concern that
the significance of the coefficient could be an artifact of possible multicollinear-
ity, rather than a genuine interaction effect.
24
Finally, although the magnitude
of the effect appears fairly large (and would therefore argue for taking this
effect into account), this is because of scaling differences; the standardized coef-
ficient for M2 (reported in Table A.7) reveals that it exerts an effect of a similar
magnitude the short-term interest rate, which is taken into account for in the
uninteracted model.
As a final point, we do note that the coefficient on the (lagged) growth
differential and yield curve in (I6 ) is marginally significant, as is the interaction
of VIX and QE in (I3 ). However, these enter with the opposite sign. Since
our goal is, in any case, not to identify the effects of QE via the observable
fundamentals, we simply recognize that this result would tend to downward
bias our estimates of the uninteracted effects.
6 Robustness of the baseline
6.1 Additional and alternative controls and estimators
We test the sensitivity of our baseline by several ways. Our first set of tests incre-
mentally introduces additional controls that correspond to: (R1 ) the global level
of saving (to account for the quantity of investable funds); (R2 ) the (lagged)
ratio of trade to output (to account for economic openness); (R3 ) the (lagged)
ratio of private credit to output (to account for variations in the level of financial
development); (R4 ) the (lagged) ratio of debt to GDP (to control for the exist-
ing debt burden); (R5 ) the inflation differential (to allow for possibility excess
inflation may reduce the value of investment in any given economy);
25
(R6 ) the
(lagged) real exchange rate (which allows for exchange rate differentials to affect
inflows).
The left panel of Table 4 presents the results from these tests, using the
extended specification with a single QE episode indicator (specification (B4 ) of
Table 1).
Note that the inclusion of additional variables does not alter the qualitative
message from our baseline results. Moreover, the additional controls do not
generally improve the fit of the model substantially, nor do the coefficients for
24
And as discussed earlier, we favor the T-bill rate since, as a price signal, it offers a
potentially faster-reacting measure of effects via the liquidity channel.
25
Note that since the variables in our baseline are measured in real terms, this only captures
the residual effect that large inflation differentials may exert on inflows, rather than a standard
adjustment for variables expressed in nominal terms.
20
Table 4: Robustness regressions for gross financial inflows, unbalanced quarterly panel, 2000Q1–2013Q2
R1 R2 R3 R4 R5 R6 R7 R8 R9 R10 R11 R12
Additional controls Alternative measures Alternative estimators
Lagged inflows 0.466 0.462 0.446 0.465 0.464 0.344 0.472 0.461 0.479 0.464 0.421 0.421
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.03)
∗∗∗
(0.03)
∗∗∗
(0.03)
∗∗∗
(0.03)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.05)
∗∗∗
(0.06)
∗∗∗
All QE 0.029 0.030 0.028 0.032 0.033 0.044 0.031 0.026 0.038 0.021 0.027 0.027
episodes (0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.02)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗∗
(0.01)
∗∗
(0.01)
∗∗∗
QE tapering -0.063
(0.02)
∗∗∗
Liquidity channel
Short-term -0.017 -0.016 -0.016 -0.018 -0.018 -0.023 -0.007 -0.013 -0.008 -0.016 -0.016
rate (0.01)
∗∗
(0.01)
(0.01)
∗∗
(0.01)
(0.01)
(0.02) (0.01) (0.01)
(0.01) (0.00)
∗∗∗
