WPS8354 Policy Research Working Paper 8354 Gross Capital Flows, Common Factors, and the Global Financial Cycle Luis-Diego Barrot Luis Serven Development Research Group Macroeconomics and Growth Team February 2018 Policy Research Working Paper 8354 Abstract This paper assesses the international comovement of gross —although global factors play a bigger role for outflows than capital inflows and outflows using a two-level factor model. for inflows. The commonality of flows reflects a global cycle, Among advanced and emerging countries, capital flows summarized by a small set of variables (the VIX, the U.S. real exhibit strong commonality: common (global and country interest rate and real exchange rate, U.S. GDP growth, and group-specific) factors account, on average, for close to half world commodity prices) that explain much of the variance of their variance. There is a contrast across groups: common of the estimated factors—especially the global factors. Over factors dominate advanced-country capital flows, while time, the quantitative role of the common factors exhib- idiosyncratic factors dominate emerging-country flows its a “globalization” stage up to 2007, during which they and, especially, developing-country flows. The reason is the acquire growing importance, followed by a phase of “deglo- much larger role of global factors among advanced coun- balization” post-crisis. Greater financial openness, deeper tries. Importantly, these findings apply to both inflows and financial systems, and more rigid exchange rate regimes outflows: their respective common factors are very similar amplify countries’ exposure to the global financial cycle. This paper is a product of the Macroeconomics and Growth Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at lserven@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Gross Capital Flows, Common Factors, and the Global Financial Cycle Luis-Diego Barrot and Luis Servény JEL: F32, F36, G15 Keywords: Capital Flows, Comovement, Common Factors, Global Financial Cycle We thank Eugenio Cerutti, In Choi, Sebnem Kalemli-Ozcan, Andy Rose and Sergio Schmukler for useful comments. We are also grateful to Jorg Breitung and Sandra Eickmeier, and In Choi and Yun Jung Kim, for kindly sharing their computer code. Any remaining errors are ours only. The views expressed here do not necessarily re‡ect those of the World Bank, its Executive Directors, or the countries they represent. y The World Bank. email addresses: lbarrotaraya@worldbank.org, lserven@worldbank.org 1 Introduction There is wide consensus that capital ‡ ows can have major consequences for macroeconomic and …nancial stability in both source and destination countries. Understanding better their determinants is therefore a …rst-order priority for both academics and policy makers. Fol- lowing the global crisis, increasing attention has focused on the global factors that may drive in‡ ows and out‡ ows across the world, and on the implications of the international comove- ment of capital ‡ ows for the ability of national policy makers to shelter their economies from global …nancial shocks. These questions had been …rst addressed in the context of capital ‡ ows to emerging markets by an empirical literature that drew a distinction between what have come to be termed ’ push’factors, a¤ecting capital ‡ ows across many countries, and ’ pull’factors driving country-speci…c ‡ ows (e.g., Calvo, Leiderman and Reinhart 1996). The distinction has been revived in a more recent literature stressing the high degree of synchronization of …nancial conditions across the world. Along these lines, Forbes and Warnock (2012) …nd that ex- treme capital ‡ ows episodes are driven by a small set of global variables – notably global risk aversion, as captured by the VIX index, along with global GDP growth and global inter- est rates. In turn, in‡uential contributions by Rey (2013) and Miranda-Agrippino and Rey (2015) conclude that one latent global factor accounts for much of the variance of capital ‡ ows and risky asset returns around the world. They interpret this result as evidence of a global …nancial cycle, driven by investors’time-varying risk aversion –although they do not o¤er an explicit assessment of its quantitative importance. A closely-related view (Bruno and Shin 2015b) stresses the role for capital ‡ global liquidity’ ows of ’ , which involves the transmission of credit conditions in global …nancial centers to the rest of the world through cross-border ‡ ows re‡ ecting changes in global banks’leverage, itself largely driven by chang- ing risk perceptions summarized by the VIX. More recently, Cerutti, Claessens and Rose (2017) …nd that panel and individual-country regressions of di¤erent types of capital ‡ ows on selected global variables –including latent global factors speci…c to the ‡ ow type –rarely account for more than a quarter of the variation in the dependent variable. The literature has also examined the trends over time in the role of common factors driving international …nancial conditions. Over the long term, the degree of international comovement of asset prices (Jordá et al 2017) as well as capital in‡ ows (Reinhart, Reinhart and Trebesch 2017) appears to have been on the rise. Particular attention has been paid to the question whether the role of global factors changed with the global …nancial crisis and the ensuing ’ deglobalization’ , as termed by some observers. Fratzscher (2012) argues that common factors played the leading role in the ‡ uctuations of net capital ‡ ows during the 1 2007– 08 …nancial crisis, but domestic factors became dominant in the post-crisis. In turn, Avdjiev et al (2017a) …nd signi…cant changes in the sensitivity of international bank and bond ‡ ows to particular global variables: after the crisis, the impact of US monetary policy changes on both types of ‡ ows increased, while the responsiveness of cross-border loan ‡ ows to global risk conditions declined. Several papers address the question of what determines the extent to which a coun- try’s capital ‡ ows re‡ ect world …nancial conditions – and thereby the country’ s exposure to global …nancial shocks. Financial openness has been singled out as a likely contributing factor, as has …nancial depth. For example, Bruno and Shin (2015b) …nd that global fac- tors have a larger impact on cross-border bank ‡ ows in more …nancially open countries with big- ger banking ‡ ows. The composition of investors and/or ‡ ows may also matter, because some may be more sensitive than others to common factors. Thus, Raddatz and Schmukler (2012) stress the role of mutual funds in propagating shocks across countries. Cerutti, Claessens and Puy (2017) argue that the foreign investor base –as re‡ ected by the relative importance of global banks and international mutual funds –matters for the sensitivity of emerging-market portfolio and bank in‡ ows to global variables. The literature has paid particular attention to the role of the exchange rate regime. Rey (2013) and Miranda-Agrippino and Rey (2015) argue that the worldwide reach of the global …nancial cycle renders the Mundellian Trilemma – governing the choice among exchange rate regime, monetary independence, and capital account openness –just a dilemma, so that with an open capital account the exchange rate regime ceases to matter for the international transmission of global …nancial conditions. Cerutti, Claessens and Rose (2017) …nd that the (limited) explanatory power of global factors and push variables in capital ‡ ow regressions is not sensitive to the exchange rate regime, while Cerutti, Claessens and Puy (2017) conclude that more ‡ exible regimes augment, rather than reducing, the e¤ects of global conditions on emerging-market capital in‡ ows. In contrast, Obstfeld, Ostry and Qureshi (2017) conclude that, among emerging markets, the response of gross capital in‡ ows (speci…cally, portfolio and other investment in‡ ows) to global risk, as measured by the VXO index, is signi…cantly greater under …xed exchange rate regimes than under more ‡ exible regimes.1 In this paper we investigate the contribution of common shocks to the observed patterns of gross capital ‡ ows around the world, using a large cross-country annual data set spanning nearly four decades. In contrast with most of the recent literature, which focuses on ‡ ows disaggregated by type,2 our focus is on aggregate gross ‡ ows. This allows us to assess the 1 Bekaert and Mehl (2017) also …nd that countries with more rigid exchange rate regimes tend to exhibit signi…cantly higher degrees of interest rate pass-through than countries with more ‡exible regimes. 2 Recent examples include Byrne and Fiess (2016); Eichengreen, Gupta and Masetti (2017); Cerutti, Claessens and Puy (2017); and Cerutti, Claessens and Rose (2017). 2 overall sensitivity of countries’cross-border ‡ ows to common shocks, and thus the quantita- tive relevance of the global …nancial cycle for total capital ‡ows –a …rst-order question from the macroeconomic perspective. Also, while most previous literature has focused on gross in‡ows, we consider both aggregate in‡ ows and out‡ ows. This allows us to examine their commonality, as well as the di¤erences in the extent to which they re‡ ect the global cycle. The analysis is conducted in the framework of a two-level latent factor model that com- bines global factors a¤ecting all countries with factors a¤ecting speci…c groups of countries – advanced, emerging and developing. The two-level factor model setting permits disen- tangling the common shocks with global reach from those a¤ecting only particular country groups. The variance decompositions obtained from the factor model provide a direct as- sessment of the quantitative importance of the global cycle for capital ‡ ows across the world as well as across particular country groups. The model is estimated using a recently-developed extension of the standard principal- component approach to the multi-level setting. This approach avoids unnecessary parameter restrictions often imposed in earlier literature, and o¤ers the key advantage of its computa- tional simplicity. Recursive estimation of the factor model over moving time samples allows us to assess changes over time in the overall contribution of common shocks – rather than just that of particular variables, as considered by the existing literature. To determine the observable counterparts of the latent factors, we analyze the relation between the estimated factors and global variables stressed in the ’ push vs pull’literature. Likewise, we explore the determinants of countries’ exposure to the global cycle by relat- ing the estimated factor loadings to variables describing countries’ structural and policy framework, including in particular the exchange rate regime. We …nd that capital ‡ ows exhibit a considerable degree of commonality among advanced and emerging countries. For both in‡ ows and out‡ ows, information criteria indicate the presence of one global factor plus one group factor for each of the two groups.3 On average, the estimated common factors account for just under half of the observed variation of gross capital ‡ ows of advanced and emerging countries. However, there are clear di¤erences across 3 Cerutti, Claessens and Puy (2017) also employ a two-level latent factor model to examine disaggregated ‡ ows to advanced and emerging countries. Like us, they …nd a common factor driving gross in‡ ows to emerging countries but, in contrast with our results, they do not …nd a common factor behind in‡ ows to advanced economies, nor a global factor a¤ecting in‡ ows to both groups of countries. One possible reason for the discrepancy ia that we use a di¤erent dataset (aggregate annual data rather than their quarterly dis- aggregated data) and estimation method. Still, in our sample in‡ ows to advanced countries display stronger commonality (as re‡ ected by cross-country correlations) than those to emerging or developing countries. Using quarterly data, Advjiev et al (2017b) likewise …nd that aggregate debt in‡ ows to advanced countries react more strongly to global shocks than do in‡ ows to emerging countries. Further, as discussed below, measures of cross-sectional dependence computed on our data strongly indicate the presence of common factors in advanced-country in‡ ows (as well as out‡ows). 