(0.01)
Money supply -0.009 0.077 -0.056 0.080 0.067 0.090 0.226 0.007 0.078 -0.099 -0.099
(0.24) (0.24) (0.20) (0.29) (0.29) (0.45) (0.22) (0.25) (0.06) (0.10) (0.26)
Portfolio balance channel
Yield curve -0.021 -0.021 -0.018 -0.021 -0.022 -0.027 -0.012 -0.015 -0.013 -0.018 -0.018
(0.01)
∗∗
(0.01)
∗∗
(0.01)
∗∗
(0.01)
(0.01)
(0.02) (0.01) (0.01)
(0.01) (0.01)
∗∗∗
(0.01)
∗∗
Interest rate -0.000 -0.000 -0.000 -0.000 0.000 -0.000 -0.000 -0.001 -0.000 -0.000 -0.000
differential (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Interest rate -0.000
spread (0.00)
Global PMI -0.001 -0.001 -0.001 -0.001 -0.001 -0.001 -0.002 -0.002 -0.002 -0.001 -0.001
(0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Growth 0.001 0.001 0.001 0.002 0.002 0.002 0.001 0.001 0.001 0.001 0.001 0.001
differential (0.00)
(0.00)
(0.00)
∗∗
(0.00)
∗∗
(0.00)
∗∗
(0.00) (0.00)
∗∗
(0.00)
∗∗
(0.00) (0.00) (0.00)
(0.00)
Confidence channel
VIX -0.002 -0.002 -0.002 -0.002 -0.002 -0.002 -0.002 -0.001 -0.002 -0.002 -0.002
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗
(0.00)
∗∗∗
(0.00)
∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗∗
Additional controls
Global saving 0.055 0.064 0.081 0.085 0.116
(0.06) (0.06) (0.08) (0.08) (0.10)
Trade/GDP 0.000 -0.000 -0.000 -0.000 -0.000
(0.00) (0.00) (0.00) (0.00) (0.00)
Credit/GDP 0.000 0.000 0.000 0.002
(0.00) (0.00) (0.00) (0.00)
Debt/GDP 0.011 0.010 -0.056
(0.02) (0.02) (0.07)
Inflation 0.000 0.002
differential (0.00) (0.00)
Real exchange 0.000
rate (0.00)
Global factor 0.009
(0.00)
∗∗∗
Basic controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Estimator LSDV LSDV LSDV LSDV LSDV LSDV LSDV LSDV LSDV LSDV FE SCC-FE
Adj. R
2
0.369 0.374 0.373 0.379 0.379 0.423 0.374 0.347 0.373 0.362 0.368
R
2
(within) 0.374 0.381 0.380 0.388 0.388 0.437 0.380 0.351 0.378 0.365 0.374 0.374
R
2
(between) 0.526 0.529 0.549 0.550 0.551 0.559 0.530 0.523 0.530 0.526 0.526
N (countries) 1,925 (60) 1,838 (57) 1,665 (53) 1,435 (47) 1,435 (47) 947 (27) 1,925 (60) 2,286 (61) 1,926 (60) 1,925 (60) 1,925 (60) 1,925 (60)
All level variables are in logarithmic form, but rates, indices, and indicator variables are untransformed. Bootstrapped standard errors (with 100 replications) are reported in parentheses, with the
exception of the final two specifications, where heteroskedasticity and autocorrelation-robust and Driscoll-Kraay standard errors are reported, respectively. A time trend, country fixed effects, constant
term, and basic controls from the extended specification were included in the regressions, but not reported.
indicates significance at 10 percent level,
∗∗
indicates significance at 5 percent level, and
∗∗∗
indicates significance at 1 percent level.
these controls generally enter with significant coefficients.
26
They do, however,
erode the size of the sample (dramatically so in the case where the real exchange
rate is included).
The second suite of tests allows for alternative measures for a number of our
variables of interest, which comprise the middle panel of Table 4. As in the
first set of checks, we apply the alternative measures to our original extended
specification (B4 ).
The first alternative, reported in column (R7 ), considers a measure of the
third QE episode that includes not just a single indicator that corresponds to
the three quarters from 2012Q4 through 2013Q2, but also includes an additional
indicator for the period where there were anticipations of a tapering of QE (due
to interpretations of forward guidance issued by the Federal Reserve).
27
Inter-
estingly, expectations of tapering were associated with a significant reduction
in inflows. This reduction was not just statistically but economically impor-
tant: indeed, the coefficient on the variable is almost twice as large as average
confidence effects over all prior QE episodes.
One reservation with regard to the interest rate differential we rely on in our
baseline is that our reliance on interest rate differentials may not reflect true
fluctuations in cross-country costs of capital, since the country-specific nominal
interest rate we rely on to compute our interest rate differential is only one
possible measure of the average cost of capital. To establish whether such mea-
surement issues may explain the insignificance of the interest rate differential in
the baseline, column (R8 ) substitutes the baseline interest rate differential with
the interest rate spread computed from a richer array of fixed income instru-
ments. The coefficient on this measure is still negligibly small and statistically
insignificant, which suggests that mismeasurement is not at the heart of the
insignificant coefficient for the interest rate differential in our baseline.
A third alternative set of measures that we explore allows for the fact that
unconventional monetary policies were not the sole domain of the Federal Re-
serve, but were more or less simultaneously pursued by the Bank of England (via
the Asset Purchase Facility), the Bank of Japan (via its Asset Purchase Pro-
gram), and the European Central Bank (through its Securities Market Program
and Outright Monetary Transactions
28
). Consequently, we substitute all our
U.S.-centric controls with their weighted-average equivalents from these coun-
tries (which we collectively refer to as the G4).
29
As is clear from the results in
26
The domestic credit/GDP ratio does enter with a marginally significant coefficient in the
final specification, but the sample size is substantially smaller. For this reason, we play down
these results, but note that the sign of the coefficient does comport with a priori expectations
(higher levels of financial development are associated with larger inflows).
27
Given our quarterly frequency, this effectively amounts to including an additional fixed
effect for the period 2013Q2.
28
There is some dispute as to whether the ECB’s Long-Term Refinancing Operations consti-
tute a form of quantitative easing; we stay with the convention here and exclude this program
as a form of QE. Note as well that while the SMP has resulted in a substantial expansion of the
ECB balance sheet, the OMT has in fact never been used, despite widespread acknowledgment
that the program engendered confidence effects.
29
For the episode indicator, we drew on qualitative information in Neely (2013) concerning
22
column (R9 ), our main qualitative conclusions are unaffected by this change.
A final alternative measure we consider collapses all global variables into a
single global factor, and substitutes this for all the global variables in (1).
30
This
approach has some precedence in the international macroeconomics literature
(Albuquerque, Loayza & Serv´en 2005; ose, Otrok & Whiteman 2003). The
principle is that global variables are driven by some underlying, unobservable
common factor, and that controlling only for observables may still omit some
time-varying global component. The tradeoff—and the reason why we choose
not to employ this method for our baseline—is that, given our interest in trans-
mission channels, it is difficult to establish the precise contribution of each global
variable using a single global factor. Moreover, the Kaiser-Meyer-Olkin test of
sampling adequacy indicate that the underlying variables are sufficiently distinct
that partial correlations between them are low, and hence are not particularly
well-suited for factor analysis. Keeping these caveats in mind, introducing a
global measure nevertheless serves as a useful counter-check for the importance
of the sum total of global effects.
This is reported in column (R10 ). The global factor is statistically significant
and fairly large—a one standard deviation increase in the measure is associated
with an increase in inflows of approximately 0.01 percent—but relying on the
global factor alone is likely to underestimate the total effect of QE on inflows.
To see this, note that a one standard deviation increase in each of the three
significant channels in the simplest parsimonious specification (B1 ) in Table 1
yields an increase of inflows of 0.07 percent, an estimate seven times larger.
The rightmost panel of Table 4 offers two alternative estimation methods
for (1). Column (R11 ) provides estimates using na¨ıve OLS with fixed effects
and heteroskedasticity- and autocorrelation-robust clustered errors, to deter-
mine the importance of our correction for Nickell (1981) bias. As is clear, our
quantitative results are substantially unchanged even when the bias remains
uncorrected; most coefficient estimates obtained using OLS differ only from the
third decimal place onward.