3 countries, as well as country groups. Across countries, common factors are virtually irrel- evant for, e.g., New Zealand or Pakistan, but account for the bulk of the variance in, e.g., Norway or India. Across groups, common factors dominate advanced-country capital ‡ ows –on average they contribute close to two-thirds of the variance –while local (idiosyncratic) factors dominate emerging-country capital ‡ ows and, especially, developing-country ‡ ows. The di¤erence is primarily due to the action of the global factors: their average variance contribution is over twice as large among advanced countries as among the rest.4 Importantly, these results apply to both gross in‡ ows and gross out‡ ows. In most cases, their respective latent factors show a large positive correlation. This is particularly the case for the global factors, and also for the advanced-country group factors. Developing countries represent the exception to this rule. However, across all country groups we also …nd that out‡ ows re‡ ect more strongly than in‡ ows the action of global factors, while group factors a¤ect gross in‡ ows more strongly than gross out‡ ows. These results suggest that a signi…cant part of the variation of gross capital ‡ows over the cycle may re‡ect out‡ ows from a relatively large number of countries into a relatively smaller number of countries – e.g., safe havens during global busts, and high-return risky destinations during global booms. The estimated common factors (global and group-speci…c, for in‡ ows as well as out‡ ows) are robustly negatively correlated with the VIX and similar risk proxies. Much of their variation (70 to 80 percent in the case of the global factors, and around 50 percent in the case of the group factors) can be explained by a small set of global variables –the VIX, U.S. interest rates, the U.S. real exchange rate, U.S. GDP growth, and world commodity prices. Thus, through the common factors, these fundamental variables indirectly account for close to half the variance of advanced-country in‡ ows and out‡ ows, and around one-quarter of the variance of emerging-country in‡ ows and out‡ ows. Over the available time sample, the quantitative role of common factors –as measured by their variance contribution –displays a cyclical pattern, especially marked among advanced countries. During an initial ’ globalization’stage up to the 2008 crisis, common factors gain growing importance. In particular, global factors become increasingly dominant, partly at 4 Our results appear to stand in contrast with those of Cerutti, Claessens and Rose (2017), in that we …nd a quantitatively larger role of common factors. However, their analysis di¤ers from ours in important respects. First, they focus on ‡ ows disaggregated by ‡ ow type, while we focus on aggregate gross ‡ ows. Second, they consider only global factors a¤ecting all countres, while we also consider group factors speci…c to advanced, emerging and developing countries. And third, country coverage is also di¤erent. Our focus is primarily on advanced and emerging countries, while theirs is esentially on emerging and developing countries. Additionally, country samples di¤er, primarily in that theirs includes some 20 former socialist economies whose annual time series are shorter than required for our analysis (they use quarterly data), while ours includes six emerging markets and twenty-four developing countries (mainly in Africa) omitted from theirs. 4 the expense of group factors. After the crisis, the trend goes in reverse and ’ deglobalization’ takes place: the role of common factors and, in particular, global factors, experiences a decline. Lastly, we …nd that countries’ exposure to global cycles, as captured by the estimated loadings on the global factors, is signi…cantly related to their structural and policy framework. Financial openness and …nancial depth amplify the impact of global cycles on both the in‡ ow and the out‡ ow sides of capital ‡ ows. Additionally, a less-‡ exible exchange rate regime has the same e¤ect on the gross in‡ ow side. The latter result implies that exchange rate pegs reduce, to an economically signi…cant extent, countries’ ability to insulate their …nancial conditions from changes in global …nancial centers – suggesting that, in spite of the global …nancial cycle, the Mundellian Trilemma continues to characterize the trade-o¤s in the choice of exchange rate regime, capital account openness, and monetary policy autonomy. The rest of the paper is organized as follows. Section 2 describes the two-level latent factor model employed in the estimation. Section 3 turns to the data. Empirical results are presented in Section 4. We …rst report the factor model estimates, and then discuss how the factors relate to global ’push’variables. Next, we consider how commonality changes over time, and investigate the determinants of countries’exposure to global conditions. Lastly, we extend the empirical setting to encompass developing countries. Finally, Section 5 concludes. 2 Analytical framework Our primary objective is to analyze the comovement of gross capital ‡ ows across countries, and assess the respective roles of global, group and idiosyncratic factors. For this purpose, we study the behavior of capital in‡ ows and out‡ ows in a large panel dataset. To capture the cross-sectional dependence of capital ‡ ows, we use a latent factor model. Speci…cally, we consider a simple two-level factor model: 0 0 ym;it = m;i Gt + ( m;i ) Fm;t + um;it i = 1; :::Nm ; m = 1; :::; M ; t = 1; :::T: (1) Here ym;it denotes the chosen measure of gross capital in‡ows or out‡ows for the i-th country of group (or region) m over period t; Gt is a set of rG unobserved common (world) factors, and Fm;t is a set of rm unobserved factors speci…c to group m. In turn, m;i and m;i are the respective factor loadings, and um;it is an idiosyncratic component that may be heteroskedastic and serially and/or cross-sectionally (weakly) correlated. Stacking all the observations for group m at time t, we can write 5 Ymt = m Gt + m Fmt + umt (2) P and combining capital ‡ows for all groups into a T N matrix Y (where N = Nm ) the model can be compactly written 0 0 Y =G +F +U (3) P where G and F respectively are (T rG ) and (T rm ) matrices of factors, and and P are (N rG ) and (N rm ) matrices of global and group factor loadings respectively. In particular, is block-diagonal, with the m-th block containing the loadings of the Nm countries in the m-th group on their rm group factors. The model as written is static, with factors a¤ecting the dependent variable only con- temporaneously. However, it can be reinterpreted as a dynamic factor model with lagged e¤ects of the factors, by expressing the lags as additional static factors. As the factors and their loadings are unobserved, both need to be estimated from the data. With the model as written, they cannot be identi…ed without imposing additional 0 restrictions. A set of restrictions that yields exact identi…cation is the following: (i) GTG = IrG 0 F 0 0 and Fm T m = Irm for all m; (ii) and m m , m = 1; :::; M; are diagonal matrices, and (iii) 0 Fm G = 0 for all m. Restrictions (i) impose a normalization of the factors, in particular forcing the global factors to be mutually orthogonal, and similarly for the group factors of any given group. In turn, (ii) uniquely determines the rotation of the factors; (i) and (ii) are commonly imposed in single-level factor models. Lastly, (iii) imposes orthogonality between global and group factors. These restrictions su¢ ce to uniquely identify the factors and loadings, up to a sign change. In the empirical estimation we determine the sign of each factor by requiring that the loading of the respective group’ s largest country be positive.5 Importantly, there is no need to impose orthogonality between the group factors of di¤erent groups to identify the model, in contrast with what is often done in Bayesian analyses of multi-level factor models. Such restriction may or may not hold in practice, and imposing it leads to an overidenti…ed model. Indeed, imposing the restriction when it does not hold would result in inconsistent estimates. We …nd below that group factors are not mutually orthogonal in our data, although the magnitude of their correlation is fairly modest. In contrast with single-level factor models, which can be estimated by straightforward application of principal component analysis, estimation of the multilevel model (3) faces 5 An alternative is to set the sign of each factor so that the majority of its loadings are positive. In our case, this rule leads to exactly the same sign choices as the one in the text. 6 the di¢ culty that the matrix of group factor loadings contains zero restrictions. This prevents a standard principal-components approach, which cannot separately identify G and F . Most previous literature has confronted this issue employing Bayesian techniques (e.g., Kose et al 2003). A recently-developed alternative, which we shall follow below, builds on an extension of the principal component approach to multi-level models; see e.g., Wang (2014), Breitung and Eickmeier (2016) and Choi et al (2017). Compared with Bayesian estimation, these methods are computationally much simpler, as they just involve a sequence of OLS regressions. Their objective is to minimize the sum of squared residuals 0 0 0 0 0 SSR(G; F; ; ) = tr Y G +F Y G +F with respect to G; F; and ; subject to the identifying restrictions listed above: Estimation proceeds in iterative fashion: starting from an initial estimate of the global factors, estimates of the group factors are computed for each country group. With these, an updated estimate of the global factors can be obtained, and the process is repeated until convergence. Below we follow the sequential least squares approach of Breitung and Eickmeier (2016) and Choi et al (2017), with the initial estimate chosen through canonical correlation analysis.6 When N and T are both large, principal components estimators are consistent under general forms of heteroskedasticity and (weak) serial and cross-sectional correlation of the idiosyncratic components. Further, their small-sample performance in the multi-level factor setting is quite satisfactory (Choi et al 2017).7 However, if the idiosyncratic components are not iid; a more e¢ cient estimator may be available. Speci…cally, letting = E (ut u0t ); where ut = (u(1;1);t ; :::; u(M;NM );t )0 ; Choi’ s generalized principal component estimator (GPCE) is obtained by minimizing 0 0 0 0 0 1 tr Y G +F Y G +F with respect to G; F; and (Choi 2012). If is not known, a feasible estimator (FGPCE) can be computed using a consistent estimate ^ . The FGPCE may allow signi…cant e¢ ciency gains regarding the estimated factors and common components: This is relevant in our case because we consider a large sample of countries that display a good deal of heterogeneity in 6 candi- In essence, the initial estimate of the global factors is constructed through linear combination of ’ date’group factors. The linear combinations are chosen so as to maximize the correlation with the candidate group factors across all groups. 7 Performance is particularly robust for the estimates of the global factors, regardless of sample sizes in either dimension, and irrespective of the properties of the idiosyncratic components. In turn, the performance of the estimates of the regional factors is signi…cantly a¤ected by Nm ; the size of the groups. As Nm grows, performance improves signi…cantly, especially if the idiosyncratic components are not iid. 7 terms of the volume and variability of capital ‡ ows, so that cross-sectional heteroskedasticity in particular is a concern. Below we use the HAC covariance estimator of Andrews and Monahan (1992) to estimate ^ from the residuals of a …rst-round estimation. The above discussion assumes that the numbers of global and group factors are known a priori, which in practice is rarely the case. Following Choi and Jeong (2017), we determine the appropriate number of factors using the ICp2 , BIC; and HQ criteria, as adapted to the multi-level setting by Choi et al (2017).8 3 Data To study the global and group patterns of capital ‡ ows, we assemble a large cross-country dataset drawing from the International Monetary Fund’ s Balance of Payments Statistics (BoP). After dropping countries with incomplete data and very small economies, we end up with a balanced panel of 85 countries covering the years 1979-2015, with a total of 3,145 observations.9 We classify the countries into three groups: advanced (19 countries), emerging (28) and developing (38). The list of countries and their grouping are given in Table A1 in the appendix. Following Broner et al (2013), we construct two measures of capital ‡ ows from the BoP data: i. Capital in‡ ows by foreign agents (CIF): the sum of direct investment in the report- ing economy, portfolio investment liabilities, and other investment liabilities. ii. Capital out‡ ows by domestic agents (COD): the sum of direct investment abroad, portfolio investment assets, other investment assets, and international reserve assets. These measures of ‡ ows relate to the assets and liabilities of the reporting country’ s residents vis-a-vis non-residents. CIF is recorded as capital in‡ows to the reporting economy by foreign agents, with a positive entry indicating an increase in foreigners’ holdings of domestic assets. Similarly, COD records ‡ ows from the reporting economy, with a positive value denoting an increase in the holdings of foreign assets by domestic agents. Hence a positive COD represents a capital out‡ ow by domestic agents, while a negative COD means capital repatriation. 8 We standardize the data prior to estimation subtracting the country-speci…c mean and dividing by the country-speci…c standard deviation, as recommended by Choi and Jeong (2017). 9 Speci…cally, we download capital ‡ ows data from 1945 to 2015 for 196 countries. The data is heavily unbalanced, with some countries possessing very few observations. We construct a balanced panel comprising 98 countries with complete data from 1979 to 2015. We exclude from this sample 13 countries with population fewer than 500,000 in 2005. In addition to dropping very small countries, this also has the e¤ect of removing from the dataset several o¤shore …nancial centers and tax havens that display an extremely high volume of …nancial ‡ ows. We are left with 85 countries. 