31
Finally, one may be concerned about the pos-
sibility that cross-sectional dependency may contaminate our estimates; this is
especially pertinent given how the response of portfolio capital to QE may be
subject to herding behavior. Column (R12 ) corrects for this possibility by es-
timating spatial correlation-consistent standard errors using the methodology
suggested by Driscoll & Kraay (1998). As before, our qualitative findings are
substantially unaffected.
G4 central bank unconventional monetary policy actions, and coded additional quarters as
QE periods if at least two additional central banks engaged in QE.
30
We construct this factor from the varimax orthogonal rotation of the first principal com-
ponent of the vector of global variables. We also considered an alternative, the proportion-
weighted sum of the first three principal components (all possessed eigenvalues greater than
unity). Both methods produced qualitatively similar results, which are available on request.
31
Estimates for all specifications in the baseline are provided in the appendix.
23
6.2 Decomposition of aggregate flows
Not all flows are created equal, and different forms of financial flows can be
expected to respond differently to the effects of QE. The theoretical literature
has long recognized that the determinants of portfolio flows are fundamentally
distinct from those of FDI (Kraay, Loayza, Serv´en & Ventura 2005; Smith &
Valderrama 2009), and empirical work has corroborated the importance of ac-
counting for global drivers for the former (Fratzscher 2012) and country-specific
factors for the latter (Alfaro, Kalemli-Ozcan & Volosovych 2008; enassy-Qu´er´e,
Coupet & Mayer 2007; Busse & Hefeker 2007; Dailami, Kurlat & Lim 2012).
Here we break down our dependent variable—aggregate gross inflows—to
obtain greater insight into whether specific channels may be more operative then
others, depending on the financial flow. Inflows are decomposed into portfolio,
loans, and FDI. By relying on an alternative measure—gross fund inflows—we
are further able to separate portfolio flows into equity and bond purchases.
Our estimates are reported in Table 5, again relying on the extended spec-
ification with a single QE episode indicator—specification (B4 )—for each con-
stituent flow: (D1 ) portfolio; (D2 ) loans; and (D3 ) FDI. In column (D4 ), we
first report—for comparison purposes—total gross fund inflows, before this is
separated into portfolio bond and equity flows, in columns (D5 ) and (D6 ),
respectively.
32
We draw several conclusions from this exercise.
First, and most remarkable, is the distinction between FDI and both port-
folio and loan flows. For FDI inflows, exempting the lagged dependent variable,
only the institutional investor variable and GDP enter with a (marginally) sig-
nificant coefficient.
33
This result underscores the importance of political and
institutional risk as determinants of FDI, which has ample support in the lit-
erature (Alfaro et al. 2008; enassy-Qu´er´e et al. 2007; Busse & Hefeker 2007;
Dailami et al. 2012). This result also corroborates with evidence from gravity-
type models of FDI (which finds larger FDI flows between bilateral pairs with
larger pairwise GDP), and the more general stylized fact that gross FDI inflows
tend to be countercyclical and the least volatile among different financial flows
(Contessi et al. 2013). It is nevertheless useful for us to recognize that the insen-
sitivity of FDI inflows to global variables applies even when considering gross
(rather than net) flows, and that QE has had little impact on this large and
stable source of developing country cross-border finance. In contrast, both port-
folio capital and bank lending respond more to both global drivers. A related
observation is that portfolio and loan flows react mainly to the global “push”
factors (due to economic conditions in high-income countries), as opposed to
32
Since the decomposed series are generally less persistent than the aggregate flows data
(with comparable average time periods), the magnitude of bias between the different approx-
imations is virtually identical (Bruno 2005). Accordingly, we relax the bias correction to
just O
1
T
for the estimates in Table 5. Since the panels in the decomposed series also in-
clude samller samples and can be more unbalanced, we also increase the number of bootstrap
replications to 200.