8 Importantly, both CIF and COD are net concepts. They do not represent gross purchases of domestic assets by foreign residents, or gross purchases of foreign assets by domestic residents. However, in keeping with common usage, we shall refer to them somewhat loosely as "gross in‡ ows" and "gross out‡ ows", respectively. Like Broner et al (2013), we scale capital ‡ows by trend GDP.10 We use trend rather than actual GDP to prevent the short-term cross-country comovement of GDP found in the data from distorting the estimates of the common factors and common components of capital ‡ ows. To study the covariates of the global and group factors, below we use a set of variables commonly employed to capture world real and …nancial conditions. Speci…cally: (i) global risk, as measured by the CBOE Volatility Index (VIX) and similar measures; (ii) the global short-term interest rate, given by the FED e¤ective federal funds rate minus the U.S. GDP in‡ ation rate; (iii) global growth, as measured by the real GDP growth rate of the U.S. (although we experiment also with G7 and world GDP growth); (iv) the real e¤ective ex- change rate of the U.S. dollar; and (v) an index of real commodity prices –speci…cally, the UNCTAD index for metals and minerals, de‡ ated by the U.S. GDP de‡ ator. Likewise, to assess the covariates of the factor loadings, we use a set of variables cap- turing countries’structural and policy framework. These include: (i) …nancial openness, as measured by the Chinn-Ito index; (ii) trade openness, given by total exports plus imports as a percentage of GDP, and expressed in log terms; (iii) …nancial depth, measured by domestic credit to the private sector as a percentage of GDP; and (iv) the exchange rate regime, sum- marized by the index of de facto exchange rate arrangements of Ghosh, Ostry and Qureshi (2015). Table A2 in the appendix gives the details on the data sources. Figure 1 shows the time path of aggregate ‡ ows, for the full country sample as well as each of the three country groups, expressed as percent of the respective group’ s trend GDP. The full-sample ‡ ows reveal two facts. First, in‡ows and out‡ ows exhibit a high degree of comovement. Second, both in‡ ows and out‡ ows display pronounced cycles. Inspection of the …gures corresponding the three country groups shows that the comovement between in‡ ows and out‡ ows is particularly noticeable among advanced countries. Also, the exact timing of the capital ‡ ow cycles varies across country groups. Advanced-country ‡ ows show an upward trend starting in the mid-1990s. They peak at around 25 percent of trend GDP just before the global crisis of 2007-2008, and collapse thereafter. In turn, emerging-market ‡ows exhibit wide swings, with sharp drops at the time of the 1981 debt crisis and the East Asia crisis of the late 1990s. They subsequently peak in 2006, and fall sharply afterwards. Finally, 10 Trend GDP is calculated applying the Hodrick-Prescott …lter, using a parameter of 100, to the series of nominal GDP in U.S. dollars. 9 developing countries’in‡ ows and out‡ ows experience a steep fall in the 1980s, and recover thereafter. They peak in 2011, after the global …nancial crisis, and decline subsequently. Table 1 reports descriptive statistics. The …gures in the table are group averages of the underlying country data. Like with the aggregates in Figure 1, advanced countries exhibit by far the largest gross ‡ ows as percent of trend GDP. Moreover, as noted by Broner et al (2013), their in‡ows and out‡ ows are highly synchronized –their average contemporaneous 11 correlation exceeds 0.90. The correlation between CIF and COD is smaller for emerging and, especially, developing countries. The individual-country data, shown in Table A3 in the appendix, reveal that all of the 19 advanced-country in‡ ow-out‡ ow correlations are signi…cantly positive.12 This is also the case for 22 out of 28 emerging countries, but only for 10 out of 30 developing countries. On the other hand, advanced countries show the smallest degree of capital ‡ow variability, as measured by the coe¢ cient of variation, while developing countries show the largest, for both in‡ ows and out‡ ows. 4 Empirical results To check the suitability of a latent factor model for characterizing the patterns of capital ‡ows, we …rst assess their degree of international comovement. Table 2 reports cross-country correlations of gross capital ‡ ows. The numbers shown in the top panel of the table are averages of individual-country …gures. In each block, the diagonal entries correspond to within-group correlations (i.e., the average of all the intra-group pairwise cross-country cor- relations), while the o¤-diagonal entries are average between-group correlations.13 In turn, the bottom panel of the table indicates what fraction of the total number of individual cor- relations underlying each of the averages in the top panel are signi…cantly di¤erent from zero. All but one of the average within-group correlations are positive, with the negative entry corresponding to the developing country group’ s in‡ ow-out‡ ow correlation. For both in‡ows and out‡ ows, the highest average within-group correlation corresponds to advanced countries – they are the countries whose ‡ ows co-move most strongly. The …gures shown in the 11 The correlation is similarly high (0.86) if ‡ows are expressed in …rst di¤erences rather than ratios to trend GDP. In a disaggregated analysis of capital ‡ows by economic sector, Avdjiev et al (2017b) show that the positive correlation between in‡ ows and out‡ ows, especially in advanced countries, is primarily due to banks, although the in‡ ows and out‡ ows of corporates and government also show positive (but smaller) correlation. On this issue see also Davis and van Wincoop (2017). q 2 12 The standard error of a correlation r is approximated as 1 r T 1: 13 The average in‡ow-out‡ ow correlations in the table exclude the within-country correlation. On the other hand, all the qualitative conclusions in the text continue to hold if ‡ ows are expressed as …rst di¤erences rather than ratios to trend GDP. 10 bottom panel of the table con…rm this view –between 80 and 90 percent of all the pairwise advanced-country correlations are signi…cant, a much higher percentage than for the other groups. Moreover, the large value of the advanced-country in‡ ow-out‡ ow cross-country correlation (.43) suggests that among advanced countries gross in‡ ows (out‡ ows) frequently come from (go to) other advanced countries. At the other extreme, the smallest within-group correlations (including the one negative entry) correspond to developing countries, who also tend to exhibit the smallest percentages of signi…cant correlations in the bottom panel of the table. Finally, the data also suggest that gross out‡ows exhibit stronger commonality than gross in‡ ows, as shown by the fact that, in both panels of the table, all but one of the entries in the southeast block exceed the corresponding entries in the northwest block. As for the average between-group correlations also shown in the top panel of the table, in all three blocks the largest one corresponds to the advanced country-emerging country pair, suggesting that the capital ‡ ows of these two groups are the most likely to share common factors. In contrast, the smallest values correspond to developing countries, which also account for the three negative o¤-diagonal entries in the table. Overall, the information in Table 2 suggests that gross capital ‡ ows exhibit a good deal of commonality among advanced and emerging markets – and especially in the former. In contrast, developing countries’capital ‡ ows show less commonality with the ‡ ows of other countries –whether developing, emerging, or advanced. These results indicate the presence of cross-sectional dependence in capital ‡ ows, espe- cially among advanced and emerging countries, but do not tell us if the dependence is strong or weak. Strong dependence arises from pervasive common factors, i.e., factors that a¤ect many countries. Weak dependence re‡ ects localized interactions, e.g., bilateral …nancial linkages. The distinction is important, because standard factor models provide a suitable characterization of the former but not the latter form of dependence.14 To assess if dependence is strong or weak, we turn to the exponent of cross-sectional dependence (Bailey et al 2015). It can be viewed as a measure of the rate at which factor loadings (fail to) die o¤ as cross-sectional sample size grows. The exponent ranges between zero and one, with a value of 1 indicating the presence of strong dependence. Table 3 reports the computed values for the three country groups, along with the 95 percent con…dence bands. For both advanced and emerging countries, and for both in‡ ows and out‡ ows, the exponents 14 Strong and weak cross-sectional dependence are de…ned in terms of the rate at which the largest eigen- value of the covariance matrix of the cross-section units rises with the number of the cross-section units. In the cross-sectional dependence literature, strong dependence is typically modeled with factor models, while weak dependence is modeled with spatial models. Estimation of standard factor models on weakly cross-sectionally dependent data is likely to yield inconsistent estimates; see, e.g., Onatski (2012). 11 of cross-sectional dependence exceed 0.95, and the con…dence regions include 1. This con…rms the presence of strong cross-section dependence in the gross in‡ ows and out‡ ows of these 15 two country groups. In contrast, for developing countries the evidence is less supportive of strong dependence. The estimated exponent of cross-sectional dependence is just above 0.8 for both CIF and COD, and the 95 percent con…dence region does not reach up to 0.90 in either case. This result is in line with the weaker commonality found in Table 2 for developing countries’ capital ‡ ows. It also raises doubts on the suitability of a factor model to describe the patterns of capital ‡ ows of this country group. In light of this evidence, the analysis below focuses primarily on advanced and emerging countries. Developing countries are considered in a subsequent extension. 4.1 Factor model estimates We turn to estimation of the global and group factors. To that end, we estimate the model (1) for gross in‡ ows and gross out‡ ows. In each case, we compute the three information criteria mentioned earlier (ICp2 , BIC and HQ) for speci…cations ranging from 1 to 3 global factors and 1 to 3 group factors per group. All three criteria select one global and one group factor per country group (see Table A4 in the Appendix). As noted, we determine the sign of each factor so that the largest economy in the respective group carries a positive loading. For advanced and emerging countries, respectively, the largest economies are the U.S. and China. For the global factors, we set the sign using again the U.S. loading. By construction, the estimated factors have zero mean and unit variance. Figure 2 plots the estimated global factors for gross in‡ ows and out‡ ows. Both display a steep rise since the mid-1990s until the inception of the global crisis, and a sharp decline afterwards. This pattern roughly matches the one shown in Figure 1(a) for aggregate ‡ ows. Figure 3 reports the group factors. The advanced-country factors (panel (a)) exhibit a gradual rise since the mid- to late 1980s, with a hump in the late 1990s, and an abrupt collapse in 2007, especially marked in the case of out‡ ows. In turn, the emerging-country factors (panel (b)) display wide swings coinciding with the debt crisis of the 1980s and the Asia-Russia crises of the 1990s. Those ‡ uctuations are as wide (wider, in the case of in‡ows) as the ‡ uctuations seen at the time of the global crisis. The common factors exhibit considerable persistence, more so in the case of the global than the group factors. The …rst-order autocorrelation coe¢ cient of the global factors is .79 for CIF and .83 for COD. For the group factors, the values range from .67 and .68 for 15 This conclusion is in contrast with that of Cerutti et al (2016), who fail to …nd common factors in advanced countries’capital ‡ ows. 12 the advanced- and emerging-country CIF factors respectively, to .28 and .21 for the COD factors. Standard ADF tests reject at the 5 percent con…dence level the null of a unit root for the emerging-country group factors, as well as the advanced-country COD factor. For the remaining factors (global CIF and COD, and advanced-country CIF), both ADF and KPSS tests fail to reject their nulls of non- and stationarity, respectively. However, unit root tests allowing for a break in constant and trend in the run up to the global crisis do reject the null of a unit root at the 5 percent level in all three cases.16 We conclude that all the common factors are stationary.17 It is also apparent from Figure 2 that the global CIF and COD factors follow very similar patterns. Their correlation coe¢ cient equals 0.95. Table 4, which reports the correlation pattern of the group factors, shows that the advanced-country CIF and COD factors are also highly correlated – their correlation equals 0.82. Overall, this implies that advanced- country in‡ ows and out‡ ows are driven essentially by the same common shocks. In turn, the emerging-country group factors are also positively correlated, but to a more limited extent – their correlation is just 0.34. The table also shows that the group factors are not signi…cantly correlated across groups. This applies to both in‡ ow factors and out‡ ow factors, as well as the cross-group in‡ ow-out‡ ow correlation. The estimated factor model provides a satisfactory account of the observed cross-country comovement of capital ‡ ows. The exponents of cross-sectional dependence of the CIF and COD residuals from the estimation equal 0.40 and 0.33, respectively, and their 95 percent con…dence regions reach up to 0.50 and 0.37, far below 1. Thus, once the common factors have been removed, the residuals show no traces of strong cross-sectional dependence. Also, panel unit root and stationarity tests indicate that the residuals are stationary (see Table A5 in the Appendix). The sensitivity of each country’ s gross ‡ows to global and group-speci…c common factors is given by its factor loadings ( m;i and m;i in equation (1)). Figures 4 and 5 show the estimated global and group factor loadings, respectively. The dots denote the 95-percent con…dence bands. For advanced countries, Figure 4 shows that the global factor loading estimates, for both in‡ ows and out‡ ows, are positive in every case, except for the COD global factor loading in the case of New Zealand. Further, all the loadings, except New Zealand’ s (for both CIF and COD) are signi…cant at the 5 percent level.18 For emerging markets, 16 When the test procedure endogenously selects the date of the break, it is placed at some point in the 2005- 2007 interval, depending on the exact speci…cation of the test and the common factor under consideration. 17 It is important to note that, even if the factors were I(1), estimates of the factor model in levels (as computed here) should perform better than estimates of the model in …rst di¤erences, as long as the idiosyn- cratic components are I(0); see Choi (2017) and Ergeman and Rodriguez-Caballero (2016). Such condition holds in our case, as noted later in the text. 18 Inference on the loadings is based on Choi (2012). His analysis applies to single-level factor models. 13 twenty-one of the estimated loadings on the global CIF factor are positive, and fourteen of them are statistically signi…cant. The remaining seven are negative, although only one of them (which corresponds to the Philippines) is statistically signi…cant. In turn, all but two of the emerging-country COD global factor loading estimates are positive and signi…cant. The exceptions are Uruguay and Pakistan, whose loadings are small and insigni…cant. In general, the largest global factor loadings are found among the advanced countries (there are some exceptions, such as India). In other words, advanced countries are more exposed than emerging countries to the global factors. For in‡ ows, the median global factor loading is .53 for advanced countries, and .25 for emerging countries. For out‡ ows, the median estimates are .54 and .48, respectively. Further, countries with large loadings on the CIF global factor also tend to exhibit large loadings on the COD global factor: the correlation between both sets of loadings is 0.62, although it is larger for advanced countries (.72) than for emerging countries (.45). Figure 5 presents similar information for the group factor loadings. Sixteen advanced countries exhibit positive loadings on the CIF group factor, of which all except New Zealand’ s are signi…cant. The remaining three loadings (corresponding to Canada, Japan and Aus- tralia) are signi…cantly negative. The same three countries, plus Norway and Finland, exhibit negative loadings on the COD group factor, although only Australia’ s is signi…cantly nega- tive. The other fourteen loadings on the COD group factor are all positive, and all signi…cant except for that of the Netherlands. Interestingly, the largest loading for both CIF and COD belongs to the U.K., perhaps re‡ ecting its role as …nancial center. In turn, all of the emerging-market CIF group factor loadings are positive, and twenty- one of them are statistically signi…cant. As for the COD group factor loading estimates, twenty-tree are positive, of which …fteen signi…cantly so. The remaining …ve are negative, although none is signi…cant. Interestingly, for both CIF and COD the smallest (or negative) group factor loading estimates tend to be found among Middle Eastern countries, while the largest ones are found among East Asian and Latin American emerging markets. Unlike with the global factor loadings, emerging countries exhibit group factor loadings roughly as large or even larger than those of advanced countries. The median loading on the group CIF factor is .39 for advanced countries and .40 for emerging countries. For COD, the corresponding …gures are .31 and .24. In fact, the largest group factor loading corresponds to the Philippines in the case of CIF, and to Thailand in the case of COD. Across countries, CIF and COD group factor loadings show a strong positive correlation, more so for advanced countries (the correlation equals .83) than for emerging countries (.73) – the However, Wang’ s (2014) results suggest it should apply also to the multi-level setting, although a formal proof is not available at present. 14 same patterns found for the global factor loadings. The clear conclusion is that countries’ exposure to the international drivers of gross in‡ows goes hand-in-hand with their exposure to the international drivers of gross out‡ows. 4.2 The variance contribution of common factors The orthogonality conditions imposed to identify the factor model allow a straightforward de- composition of the variance of each country’ s capital ‡ows into three orthogonal components: a global component, a group component, and a country (or idiosyncratic) component.19 Ta- ble 5 summarizes the results for both in‡ ows and out‡ ows.20 The …gures shown are averages of the individual-country results. Four facts stand out. First, the global and group-speci…c common factors account, on average, for close to half the variance of gross capital ‡ ows. Second, global factors play a bigger role than group factors. This is particularly the case for gross out‡ ows; in contrast, gross in‡ ows are more strongly a¤ected than gross out‡ ows by group factors. Together, the latter two observations suggest that a major part of the variation of gross capital ‡ ows over the cycle may re‡ ect out‡ ows from a relatively large number of countries and into a relatively small number of countries – possibly following a ’ risk-on / risk-o¤’ pattern according to which capital ‡ ows into safe havens as investors run for cover during global busts, and into high-return destinations as international investors engage in search for yield during global booms. Third, there is a major di¤erence between advanced and emerging countries. Common factors contribute a much bigger share of the variance among the former countries (averaging almost 60 percent for in‡ ows and 64 percent for out‡ ows) than among the latter (35 and 37 percent, respectively). Put di¤erently, local factors dominate emerging-country capital ‡ ows, while common factors dominate advanced-country ‡ ows. Fourth, the di¤erence is primarily due to the global factors: their contribution to the overall variance of ‡ows is much larger for advanced economies than for emerging countries. This applies to both in‡ ows and out‡ ows: among advanced countries, global factors account for 38 percent of the variance of in‡ ows and 47 percent of the variance of out‡ ows, while the corresponding …gures for emerging countries are just 15 percent and 25 percent, respectively. In contrast, the role of group 19 Because group factors are not mutually orthogonal in our setting, the variance contribution of each group factor can be in principle further decomposed into the portion attributable to the factor’s component uncorrelated with other groups’factors, and that attributable to the component correlated with the factors of other groups However, the cross-group correlation between group factors is su¢ ciently low that the latter contribution is virtually negligible, and therefore we do not report such additional decomposition. 20 Cerutti, Claessens and Puy (2016) carry out a similar decomposition for di¤erent types of gross in‡ ows to a set of emerging markets. 15 factors is, on average, not very di¤erent across the two sets of countries. As a result, the variance contribution of global factors far outweighs that of group factors among advanced countries, but not among emerging countries – in fact, the opposite happens in the case of emerging-country gross in‡ ows. Importantly, these results are not due to the fact that the advanced-country group includes the leading global …nancial centers (the U.S., U.K., Germany and/or Japan). Table A6 in the Appendix shows that excluding them has little e¤ect on the average variance contributions shown in Table 5. Figure 6 reports the individual-country variance decomposition results underlying Table 5. The respective roles of common and idiosyncratic factors exhibit considerable variation across countries, even within the same country group. Among advanced countries, common factors contribute the bulk of the variance of both gross in‡ ows and out‡ ows in the Nether- lands and Norway, but play only a modest role in New Zealand, Japan and Finland, where idiosyncratic factors dominate. Overall, common factors account for at least one-fourth of the variance of both in‡ ows and out‡ ows in all countries except New Zealand. The contribution of the global factors in particular also shows substantial variation across countries. They play virtually no role in New Zealand, but account for over 80 percent of the variance in the Netherlands, for in‡ ows as well as out‡ ows in both cases. Further, their role is not disproportionately larger in the four center countries mentioned above: in three of them, the contribution of the global factors is below the group average. The exceptions are the U.S., in the case of in‡ ows, and Germany, in the case of out‡ ows. In turn, group factors are virtually irrelevant in Finland, but play a dominant role in the U.K., again for both in‡ ows and out‡ ows. Emerging markets also display considerable heterogeneity along these dimensions. In some countries (e.g., Egypt, Pakistan), neither global nor group factors play any signi…cant role for in‡ ows or out‡ ows. In contrast, in several major emerging markets they account for half the variance or more (e.g., Korea, India). The global factor contributes less than 10 percent of the variance of gross in‡ ows in over half the countries in the group. At the other extreme, it accounts for two-thirds of the variance of Turkey’ s in‡ows and China’ s out‡ ows. Interestingly, among Middle Eastern economies –Egypt, Jordan, Kuwait, Oman, Saudi Arabia –as well as Cyprus and Turkey, the contribution of group factors tends to be quite small, suggesting that, once global forces are taken into account, their capital ‡ ows do not have much in common with those of other emerging markets. 16 4.3 Common factors and the global …nancial cycle The variance share of common factors shown in Table 5 can be interpreted as the contribution of the international cycle to the variation of capital in‡ows and out‡ ows. A separate question is what drives that cycle, and how it is re‡ push vs pull’ ected in observable variables. The ’ literature, going back to Calvo, Leiderman and Reinhart (1996), as well as the more recent literature on the global …nancial cycle (starting with Rey 2013), have stressed the role of a few global (or …nancial-center) variables capturing …nancial conditions worldwide as key drivers of the commonality of capital ‡ ows. Global risk, as summarized by the VIX and similar measures, is often taken as a sort of summary statistic of the ’ global …nancial cycle’ (Rey 2013, Miranda-Agrippino and Rey 2015), and typically found to be negatively related to capital ‡ ows, especially in‡ ows to emerging markets.21 However, other global …nancial variables capturing advanced-country monetary policy (such as the U.S. short-term interest rate and/or the term premium), as well as global growth, are often found to a¤ect capital ‡ ows to multiple countries, even after controlling for the VIX or other risk proxies (e.g., Avdjiev et al 2017a; Cerutti, Claesens and Ratnovsky 2017). Bruno and Shin (2015b) argue that the same should apply to the U.S. real exchange rate: because of the dominant role of the dollar as currency of denomination of …nancial contracts worldwide, dollar appreciation constitutes a tightening of global …nancial conditions. To explore the fundamental covariates of the global …nancial cycle, Table 6 reports re- gressions of the estimated common factors on selected global variables (or U.S. variables, as they pertain to the world’ s leading …nancial center).22 The top block of the table reports univariate regressions of the factors on the VIX and other risk proxies. Because the VIX is not available prior to 1990, we also use its predecessor the VXO, which is available since 1986, as well as the BAA 10-year corporate spread, which represents a commonly-used measure of market risk premia. Consistent with earlier literature, all the factors exhibit a signi…cant negative correlation with the three measures of risk. This applies to the global factors as well as the advanced-country and emerging-country group factors, and to both in‡ ows and out‡ ows. The largest coe¢ cient estimates are found in the regression with the advanced- country COD factor as dependent variable, suggesting that risk particularly discourages out‡ ows from these countries, over and above its e¤ects on other ‡ ows. 21 See Forbes and Warnock (2012), Broner et al (2013), Bruno and Shin (2015a,b), Cerutti, Claessens and Ratnovsky (2017), Avdjiev et al (2017a,b), Eichengreen, Gupta and Masetti (2017), and Obstfeld, Ostry and Qureshi (2017). 