33
Both variables consistently enter with a positive and significant across our baseline and
robustness specifications.
24
25
Table 5: Regressions for financial inflows, by type, unbalanced quarterly panel,
2000Q1–2013Q2
D1 D2 D3 D4 D5 D6
Portfolio Loans FDI Gross fund Bonds Equity
Lagged inflows 0.261 0.307 0.597 -0.088 0.294 -0.011
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.04)
∗∗
(0.03)
∗∗∗
(0.03)
All QE 0.018 0.021 -0.003 0.061 0.015 0.044
episodes (0.01)
∗∗∗
(0.01)
∗∗∗
(0.01) (0.02)
∗∗∗
(0.02) (0.03)
Liquidity channel
3M T-bill -0.015 -0.008 0.004 -0.080 -0.089 -0.053
rate (0.01)
∗∗
(0.01)
(0.01) (0.02)
∗∗∗
(0.02)
∗∗∗
(0.03)
∗∗
Money supply 0.015 -0.071 0.056 -1.110 -2.120 -0.589
(0.19) (0.16) (0.26) (0.65)
(0.45)
∗∗∗
(0.66)
Portfolio balance channel
Yield curve -0.020 -0.002 0.005 -0.090 -0.065 -0.064
(0.01)
∗∗∗
(0.01) (0.01) (0.03)
∗∗∗
(0.02)
∗∗∗
(0.03)
∗∗
Interest rate -0.000 -0.000 -0.000 -0.001 -0.002 -0.000
differential (0.00) (0.00) (0.00) (0.00) (0.00)
(0.00)
Global PMI -0.001 -0.001 -0.001 0.008 0.003 0.004
(0.00) (0.00) (0.00) (0.01) (0.00) (0.01)
Growth 0.001 0.001 0.000 0.001 -0.000 -0.001
differential (0.00)
(0.00) (0.00) (0.00) (0.00) (0.00)
Confidence channel
VIX -0.002 -0.000 -0.000 -0.002 -0.006 -0.000
(0.00)
∗∗∗
(0.00) (0.00) (0.00) (0.00)
∗∗∗
(0.00)
Basic controls
GDP 0.009 0.110 0.070 -0.060 0.020 0.039
(0.03) (0.02)
∗∗∗
(0.04)
(0.09) (0.07) (0.08)
Developing 0.004 0.000 -0.001 0.014 0.023 0.007
GDP growth (0.00)
∗∗∗
(0.00) (0.00) (0.01)
∗∗∗
(0.00)
∗∗∗
(0.01)
High-income -0.001 0.002 0.004 -0.011 -0.017 -0.007
GDP growth (0.00) (0.00) (0.00) (0.01) (0.01)
∗∗∗
(0.01)
Country 0.001 0.001 0.002 0.002 0.001 0.000
rating (0.00)
∗∗∗
(0.00)
∗∗∗
(0.00)
∗∗
(0.00) (0.00) (0.00)
Crisis period -0.002 -0.043 -0.005 0.024 -0.043 0.032
(0.01) (0.01)
∗∗∗
(0.02) (0.04) (0.03) (0.05)
Post-crisis 0.024 -0.025 -0.010 0.038 -0.061 0.050
period (0.01)
(0.01)
∗∗
(0.02) (0.05) (0.04) (0.05)
Adj. R
2
0.157 0.032 0.399 0.054 0.193 0.005
R
2
(within) 0.164 0.037 0.403 0.070 0.203 0.018
R
2
(between) 0.572 0.209 0.854 0.450 0.562 0.042
N (countries) 1,925 (60) 3,460 (85) 2,419 (63) 974 (31) 1,220 (39) 1,185 (37)
All level variables are in logarithmic form, but rates, indices, and indicator variables are untrans-
formed. Bootstrapped standard errors (with 100 replications) are reported in parentheses. A time
trend, country fixed effects, and constant term were included in the regressions, but not reported.
indicates significance at 10 percent level,
∗∗
indicates significance at 5 percent level, and
∗∗∗
indicates
significance at 1 percent level.
country-specific “pull” variables.