22 Because of the high persistence of the factors, to avoid spurious inferences the exercises are run with the variables expressed in …rst di¤erences. This tends to weaken the explanatory power of the regressions, so they o¤er a conservative view on the strength of the relations under consideration. At the same time, the samples are short enough that the results should be taken with caution. 17 The bottom block of Table 6 reports regressions of the factors on the VIX plus other global variables. The main purpose of the regressions is to assess their ability to account for the variation of the common factors, rather than to establish the sign or magnitude of particular coe¢ cients. The additional explanatory variables are the main ones used in the recent literature, and intend to provide a minimalist representation of the global …nancial cycle: the real short-term U.S. interest rate, the U.S. real exchange rate, and the U.S. real growth rate.23 We also include global commodity prices (measured by the world price of metals and minerals in real terms), whose links with global capital ‡ ows have been stressed 24 by Reinhart, Reinhart and Trebesch (2017). The results in Table 6 show that the estimated coe¢ cient on the VIX remains negative and signi…cant in all six columns. Moreover, the magnitude of its coe¢ cient is not much a¤ected by the presence of the additional variables. In turn, the real interest rate carries in most cases a positive coe¢ cient, but it is signi…cant only in the advanced-country CIF factor regression. The signs of the coe¢ cients of the other regressors are more heterogeneous. The U.S. real exchange rate carries a signi…cant positive sign in the global CIF factor regression, and signi…cant negative coe¢ cients in the group CIF factor regressions. Growth has a positive and signi…cant e¤ect on the global CIF and COD factors, as well as the advanced- country CIF factor, but a negative one on the emerging-country COD factor. Lastly, the commodity price index carries signi…cant coe¢ cients in all regressions except for that of the emerging-country CIF factor. The coe¢ cients are positive in the global factor regressions, and negative in the rest. Overall, the R2 of these augmented regressions indicate that a small set of variables capturing real and …nancial conditions worldwide can account for the bulk of the variation of the global factors –over 70 percent for the CIF factor, and over 80 percent for the COD factor. They also account for a respectable portion of the variation of the group factors – ranging from 45 percent for the emerging-market factors, to 55 percent for the advanced- country COD factor. Combining these results with the variance decomposition in Table 5, we can infer the extent to which this handful of fundamental variables can explain the variation in capital ‡ows across the world through the global and group factors combined. Simple calculations show that they account for about 50 percent of the variance of advanced- country out‡ ows, and about 40 percent of the variance of advanced-country in‡ ows. For emerging-country ‡ ows, the …gures are smaller – around 25 and 20 percent for COD and CIF, respectively. However, we should keep in mind that these are conservative estimates, 23 The results are very similar if the growth rate U.S. GDP is replaced with that of G7 GDP or world GDP. 24 Augmenting these regressions with other variables summarizing global …nancial conditions, such as the U.S. term spread or the TED spread, yields very small increases in explanatory power, along with insigni…cant parameter estimates for the additional regressors. 18 given that the regressions in Table 6 are run in di¤erences, which likely tends to understate their explanatory power. 4.4 deglobalization’post-crisis? Has there been a ’ It seems plausible that the global trend towards …nancial integration witnessed over the last quarter-century should be re‡ ected in a growing e¤ect of common factors on capital ‡ ows.25 However, it has been argued that the rising trend was interrupted by the global crisis, and followed by what has been termed ’ …nancial deglobalization’ – attributed to regulatory and other policy measures discouraging, in particular, cross-border bank lending (Rose and Wieladek 2014; Forbes, Reinhart and Wieladek 2017). Our setting allows a straightforward assessment of changes in the reach of common factors over time, by examining the time pattern of their variance contribution. To do this, we re- estimate the factor model over rolling time samples. Speci…cally, we use 20-year windows, starting with 1979-98, shifting them forward one year at a time, so that the …nal estimation is done over the 1996-2015 sample. This yields eighteen di¤erent estimates of the factor model. For each one of them, we compute the variance decomposition as done above. Because the time samples underlying these estimates are short, the results need to be taken with some caution. Figure 7 plots the time path of the respective variance contributions of global, group and local factors that results from these rolling estimates. The decomposition pertaining to each estimation window is denoted by the window’ s …nal year. Like in Table 5, for each window the …gure reports group averages for advanced and emerging countries. Among advanced countries, for both CIF and COD the role of common factors grows markedly over the …rst ten windows. For CIF, the average variance contribution of common factors rises from 45 percent to just over 70 percent; for COD, from 55 percent also to just over 70 percent. Most of this increase is concentrated in the …rst few years, and is attributable to the rising importance of the global factors, whose contribution grows partly at the expense of that of the group factors. For both in‡ ows and out‡ ows, the variance share of global factors peaks at over 60 percent in the window ending at the onset of the global crisis (2007). At its peak, the combined variance contribution of common factors reaches 72 percent for both CIF and COD. Thereafter, as the subsequent estimation windows start including post-crisis years, the share of the common factors – especially the global factors – in the variance of both CIF and COD enters a period of gradual decline. In the context of the factor model, this literally is a ’deglobalization’. In the …nal window of the sample 25 This was the case for FDI over the 1990s: Albuquerque, Loayza and Servén (2006) …nd that external …nancial liberalization was re‡ected in a rising contribution of global factors to the variation of FDI. 19 (1996-2015), common factors account for 60 percent of the variance of CIF and COD, with global factors contributing close to two-thirds of that total.26 Emerging countries exhibit similar ‡ uctuations, but on a reduced scale. In their case, the variance contribution of common factors remains consistently smaller than in advanced countries, for both CIF and COD. Their variance share remains relatively ‡ at until the mid- 2000s, and then rises until the global crisis – especially in the case of out‡ows – re‡ ecting an increasing role of global factors. At the peak, common factors account for just under 50 percent of the overall variance of both in‡ ows and out‡ ows. Like in advanced countries, the contribution of the common factors declines post-crisis – although in this case there is no noticeable change in the relative shares of global and group factors. In the …nal window of the sample, the global and group factors combined account for some 40 percent of the overall variance of both CIF and COD. Unlike in advanced countries, where global factors play by far the biggest role for both in‡ ows and out‡ ows, among emerging countries the relative roles of global and group factors vary depending on the direction of the ‡ ows. The former factors outweigh the latter for out‡ ows, but the opposite happens with in‡ ows. 4.5 What shapes the role of the global factors? The global crisis has prompted renewed interest in the factors that shape countries’exposure to common shocks through their capital ‡ ows. As Figures 5 and 6 show, exposure varies a lot across countries. This raises the question of what drives such heterogeneity. In this regard, structural and policy features such as capital account openness and …nancial development have received considerable attention. In addition, the ability of ‡ exible exchange rates to provide insulation from the global …nancial cycle has attracted an active debate following Rey’ s (2013) in‡ uential work. As a …rst illustration of these issues, we present variance decompositions of capital ‡ows over time, analogous to those shown in Figure 7, but grouping countries according to par- ticular features of their structural and policy framework. Figure 8 does this distinguishing between countries exhibiting high and low degrees of …nancial openness. For each estimation window, countries are allocated to either group depending on whether their average …nancial openness (as measured by the Chinn-Ito index) over the window in question is above or below the overall sample median. The pattern of rise and subsequent fall of the variance contribution of common factors found in Figure 7 is apparent in the group of more …nancially 26 The rise and fall in the contribution of common factors before and after the global crisis are consistent with the changing composition of ‡ ows. The reason is that the pre-crisis capital ‡ow boom, as well as the post-crisis collapse, were led by portfolio and, especially, bank ‡ows, which are commonly found to be the ‡ow type most responsive to global variables. 20 open countries, but not in the group of less …nancially open countries. Further, the common factors play a larger role in the former countries than in the latter. This applies to both in‡ ows and out‡ ows, and it is primarily due to the variance contribution of the respective global factors, which is consistently bigger among the former group than among the latter. In turn, Figure 9 highlights the role of the exchange rate regime using the de facto classi…cation compiled by Ghosh, Ostry and Qureshi (2015), which distinguishes between …xed, intermediate, and ‡ oating regimes. For the purposes of the …gure, we consolidate the three-way classi…cation into two groups, again depending on how each country’ s average degree of ‡ exibility over each window compares to the full-sample median; this yields two pegs’and ’ groups of countries that, in a slight abuse of language, we label ’ ‡oats’. The graphs show that the in‡ ows and out‡ ows of countries on pegged regimes consistently re‡ ect the action of common factors to a greater extent than do the in‡ ows and out‡ows of countries on ‡ oating regimes. This is particularly the case for the global factors, whose variance contribution is up to twice as big in the former group than in the latter. The di¤erence between the two groups along this dimension is especially large in the case of gross out‡ ows. Moreover, while the trend of rise and fall of the quantitative role of the common factors around the global crisis a¤ects both country groups, it is much more pronounced among countries with pegs than among those with ‡ oating regimes. Finally, Figure 10 turns to …nancial depth, classifying the countries into high and low …nancial depth groups following the same procedure as in the preceding …gures. Greater …nancial depth is associated with a bigger role of common factors, for both in‡ ows and out‡ ows, which again is primarily due to the larger variance contribution of the global factors – although the di¤erence between both country groups along this dimension seems to have narrowed after the global crisis. These …gures highlight some ingredients that shape the e¤ect of common factors across countries. However, since they consider one ingredient at a time, they may not convey an accurate picture of the respective role of each one of them. An alternative way to do this is by regressing the estimated factor loadings, which capture the impact of the factors on capital ‡ ows, on suitable measures of countries’policy and structural features.27 In principle, we could run cross-sectional regressions using as dependent variable the full- sample estimates of the loadings. However, these pertain to the full 37-year sample, a time span over which any candidate explanatory variables have surely undergone major changes, which would then obscure their relationship with the loadings in a pure cross-section. To remedy this, we opt instead for using the estimates of the loadings obtained from the moving- 27 Cerutti, Claessens and Puy (2016) report a similar exercise for portfolio and banking in‡ows to emerging markets. 21 window estimation. This o¤ers two advantages: …rst, the windows cover a shorter time span (20 years), which partly mitigates (although it certainly does not eliminate) the concern with the variation of the explanatory variables over time. Second, it allows us to build a panel combining the factor loading estimates from the di¤erent windows, so that the estimation can exploit, at least to some extent, the time variation of the regressors across windows. Table 8 reports the estimation results, for both the CIF and COD global factor loadings.28 In addition to …nancial openness, …nancial depth and the exchange rate regime, we also examine the role of trade openness, which might o¤er another channel for the propagation of global …nancial conditions. The explanatory variables are averaged over the corresponding 20-year window. In particular, we compute GLS estimates, using an AR(1) speci…cation to take into account the likely persistence arising from the fact that consecutive windows share a good deal of information. The …rst …ve columns of Table 8 correspond to the regressions with the CIF global factor loadings as dependent variable. The results indicate that countries’exposure to the global forces behind capital in‡ ows rises signi…cantly with their degree of …nancial openness (column 1), as well as trade openness (column 2) and …nancial depth (column 3). A higher degree of exchange rate rigidity (i.e., a decline in the value of the exchange rate regime index) has the same e¤ect (column 4). These results are highly signi…cant, and survive when all four variables are jointly considered (column 5), except for the e¤ect of trade openness, which becomes insigni…cant. The last …ve columns of Table 8 report the results using the COD global factor loadings as dependent variable. For the most part, the estimates are not very di¤erent from those obtained with the CIF factor loadings. This is unsurprising given that, as shown earlier, the loadings on both global factors show large positive correlation. The main di¤erence is that the coe¢ cient estimate on the exchange rate regime in column 4 of the out‡ ow loadings regressions is only half as large as that in column 4 of the in‡ ow loadings regressions. Finally, when all four regressors are jointly considered, the exchange rate regime becomes insigni…cant, and only …nancial openness and …nancial depth remain statistically signi…cant. The conclusion is that, once these two variables are taken into account, the degree of exchange rate rigidity still matters for countries’sensitivity to the global forces driving capital in‡ows, but not for their sensitivity to the global forces driving capital out‡ ows.29 28 We focus on the global factor loadings, because – as Figures 4 and 5 showed – they vary markedly between advanced and emerging countries, much more so than do the group factor loadings. They also display more variation over time than do the group factor loadings, as can be inferred from the cyclical patterns shown in Figures 8-10. 29 For emerging markets, Obstfeld, Ostry and Qureshi (2017) likewise …nd a signi…cant dampening e¤ect of exchange rate ‡exibility on the responsiveness of gross in‡ows to global risk, as measured by the VXO index, but not on the response of gross out‡ ows. In contrast, Cerutti, Claessens and Puy (2017) …nd the 22 What is the economic signi…cance of the estimates shown in Table 8? To assess this question, consider the e¤ects of raising in turn each of the explanatory variables considered (except for trade openness, which is not signi…cant in the multivariate regressions) from its 25th to its 75th sample percentile. Start with …nancial openness. Simple calculations show that such a change in the capital account openness index would raise the loading on the global factor by 0.08 for in‡ ows and 0.12 for out‡ ows. With the in‡ ow and out‡ ow global factors unchanged, this in turn would increase the variance share of the global factor by 7 percentage points in the case of in‡ ows and 14 percentage points in the case of out‡ ows.30 Similar calculations show that raising …nancial depth from the 25th to the 75th percentile would increase the variance share of the global factor by 16 and 13 percentage points for in‡ ows and out‡ ows, respectively. Finally, because of the de…nition of the exchange rate regime index, raising its value from the 25th to the 75th percentile amounts to moving from a peg to a ‡ oating regime. The same calculations as with the other variables reveal that this would reduce the variance share of the global in‡ ow factor by 12 percentage points. A look at the average variance contribution of the global factors, shown in the …rst column of Table 5, helps put these calculations in perspective. Each of the experiments considered would change the variance share of global factors by as much as half of its sample average in the case of in‡ ows, and at least one-third in the case of out‡ ows. The overall conclusion is that …nancial openness, …nancial depth, and the degree of rigidity of the exchange rate regime all augment countries’ exposure to the global drivers of capital ‡ ows to an economically- signi…cant extent. 4.6 Adding developing countries So far we have focused on advanced and emerging countries, leaving aside developing coun- tries due to the lack of evidence that their capital ‡ow patterns re‡ect the action of (strong) common factors. We next extend the analysis to include the thirty-eight developing countries in Table A1 as another country group. Thus, we re-estimate the factor models for CIF and COD allowing in each case for a developing-country group factor.31 As already noted, the estimates have to be taken with extra caution, because the principal-component method un- opposite result: exchange rate ‡ exibility augments the e¤ect of common factors on emerging-country bond and bank in‡ ows. 30 To compute the change in the in‡ ow and out‡ ow loadings, we use the estimates in the last column of the respective block of Table 5. Because of the normalization imposed, the variance contribution of the factor is just given by the square of its loading. The change in the variance share of the global factor is evaluated at its median value. 31 To save space, we only provide here a brief summary of the main …ndings; detailed results are available upon request. 23 derlying our estimation approach is poorly suited to situations in which the common factors are just weakly (rather than strongly) in‡ uential. The estimated global factors, as well as the group factors for emerging and advanced coun- tries, show little change relative to those obtained when estimating the model on the sample without developing countries. The correlation of the newly-estimated factors (whether global or group-speci…c) with their respective counterparts in the two-group model exceeds .90 in all cases. In other words, they are virtually indistinguishable from those depicted in Figures 2 and 3 above, and to save space we do not show them here. Likewise, the advanced- and emerging-country loadings on the global and group factors remain virtually unchanged rela- tive to those obtained from the sample without developing countries: the correlation of the loadings from the two samples exceeds 0.97 for the global factor loadings and 0.95 for the group factor loadings. The only discernible di¤erence relative to the earlier results is the fact that advanced countries’ group factor loadings rise slightly, at the expense of their global factor loadings. The likely reason is that the addition of developing countries weakens the role of advanced countries in shaping the global factors. The latter must now also re‡ ect the patterns of developing-country gross ‡ ows, which –as shown in Table 2 –are less correlated than emerging-country gross ‡ ows with the CIF and COD of advanced countries. Figure 11 depicts the developing-country group factors. We set their sign so that the loading of the largest country (Nigeria) is positive. The CIF factor displays a peak at the onset of the debt crisis in 1981, followed by a sharp decline and a subsequent steady rising trend, interrupted by an abrupt fall in 2006. The factor’ s …rst-order autocorrelation coe¢ cient equals 0.54. The COD factor appears somewhat less persistent (its …rst-order autocorrelation is 0.43). It also shows a deep fall following the global crisis. In both cases, standard ADF unit root tests are able to reject at the 5 percent level the null of a unit root. On the other hand, the developing-country factors exhibit some distinct features. The pattern of their correlations with other factors replicates that of the between-group correla- tion of gross ‡ ows found in Table 2. In particular, the correlation between the developing- country in‡ ow and out‡ ow factors equals -0.25, in sharp contrast with the positive corre- lations between the in‡ ow and out‡ ow factors of the other country groups. Also, the CIF factor is negatively correlated with that of advanced countries, and positively correlated with that of emerging countries. In turn, the COD factor is positively correlated with both the COD and the CIF factors of advanced countries. Empirical exercises similar to those reported in Table 6 show that neither the CIF nor the COD factors of the developing-country group are signi…cantly correlated with the VIX or with the other risk proxies considered before. The same conclusion obtains from regressions similar to those in the bottom block of Table 6, adding as explanatory variables the short- 24 term U.S. real interest rate, the real exchange rate, the U.S. GDP growth rate and the relative price of commodities. The explanatory power of such regressions is very poor, with R2 under 0.1 in all but one case.32 In other words, standard ’ push’variables do not seem to play a major role as drivers of developing-country capital ‡ ows. The estimated factor loadings of developing countries also merit comment. Overall, the loadings are smaller –especially in the case of the global factors –and their signs are more heterogeneous than among advanced and emerging countries.33 Moreover, the loadings on the COD and CIF factors are virtually uncorrelated, in contrast with the large positive correlation found in the other country groups. This applies to the loadings on both the global and developing-country group factors. The conclusion is that developing-country capital ‡ ows are less re‡ ective of common external forces than are the ‡ ows of the other country groups. This just corroborates the descriptive evidence reported earlier on the limited extent of cross-sectional dependence in the developing-country data. Indeed, Table 7 con…rms this presumption. Like Table 5, it shows the decomposition of the variance of capital in‡ ows and out‡ ows, now for the three-group country sample. On average, developing countries exhibit the smallest variance contribution of the common factors, for both CIF and COD, with the di¤erence vis-a-vis the other country groups particularly noteworthy in the case of the latter. The average variance share of the idiosyncratic factors is close to 70 percent for in‡ ows, and close to 80 percent for out‡ ows. Both global and group factors –whose respective contributions are of roughly similar magnitude –play a smaller role in developing-country capital ‡ ows than they do in the ‡ ows of the other country groups. Finally, comparing Tables 5 and 7 also reveals that adding developing countries to the analysis causes some changes in the variance decomposition of the capital ‡ ows of both advanced and emerging-country groups. For the latter, the changes are minimal, for both in‡ ows and out‡ ows. For the former, the respective roles of idiosyncratic and common factors also exhibit very modest changes. However, although the total contribution of the common factors is roughly unchanged, the respective shares of global and group factors in the total variance of advanced-country in‡ ows and out‡ ows do change: for both in‡ ows and out‡ ows, 34 the share of the group factors rises at the expense of the share of the global factors. 32 To save space, the results are omitted here but are available from the authors. 33 For in‡ows, the majority of the global factor loadings (22 out of 38) are negative, thirteen of them signi…cantly so. Nine are signi…cantly positive. In contrast, all but two of the loadings on the developing- country CIF factor are positive, and 19 of them are signi…cant. In turn, all but four developing countries exhibit positive loadings on the global COD factor, and twenty-two of them are signi…cantly positive at the 5 percent level (two are negative). The signs of the loadings on the developing-country COD factor are more evenly distributed: twenty-seven are positive (of which ten signi…cant) and eleven are negative (of which four signi…cant). 