Second, comparing portfolio with loan flows, the latter clearly responds more
to the unobservable effect of QE (the coefficient on the indicator variable is 0.021
versus 0.018), suggesting that more so than for the other flows, QE operated
through channels other than the modeled channels to boost bank lending. In
contrast, measurable transmission channels for QE are routinely larger for port-
folio flows. For example, the coefficient on the yield curve (corresponding to
the portfolio rebalancing channel) is an order of magnitude as large as that for
loans; an analogous argument can be made for the short-term interest rate (for
the liquidity channel). The number of statistically significant coefficients is also
larger for portfolio flows.
34
Taken together, these first two findings strongly suggest that it is portfolio
flows, and especially bond capital, that are most sensitive to QE. In contrast,
FDI—which is the most stable component of cross-border financial flows—tends
to respond to structural, long-term determinants, such as the institutional rating
of the economy. This is consistent with our understanding that portfolio flows
react most to the various effects of conventional monetary policy—after all,
monetary policy is generally effective only in the short run, and portfolio flows
are by definition the most easily reassigned—and this bias toward shorter-term
flows evidently carries over to unconventional monetary policy as well.
Turning to the portfolio flow decompositions, we first note that the statisti-
cally significant coefficients in columns (D1 ) and (D4 ) are broadly comparable,
which lends credibility to our use of the fund inflows data. In terms of the
decomposition, it is notable that while bond flows appear to react to more
transmission channels than equity flows—debt is associated with changes in the
VIX as well as the global PMI,
35
while equity is not—the magnitude (and stan-
dard errors) of the coefficients on equity are generally larger than those for debt.
Alternatively, although bond flows are liable to react to a wider range of possi-
ble QE transmission channels, equities react more strongly to the few channels
to which they do react to.
6.3 Accounting for commonality in gross inflows
Given the strong representation of global factors our baseline specification, it is
natural to question whether how pervasive this source of commonality is in our
data. In this subsection, we probe the relative importance of our global factors
34
The relative insensitivity of bank lending to observable fundmentals in specification (D2 )
may strike some as contrary to findings, especially from the crisis literature, that loans appear
to respond to key global factors during crisis events (Adams-Kane, Lim & Jia 2012; Broner
et al. 2013). We do not see an inconsistency here, since loan flows may well respond strongly
to fundamentals in a crisis environment (which we control for), but not under non-crisis
conditions. Moreover, it is important to recognize that our results pertain to post-2000 data
only, which is our period of interest. Consequently, studies that rely on a longer span of bank
lending data may well uncover somewhat ifferent relationships.
35
This coefficient is negative, which indicates that inflows into debt decrease when global
growth prospects improve. This outcome is consistent with substitution into riskier assets
when growth outlooks turn upward.
26
using a principal components approach inspired by Longstaff, Pan, Pedersen &
Singleton (2011).
36
We construct two alternative global factors from our gross inflows data: a
factor constructed from the varimax orthogonal rotation of the first principal
component of the vector of gross inflows (which we term PC1 ), and a proportion-
weighted sum of the first three principal components (which we term PC3 ).
37
We then use the global factor—which by construction captures the common ele-
ments among cross-country gross inflows—as a dependent variable in regressions
where we include our global variables as regressors.
Table 6reports our results from this exercise. The first two columns report
(OLS) regressions for, respectively, the first and weighted principal components,
using only global variables as covariates. Since our goal is to ascertain the rela-
tive importance of these global variables in gross inflows, this pair of specifica-
tions represent our main results of interest.
38
We find that all global variables included in our extended baseline specifica-
tion contribute substantially to the common variation across cross-country gross
inflows. The adjusted R
2
is extremely high—0.92 and 0.91—and the point es-
timates of coefficients are both statistically significant, economically large, and
carry signs consistent with our earlier findings. As before, there is evidence that
these global variables operate along all three channels we consider in this paper.
Overall, these results echo the findings in the literature that global macroeco-
nomic variables dominate movements in international financial markets, as veri-
fied for international equity market returns (Baker et al. 2012), portfolio capital
flows (Fratzscher 2012), and sovereign credit default swap spreads (Longstaff
et al. 2011).
For robustness, we also consider, using the PC3 factor,
39
the incremental in-
clusion of several country-specific controls: the cross-country averaged interest
rate differential (C3 ), the average growth differential (C4 ), both differentials
(C5 ), and with both differentials as well as additional basic controls from the
extended specification (C6 ).
40
We regard these results as mainly of supple-
mentary value—in the sense that they help us better understand the relative
importance of the global factors—since it is difficult to interpret the contribu-
36
In contrast to Longstaff et al. (2011), we extract principal components directly from actual
gross inflows, rather than its correlation matrix. We then use this extracted factor directly as
our dependent variable in the analyses that follow (as an aggregate time series), rather than
perform country-specific or panel regressions.
37
The first three components had eigenvalues in the excess of 3, and cumulatively explain
slightly more than half of the cross-country variation. Including the next 7 components (all
components that have eigenvalues greater than unity) increases the explained variation to
87 percent, but at the cost of the weighted-sum component being comprised of a very large
number of (linearly uncorrelated) subcomponents, none of which singularly contribute much
to the overall variance in the data.
38
In contrast to the baseline, these specifications omit a lagged dependent variable, which
enter with an insignificant coefficient.
39
Results with the PC1 factor were qualitatively similar, although standard errors were
slightly larger. These are available on request.
40
These are total developing country GDP, the growth rate for the developing world, and
average institutional risk rating.
27
tion of an “average” interest rate or growth differential to the common global
factor in gross inflows.
With this caveat in mind, we make a two additional points with regard to
these latter specifications. First, and most important, there is little additional
gain from including country-specific explanatory variables: the improvement in
the adjusted R
2
from including these variables is miniscule, which is also verified
in the final row of Table 6, where we report the gain in the fit from including
these additional country-specific regressors.
41
Second, most of the global vari-
ables from the different channel retain their significance, which corroborates
their inclusion in our baseline analysis.
7 The effects of tapering and monetary policy
renormalization on developing economies
7.1 Normalization scenarios for unconventional monetary
policy
We operationalize the estimation of (2) by populating the vector TC
it
=
[L
it
PB
it
C
it
X
it
GFI
it
] with the (weighted average) G4 short-term interest
rate (L
it
), the G4 yield curve (PB
it
), the VIX (C
it
), real GDP growth in the
G4 and developing world (X
it
), and aggregate gross capital inflows in developing
economies (GFI
it
) (so the final VAR specification is six-dimensional).
Standard model selection metrics suggest between one and four lags; we
adopt two (k = 2) as a middle-ground compromise.
42
Formal Johansen tests
reject the presence of cointegrating relationships in the system, which supports
the estimation of the model as an unrestricted VAR. Our Cholesky ordering
assumes the following: G4 GDP growth, developing GDP growth, developing
capital inflows, the VIX, G4 short-term interest rates, and the yield curve (which
potentially responds to all other variables in real time).
To provide some sense of the overall performance of the VAR, Figure 2
presents the impulse response functions for gross inflows (as a share of GDP)
for a positive one standard deviation shock in the transmission channels and
endogenous controls in (2). The predicted deviations appear to be consistent
with what one would intuitively expect; for example, a shock to the yield curve
or VIX is associated with declines in inflows, while GDP growth is associated
with increases. In general, the shocks appear to die out after about 10 quarters.