34 Like with Table 5 above, the advanced-country variance decomposition results in Table 7 are not mate- 25 5 Concluding remarks The extent to which countries’…nancial ‡ ows are driven by global forces beyond their control remains a subject of interest for both academics and policy makers. This paper o¤ers a quantitative assessment of the role of common factors for the observed patterns of gross capital ‡ ows across a large number of countries. To do this, the paper implements a two- level latent factor model using a novel estimation method based on an extension of the standard principal-component approach. Unlike most of the previous literature, the analysis considers both in‡ ows and out‡ ows. Our results speak to the ongoing debate on the quantitative importance of the global …nancial cycle for the observed patterns of capital ‡ ows around the world. They can be summarized in four points. First, among advanced and emerging countries, capital ‡ ows exhibit a considerable degree of commonality. Common factors – speci…cally, a global fac- tor, plus an advanced-country factor and an emerging-country factor for each of in‡ ows and out‡ ows –account, on average, for just under half of the observed variation in gross capital ‡ ows. Second, commonality is particularly strong among advanced countries, where common factors are responsible for the majority (around 60 percent) of the variation. Among emerg- ing countries, just over one-third of the observed variation in capital ‡ ows is attributable to common factors. In other words, common factors dominate advanced-country capital ‡ ows, while local (idiosyncratic) factors dominate emerging-country capital ‡ ows. Third, the di¤er- ence between advanced and emerging countries regarding the contribution of common factors is primarily due to the global factor. Its variance contribution is, on average, over twice as large among the former countries than among the latter. In contrast, the contribution of the group-speci…c factor is roughly similar across the two groups of countries. Fourth, all of the preceding observations apply to both in‡ ows and out‡ ows. In fact, the latent factors for in‡ ows and out‡ ows show a large positive correlation. This is particularly true for the global factors and the advanced-country in‡ ow and out‡ ow group factors. The emerging-country in‡ ow and out‡ ow group factors also show positive, but weaker, correlation. Still, across all country groups, gross out‡ ows are more strongly a¤ected than gross in‡ ows by global factors. In conclusion, international cycles, as summarized by the estimated common factors, are responsible for much of the observed variation of capital ‡ ows, especially among advanced countries. Going one level deeper to explore the forces behind those cycles, we …nd that a good deal of the variation of the factors themselves can be explained by a handful of variables summarizing real and …nancial conditions across the world. In particular, all the rially a¤ected if leading …nancial centers are omitted from the calculation of the averages. 26 factors (global and group-speci…c, for in‡ ows as well as out‡ ows) are robustly negatively correlated with the VIX and similar indices of investor risk aversion. The VIX, along with U.S. interest rates, the U.S. real exchange rate, U.S. real GDP growth, and world commodity prices, account for a substantial share of the variance of the factors – as much as 70 to 80 percent in the case of the global factors, and between 40 and 50 percent in the case of the group factors. This means that, through the common factors, a small set of global variables drives up to half of the variance of advanced-country in‡ ows and out‡ ows, and around one-quarter of the variance of emerging-country in‡ ows and out‡ ows. Our results shed light on the trends in globalization – as measured by the common component of capital ‡ ows – before and after the global crisis. We …nd a cyclical pattern, with an initial stage (’globalization’) in which the common factors gain growing importance up to the crisis, with global factors becoming increasingly dominant, especially in advanced countries – and partly at the expense of group factors. After the crisis, the trend goes in reverse: the role of common factors and, in particular, global factors enters a stage of decline –a ’ deglobalization’phase follows. Exposure to the global cycle varies greatly across countries, and the paper also sheds light on the underlying reasons. Empirical tests …nd a signi…cant role of the structural and policy framework: countries exhibiting a higher degree of …nancial openness and/or greater …nancial depth tend to be impacted more strongly by global cycles – on both the in‡ ow and the out‡ ow side of capital ‡ ows. In addition, the exchange rate regime also matters: pegged exchange rates raise, to an economically signi…cant extent, the sensitivity of gross capital in‡ows to global cycles. Thus, in spite of the wide reach of the global …nancial cycle, the Trilemma still remains an accurate characterization of countries’choice of exchange rate regime, capital account openness, and monetary autonomy. References [1] Adrian, T., D. Stackmand and E. Vogt (2017): "Global price of risk and stabilization policies", unpublished manuscript. [2] Albuquerque, R., N. Loayza and L. Servén (2006): "World market integration through the lens of foreign direct investors", Journal of International Economics 66, 267-295. [3] Avdjiev, S., L. Gambacorta, L. Goldberg and S. Schia¢ (2017a): "The shifting drivers of global liquidity", BIS Working Paper 644. [4] Avdjiev, S., B. Hardy, S. Kalemli-Ozcan and L. Servén (2017b): "Capital in‡ows to banks, corporates, and sovereigns", NBER Working Paper 23116. 27 [5] Andrews, D. and J. Monahan (1992): "An Improved Heteroskedasticity and Autocor- relation Consistent Covariance Matrix Estimator", Econometrica 60, 953-966. [6] Bai, J. (2003): "Inferential theory for factor models of large dimensions", Econometrica 71, 135-71. [7] Bailey, N., G. Kapetanios and H. Pesaran (2015): "Exponent of cross-sectional depen- dence: estimation and inference", Journal of Applied Econometrics Swoosh’and the [8] Bekaert, G., and A. Mehl (2017): “On the global …nancial integration ’ Trilemma” , unpublished manuscript. [9] Bluedorn, J., R. Duttagupta, J. Guajardo and P. Topalova (2013): "Capital ‡ows are …ckle: anytime, anywhere", IMF Working Paper 13/183. [10] Breitung, J. and S. Eickmeier (2016): "Analyzing International Business and Financial Cycles using Multi-Level Factor Models: A Comparison of Alternative Approaches", in E. Hillebrand and S.Koopman (eds.), Advances in Econometrics vol 33. [11] Broner, F., T. Didier, A. Erce, and S. Schmuker (2013): "Gross capital ‡ows: dynamics and crises", Journal of Monetary Economics 60, 113-133. [12] Bruno, V. and H. Shin (2015a): "Capital ‡ows and the risk-taking channel of monetary policy", Journal of Monetary Economics 71, 119-132. [13] Bruno, V. and H. Shin (2015b): "Cross-border banking and global liquidity", Review of Economic Studies 82, 535-564. [14] Byrne, J. and N. Fiess (2011): “International Capital Flows to Emerging and Developing Countries: National and Global Determinants, ” Journal of International Money and Finance 61, 82-100. [15] Calvo, G., L. Leiderman and C. Reinhart (1996): "In‡ ows of capital to developing countries in the 1990s", Journal of Economic Perspectives 10, 123-139. [16] Cerutti, E., S. Claessens, and D. Puy (2017): "Push factors and capital ‡ows to emerg- ing markets: why knowing your lender matters more than fundamentals", unpublished manuscript. [17] Cerutti, E., S. Claessens, and L. Ratnovski (2017): "Global liquidity and cross-border bank ‡ ows", Economic Policy, 81-125. [18] Cerutti, E., S. Claessens, and A. Rose (2017): "How important is the global …nancial cycle? Evidence from capital ‡ ows", CEPR Discussion paper 12075. [19] Choi, I. (2012): "E¢ cient estimation of factor models", Econometric Theory 28, 274- 308. [20] Choi, I. (2017): "E¢ cient estimation of nonstationary factor models", Journal of Sta- tistical Planning and Inference 183, 18-43. 28 [21] Choi, I. and H. Jeong (2017): " Model Selection for Factor Analysis: Some New Criteria and Performance Comparisons" , Econometric Reviews [22] Choi, I., D. Kim, Y. Kim and N. Kwark (2017): "A multilevel factor model: identi…- cation, asymptotic theory and applications", Journal of Applied Econometrics (forth- coming). [23] Davis, J. and E. van Wincoop (2017): "Globalization and the Increasing Correlation between Capital In‡ ows and Out‡ows ", NBER Working Paper 23671. [24] Eichengreen, B., P. Gupta and O. Masetti (2017): "Are capital ‡ows …ckle? Increasingly ? and does the answer still depend on type?", World Bank Policy Research Working Paper 7972. [25] Ergemen, Y. and C. Rodriguez-Caballero (2016): "A dynamic multi-level factor model with long-range dependence", CREATES Research Paper 2016-23. [26] Forbes, C. and F. Warnock (2012): "Capital ‡ ow waves: surges, stops, ‡ight and re- trenchment", Journal of International Economics 88, 235-251. [27] Forbes, K., D. Reinhardt and T. Wieladek (2017): "The spillovers, interactions, and (un)intended consequences of monetary and regulatory policies”, Journal of Monetary Economics 85, 1-22. [28] Fratzscher, M. (2012): "Capital ‡ ows, push versus pull factors, and the global …nancial crisis", Journal of International Economics 88, 341-356. [29] Ghosh, A., J. Ostry and M. Qureshi (2015): "Exchange Rate Management and Crisis Susceptibility: A Reassessment", IMF Economic Review 63, 238-276. [30] Jorda, O., M. Schularick, A. Taylor and F. Ward (2017): “Global …nancial cycles and risk premiums” , unpublished manuscript. [31] Kose, A., C. Otrok and C. Whiteman (2003): "International Business Cycles: World, Region, and Country-Speci…c Factors", American Economic Review 93, 1216-1239. [32] Leamer, E. (1984): Sources of International Comparative Advantage: Theory and Evi- dence, Cambridge, MA: MIT Press. [33] Leamer, E. (1995): The Heckscher-Ohlin Model in Theory and Practice, Princeton Studies in International Finance. [34] Miranda-Agrippino, S. and H. Rey (2015): "World asset markets and the global …nancial cycle", NBER Working Paper 21722. [35] Obstfeld, M., J. Ostry and M. Qureshi (2017): "A Tie That Binds: Revisiting the Trilemma in Emerging Market Economies", IMF Working Paper 17/130. [36] Onatski, A. (2012): " Asymptotics of the principal components estimator of large factor models with weakly in‡ uential factors", Journal of Econometrics 168, 244-258. 29 [37] Raddatz, C. and S. Schmukler (2012): "On the International Transmission of Shocks: Micro-Evidence from Mutual Fund Portfolios," Journal of International Economics 88, 357-374. [38] Reinhart, C., V. Reinhart and C. Trebesch (2017): "Capital ‡ow cycles: a long, global view", unpublished manuscript. [39] Rey, H. (2013): "Dilemma not Trilemma: the global …nancial cycle and monetary policy independence", Proceedings of the Federal Reserve Bank at Kansas City Economic Symposium at Jackson Hole. [40] Rose, A. and T. Wieladek (2014): "Financial protectionism: …rst evidence", Journal of Finance 69, 2127–2149. [41] Wang, P. (2014): "Large dimensional factor models with a multi-level factor structure: identi…cation, estimation and inference", unpublished manuscript. 30 Table 1 Gross capital flows: descriptive statistics CIF COD Correlation Std. Coef. of Std. Coef. of CIF - COD Mean Mean Deviation Variation Deviation Variation All countries 7.71 10.03 1.20 6.69 10.09 1.92 0.47 Advanced 11.99 12.15 0.92 12.03 12.24 0.96 0.90 Emerging 6.16 7.25 1.16 6.07 8.97 1.40 0.48 Developing 6.39 11.19 1.42 3.87 9.73 2.99 0.19 Notes: this table reports group averages of the individual-country statistics (shown in Table A3). CIF = gross inflows; COD = gross outflows. Table 2 Gross capital flows: cross-country comovement (a) Average cross-country correlation CIF COD Flow / Region Advanced Emerging Developing Advanced Emerging Developing Advanced 0.435 CIF Emerging 0.143 0.209 Developing -0.074 0.085 0.109 Advanced 0.434 0.142 -0.087 0.488 COD Emerging 0.290 0.155 -0.009 0.315 0.251 Developing 0.133 0.089 -0.006 0.151 0.142 0.070 (b) Percentage of significant correlations CIF COD Flow / Region Advanced Emerging Developing Advanced Emerging Developing Advanced 0.87 CIF Emerging 0.55 0.58 Developing 0.58 0.53 0.54 Advanced 0.82 0.53 0.60 0.90 COD Emerging 0.77 0.51 0.51 0.81 0.69 Developing 0.54 0.47 0.40 0.59 0.56 0.42 Notes: In panel (a), entry (i,j) is the average of the pairwise correlations between the flows of countries in group i and those of countries in group j. In panel (b), entry (i,j) indicates the percentage of all the pairwise correlations underlying the average shown in panel (a) that are statistically significant at the 95 percent level, with the standard error of a correlation r 2 approximated by the square root of (1 - r )/(T-1). Diagonal entries in each block of the table correspond to within-group correlations (excluding the own-country correlation), off-diagonal elements correspond to between-group correlations. CIF = gross inflows; COD = gross outflows. 31 Table 3 Exponent of cross-sectional dependence Flow Group CIF COD Advanced 0.97 0.99 (0.83, 1.10) (0.88, 1.11) Emerging 0.96 0.95 (0.90, 1.02) (0.84, 1.05) Developing 0.86 0.82 (0.82, 0.89) (0.74, 0.89) Notes: CIF = gross inflows; COD = gross outflows. 