Our benchmark is a no-change scenario that maintains the status quo. We
use the VAR to generate a scenario where unconventional monetary policy nor-
malizes over the course of the following three years. The paths are detailed
41
This is computed from taking one minus the ratio of the adjusted R
2
of specification in
question and specification (C2 ).
42
Specifically, Hannan and Quinn Information Criterion and the Bayesian Information Cri-
terion suggest one lag, Final Prediction Error and Likelihood Ratio statistics recommend two,
and the Akaike Information Criterion points to four. In the two-lag model, all significant
eigenvalues are less than one.
28
Table 6: Regressions for principal components of gross financial inflows, bal-
anced quarterly panel, 2000Q1–2013Q2
C1 C2 C3 C4 C5 C6
1PC 3PC 3PC 3PC 3PC 3PC
All QE 0.902 1.534 0.594 1.012 0.588 0.862
episodes (0.34)
∗∗
(0.71)
∗∗
(0.24)
∗∗
(0.40)
∗∗
(0.29)
(0.39)
Liquidity channel
3M T-bill -1.727 -2.329 -1.289 -1.838 -1.283 -0.702
rate (0.52)
∗∗∗
(1.16)
(0.56)
∗∗
(0.52)
∗∗∗
(0.61)
(0.61)
Money supply -16.740 -29.119 -5.165 -19.694 -5.014 15.830
(6.29)
∗∗
(13.29)
∗∗
(8.47) (7.65)
∗∗
(10.57) (14.47)
Portfolio balance channel
Yield curve -1.869 -2.680 -1.174 -2.017 -1.166 -0.605
(0.46)
∗∗∗
(1.17)
∗∗
(0.50)
∗∗
(0.48)
∗∗∗
(0.56)
(0.76)
Interest rate 0.125 0.126 0.197
differential (0.04)
∗∗
(0.05)
∗∗
(0.09)
Global PMI 0.257 0.395 0.188 0.280 0.187 0.146
(0.07)
∗∗∗
(0.16)
∗∗
(0.06)
∗∗∗
(0.08)
∗∗∗
(0.08)
∗∗
(0.10)
Growth 0.086 -0.003 0.086
differential (0.13) (0.11) (0.26)
Confidence channel
VIX -0.085 -0.161 -0.082 -0.086 -0.082 -0.062
(0.02)
∗∗∗
(0.04)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.02)
∗∗∗
(0.03)
∗∗
Crisis dummies Yes Yes Yes Yes Yes Yes
Basic controls No No No No No Yes
Adj. R
2
0.924 0.906 0.945 0.921 0.940 0.936
Gain in fit (%) 4.3 1.7 3.8 3.3
N 24 24 24 24 24 24
All level variables are in logarithmic form, but rates, indices, and indicator variables are un-
transformed. The dependent variable is either the varimax orthogonal rotation of the first
principal component (1PC) or proportion-weighted sum of the first three principal compo-
nents (3PC). Heteroskedasticity and autocorrelation-robust standard errors are reported in
parentheses. A time trend, constant term, and crisis-related dummies (crisis period and post-
crisis) were included in the regressions, but not reported.
indicates significance at 10 percent
level,
∗∗
indicates significance at 5 percent level, and
∗∗∗
indicates significance at 1 percent
level.
in the upper panel of Table 7. Our policy normalization scenario seeks to
match growth paths for high-income and developing economies consistent with
the World Bank’s Global Economic Prospects forecast for the global economy
(World Bank 2014). This involves three assumptions for the 2014–16 period: a
gradual start to normalization of long-term interest rates, with rates rising by
66 basis points (bps) over the course of 2014, and continuing to increase gradu-
ally over the next two years, so that cumulatively there is a 166 bps by the end
of 2016. Short-term rates are also assumed to begin normalization during this
period, rising by a cumulative 170 bps over the three years. Additional unob-
served effects of QE—as captured by the QE episode indicator—are assumed to
gradually decline and taper off fully by 2016.
29