95-percent confidence intervals shown in parentheses. Table 4 Correlation of common factors CIF COD Flow / Region Advanced Emerging Advanced Emerging Advanced 1.000 CIF Emerging -0.016 1.000 Advanced 0.821 -0.068 1.000 COD Emerging -0.140 0.343 0.157 1.000 Note: CIF = gross inflows; COD = gross outflows. 32 Table 5 Variance decomposition by group (percent) (a) Gross Inflows All Advanced Emerging countries countries countries Global share 24.1 37.5 15.0 Group share 21.3 22.3 20.6 Country share 54.6 40.2 64.5 (b) Gross Outflows All Advanced Emerging countries countries countries Global share 33.8 46.7 25.1 Group share 14.3 17.2 12.2 Country share 51.9 36.1 62.7 Note: The numbers shown are averages of the individual-country estimates. 33 Table 6 Covariates of the common factors Factors Advanced Advanced Emerging Emerging Variables Global Global countries countries countries countries CIF COD CIF COD CIF COD A. Regressions on risk measures VIX (1990-2015) -0.076 ** -0.057 *** -0.095 ** -0.154 ** -0.099 *** -0.123 *** (0.033) (0.021) (0.042) (0.068) (0.012) (0.027) 2 R 0.222 0.155 0.227 0.274 0.302 0.164 VXO (1986-2015) -0.050 * -0.032 * -0.074 * -0.134 ** -0.073 *** -0.105 * (0.028) (0.019) (0.040) (0.066) (0.020) -(0.054) R2 0.147 0.075 0.212 0.312 0.252 0.183 10-year BAA spread (1979-2015) -0.598 * -0.509 *** -0.713 ** -1.032 * -0.558 *** -0.812 * (0.311) (0.165) (0.352) (0.557) (0.226) (0.428) R2 0.266 0.237 0.245 0.229 0.155 0.131 B. Multivariate regressions VIX -0.059 *** -0.036 *** -0.085 * -0.159 *** -0.084 *** -0.133 *** (0.017) (0.013) (0.045) (0.061) (0.027) (0.045) U.S. short-term real interest rate 0.043 -0.026 0.364 * 0.606 0.243 0.583 (0.089) (0.077) (0.217) (0.380) (0.195) (0.399) Log U.S. real exchange rate 5.413 *** 0.580 -10.943 * -5.108 -11.156 ** -7.639 (1.970) (2.042) (6.353) (7.310) (5.363) (7.866) U.S. real GDP growth 13.280 *** 23.275 *** 17.052 ** -3.715 -2.768 -31.188 ** (3.464) (3.431) (8.464) (12.970) (6.936) (12.317) Log world commodity price index 2.610 *** 1.863 *** -3.867 *** -4.435 ** -1.560 -4.051 ** (0.487) (0.556) (1.481) (1.904) (1.142) (1.807) R2 0.726 0.833 0.495 0.555 0.441 0.457 Notes: All variables except U.S. real GDP growth are expressed in first differences. HAC standard errors in parentheses. All regressions include a constant. *** p<0.01, ** p<0.05, * p<0.1 34 Table 7 Variance decomposition including developing countries (percent) (a) Gross Inflows All Advanced Emerging Developing countries countries countries countries Global share 19.0 31.9 16.1 14.6 Group share 20.0 26.6 19.6 17.0 Country share 61.0 41.5 64.3 68.3 (b) Gross Outflows All Advanced Emerging Developing countries countries countries countries Global share 23.9 39.7 27.2 13.5 Group share 12.9 23.2 11.4 8.8 Country share 63.3 37.1 61.4 77.7 Note: The numbers shown are averages of the individual-country estimates. 35 Table 8 Covariates of the global factor loadings CIF COD Covariates 1 2 3 4 5 1 2 3 4 5 0.251*** 0.139*** 0.274*** 0.206*** Financial openness [0.041] [0.051] [0.044] [0.050] 0.062** -0.046 0.0824** 0.033 Trade openness [0.024] [0.031] [0.036] [0.037] 0.001*** 0.002*** 0.001*** 0.001*** Domestic credit (% of GDP) [0.000] [0.000] [0.000] [0.000] -0.104*** -0.115*** -0.042** -0.042 Exchange Rate arrangement [0.0221] [0.0244] [0.0197] [0.0425] Observations 846 846 846 846 846 846 846 846 846 846 Prob > Chi2 0.000 0.010 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 Note: Panel GLS-AR(1) regressions. The dependent variable is the global factor loadings estimates over moving 20-year windows, and a constant is included. Standard errors in brackets. The explanatory variables are averages over the respective window. Financial openness is the Chinn-Ito index, trade openness is the log of total trade over GDP, the exchange rate arrangement is measured using the De Facto aggregate classification of Ghosh, Ostry and Qureshi (2015), setting Peg = 1, Intermediate = 2, and Float = 3 . *** p<0.01, ** p<0.05, * p<0.1 36 Table A1 Country List Advanced countries Emerging countries Developing countries Australia Argentina Albania Paraguay Austria Brazil Bangladesh Rwanda Canada Chile Benin Sierra Leone Denmark China Bolivia Sri Lanka Finland Colombia Botswana Sudan France Cyprus Bulgaria Swaziland Germany Egypt Cameroon Tanzania Ireland India Costa Rica Trinidad and Tobago Italy Israel Dominican Rep. Tunisia Japan Jordan Ecuador Uganda Netherlands Korea El Salvador New Zealand Kuwait Ethiopia Norway Malaysia Fiji Portugal Mexico Ghana Spain Morocco Guatemala Sweden Oman Haiti Switzerland Pakistan Honduras United Kingdom Peru Jamaica United States Philippines Lesotho Poland Madagascar Romania Malawi Saudi Arabia Mauritius Singapore Myanmar South Africa Nepal Thailand Nicaragua Turkey Nigeria Uruguay Panama Venezuela Papua New Guinea 37 Table A2 Data Sources # Series Description Source Coverage 1 Capital Flows Gross Asset and Liability Flows IMF BOP 1970-2015 UN National 2 Nominal GDP in U.S. dollars Nominal GDP in U.S. dollars 1960-2015 Accounts 3 VIX CBOE Volatility Index FRED 1990-2015 4 VXO CBOE S&P 100 Volatility Index FRED 1986-2015 Moody's Baa Corporate rate minus Treasury 5 10-year BAA spread FRED 1970-2015 10Y rate 6 U.S. short-term real interest rate Effective Fed Funds Rate minus GDP inflation FRED 1970-2015 Gross Domestic Product in constant US 7 U.S. real GDP growth IMF IFS 1970-2015 dollars 8 U.S. real exchange rate Real Effective Exchange Rate REER IMF IFS 1979-2015 Commodity Price Index - Metals and 9 World commodity price index UNCTAD 1960-2015 Minerals Chinn-Ito Index of Capital Account 10 Financial openness Chinn-Ito 1960-2014 Liberalization UN National 11 Trade openness (% of GDP) Total Exports plus Imports over GDP 1960-2015 Accounts 12 Domestic credit (% of GDP) Domestic credit to private sector over GDP WB WDI 1960-2015 Gosh, Ostry and 13 Exchange Rate regime De Facto Aggregate ERA 1980-2011 Qureshi (2015) Net Exports of Commodities over GDP 14 Commodity intensity Comtrade/UNCTAD 1960-2015 (Leamer index) 38 Table A3 Descriptive statistics by country CIF COD Correlation Countries Std. Coef. of Std. Coef. of Mean Mean CIF & COD Deviation Variation Deviation Variation Advanced Australia 8.16 3.33 0.41 3.97 2.76 0.69 0.93 Austria 9.23 11.75 1.27 9.34 11.95 1.28 0.98 Canada 7.08 2.64 0.37 5.43 3.00 0.55 0.69 Denmark 7.57 8.23 1.09 8.83 8.14 0.92 0.89 Finland 10.36 10.22 0.99 10.77 11.34 1.05 0.89 France 8.46 7.51 0.89 8.50 7.73 0.91 0.98 Germany 6.54 5.88 0.90 8.45 5.79 0.69 0.88 Ireland 60.23 69.83 1.16 57.47 68.58 1.19 1.00 Italy 5.44 4.01 0.74 4.56 3.93 0.86 0.90 Japan 2.33 2.49 1.07 4.53 2.65 0.59 0.84 Netherlands 23.88 26.35 1.10 27.29 26.87 0.98 0.99 New Zealand 6.06 3.94 0.65 1.76 3.32 1.89 0.69 Norway 8.22 9.70 1.18 13.19 13.08 0.99 0.89 Portugal 10.28 9.48 0.92 6.81 7.00 1.03 0.89 Spain 8.83 7.74 0.88 6.42 6.08 0.95 0.91 Sweden 9.48 7.32 0.77 10.46 8.47 0.81 0.90 Switzerland 12.11 17.32 1.43 20.95 19.49 0.93 0.97 United Kingdom 18.15 19.15 1.06 16.95 19.73 1.16 0.99 United States 5.78 3.64 0.63 3.49 2.63 0.75 0.88 Emerging Argentina 3.66 3.15 0.86 1.96 2.68 1.37 0.56 Brazil 4.04 2.75 0.68 1.77 2.34 1.33 0.48 Chile 9.21 5.39 0.59 6.02 5.42 0.90 0.65 China 3.63 2.55 0.70 4.97 4.40 0.88 0.75 Colombia 4.41 2.90 0.66 2.18 2.05 0.94 0.50 Cyprus 21.84 51.33 2.35 17.18 49.63 2.89 1.00 Egypt 3.55 5.72 1.61 2.93 4.47 1.53 -0.09 India 3.15 2.07 0.66 1.77 1.99 1.13 0.80 Israel 5.88 4.47 0.76 5.70 4.51 0.79 0.57 Jordan 9.01 8.36 0.93 6.46 7.89 1.22 0.66 Korea 3.68 4.08 1.11 4.05 3.32 0.82 0.19 Kuwait 1.57 7.84 4.98 20.90 45.51 2.18 0.51 Malaysia 6.21 6.21 1.00 7.33 7.58 1.03 0.20 Mexico 4.59 3.17 0.69 1.81 2.04 1.13 0.37 Morocco 4.43 3.29 0.74 1.74 2.52 1.45 0.06 Oman 3.79 4.17 1.10 6.57 9.48 1.44 0.41 Pakistan 2.84 2.20 0.78 0.75 1.81 2.42 0.04 Peru 6.06 3.56 0.59 2.74 3.56 1.30 0.51 Philippines 4.36 4.26 0.98 2.71 3.71 1.37 0.38 39 Table A3 (continued) Descriptive statistics by country CIF COD Correlation Countries Std. Coef. of Std. Coef. of Mean Mean CIF & COD Deviation Variation Deviation Variation Emerging (continued) Poland 4.89 3.64 0.74 2.35 2.38 1.01 0.49 Romania 4.29 6.51 1.52 1.64 2.63 1.60 0.58 Saudi Arabia 2.07 2.86 1.38 5.13 15.18 2.96 0.33 Singapore 32.21 35.77 1.11 43.77 40.69 0.93 0.97 South Africa 3.51 4.18 1.19 2.39 2.45 1.03 0.79 Thailand 5.03 5.85 1.16 3.96 4.24 1.07 0.41 Turkey 3.76 3.40 0.90 1.44 1.47 1.02 0.57 Uruguay 6.19 5.39 0.87 3.76 6.28 1.67 0.70 Venezuela 2.45 3.32 1.35 4.95 5.68 1.15 0.20 Developing Albania 6.19 6.66 1.08 3.64 3.78 1.04 0.50 Bangladesh 2.50 1.16 0.46 1.61 1.40 0.87 -0.08 Benin 4.13 6.64 1.61 1.32 4.09 3.09 0.31 Bolivia 7.26 5.86 0.81 3.81 5.19 1.36 -0.14 Botswana 6.75 5.81 0.86 12.77 11.97 0.94 0.06 Bulgaria 7.14 12.12 1.70 3.63 4.96 1.37 0.57 Cameroon 4.07 3.31 0.81 0.92 2.45 2.67 0.26 Costa Rica 7.15 4.77 0.67 3.02 2.38 0.79 0.60 Dominican Rep. 4.75 2.75 0.58 1.10 2.14 1.95 0.27 Ecuador 3.23 3.30 1.02 1.84 2.39 1.30 -0.06 El Salvador 5.29 6.50 1.23 0.95 3.41 3.58 0.26 Ethiopia 3.73 2.02 0.54 -0.67 2.85 -4.25 -0.17 Fiji 4.99 6.86 1.37 0.73 3.87 5.33 0.41 Ghana 5.35 4.17 0.78 0.57 1.92 3.39 0.14 Guatemala 4.60 2.29 0.50 0.78 2.43 3.13 0.66 Haiti 2.41 3.52 1.46 1.06 3.49 3.29 -0.46 Honduras 6.34 3.71 0.58 1.71 2.54 1.49 0.24 Jamaica 10.50 8.26 0.79 3.96 5.51 1.39 0.79 Lesotho 11.80 12.53 1.06 14.18 10.49 0.74 0.05 Madagascar 7.61 5.74 0.75 0.95 2.27 2.40 0.39 Malawi 4.01 9.85 2.46 0.17 2.29 13.62 0.10 Mauritius 40.95 116.95 2.86 39.16 116.12 2.97 1.00 Myanmar 3.78 2.42 0.64 1.31 2.64 2.02 -0.08 Nepal 3.05 1.84 0.60 2.37 4.36 1.84 -0.32 Nicaragua 7.00 19.62 2.80 1.51 3.04 2.02 -0.02 Nigeria 2.38 3.82 1.60 4.16 4.10 0.99 0.09 Panama 10.67 73.39 6.88 6.00 70.96 11.83 1.00 Papua New Guinea 1.23 5.23 4.24 1.39 5.57 4.01 0.07 Paraguay 2.51 3.79 1.51 0.42 3.19 7.60 -0.01 40 Table A3 (continued) Descriptive statistics by country CIF COD Correlation Countries Std. Coef. of Std. Coef. of Mean Mean CIF & COD Deviation Variation Deviation Variation Developing (continued) Rwanda 2.66 6.94 2.61 1.07 2.59 2.41 -0.13 Sierra Leone 7.28 10.11 1.39 0.77 2.78 3.63 -0.41 Sri Lanka 5.16 2.58 0.50 1.10 2.37 2.15 0.33 Sudan 5.13 3.16 0.62 1.03 1.49 1.45 0.41 Swaziland 4.88 3.78 0.78 4.03 8.01 1.99 0.07 Tanzania 4.80 4.98 1.04 0.87 1.36 1.56 -0.06 Trinidad and Tobago -1.18 8.44 -7.14 3.09 7.95 2.57 -0.28 Tunisia 6.34 2.21 0.35 2.59 2.16 0.83 0.30 Uganda 3.54 5.12 1.45 0.99 1.76 1.78 0.18 41 Table A4 Information criteria for selecting the number of factors Global Factors by CIF COD factors Group ICP2 BIC HQ ICP2 BIC HQ 1 1 3.8 7.0 7.1 3.7 7.0 7.1 1 2 6.2 7.5 7.6 6.1 7.5 7.6 1 3 8.6 7.9 8.0 8.5 7.8 7.9 2 1 5.4 7.4 7.5 5.3 7.4 7.5 2 2 7.8 7.8 7.9 7.8 7.7 7.8 2 3 10.2 8.1 8.2 10.2 8.0 8.1 3 1 7.1 7.7 7.8 7.0 7.6 7.7 3 2 9.5 8.0 8.1 9.4 7.9 8.0 3 3 11.9 8.2 8.3 11.9 8.2 8.3 Note: CIF = gross inflows; COD = gross outflows. Table A5 Panel unit root and stationarity tests CIF COD Test Null Hypothesis P-Val P-Val Im-Pesaran-Shin Ho: All panels contain unit roots 0.00 0.00 Hadri LM Ho: All panels are stationary 0.35 0.61 Notes: the specification includes two lags and no time trend. CIF = gross inflows; COD = gross outflows. 42 Table A6 Variance decomposition by group, excluding global financial centers (percent) (a) Gross Inflows All Advanced Emerging countries countries countries Global share 23.1 38.1 15.0 Group share 20.3 19.9 20.6 Country share 56.6 42.0 64.5 (b) Gross Outflows All Advanced Emerging countries countries countries Global share 33.2 48.5 25.1 Group share 13.3 15.4 12.2 Country share 53.4 36.1 62.7 Note: The numbers shown are averages of the individual-country estimates, computed excluding the U.S., U.K., Germany, and Japan. 43 Figure 1 Gross capital flows by country group (percent of trend GDP) (a) All countries 25 25 CIF COD 20 20 15 15 10 10 5 5 0 0 1978 1983 1988 1993 1998 2003 2008 2013 2018 year (a) Advanced countries 30 30 CIF COD 25 25 20 20 15 15 10 10 5 5 0 0 1978 1983 1988 1993 1998 2003 2008 2013 2018 year 44 Figure 1 (continued) Gross capital flows by country group (percent of trend GDP) (b) Emerging countries 15 15 CIF COD 12 12 9 9 6 6 3 3 0 0 -3 -3 1978 1983 1988 1993 1998 2003 2008 2013 2018 year (c) Developing countries 12 12 CIF COD 9 9 6 6 3 3 0 0 -3 -3 -6 -6 1978 1983 1988 1993 1998 2003 2008 2013 2018 year 45 Figure 2 Global factors 4 4 CIF COD 3 3 2 2 1 1 0 0 -1 -1 -2 -2 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 Year 46 Figure 3 Group factors (a) Advanced countries 4 4 CIF COD 3 3 2 2 1 1 0 0 -1 -1 -2 -2 -3 -3 -4 -4 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 Year (b) Emerging countries 4 4 CIF COD 2 2 0 0 -2 -2 -4 -4 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 Year 47 0.0 0.2 0.4 0.6 0.8 1.0 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0 -0.6 -0.4 -0.2 New Zealand New Zealand Portugal Switzerland Switzerland Portugal United Kingdom Japan Japan Canada Italy Finland Finland Denmark Spain United Kingdom United States Germany Australia Italy Austria Austria France Sweden Denmark Spain Ireland France Canada Ireland Sweden Australia Germany Norway Netherlands United States Norway Netherlands Pakistan Philippines Uruguay Thailand Egypt Malaysia Jordan Argentina 48 Mexico Mexico Cyprus Venezuela Figure 4 Morocco Korea (a) Gross Inflows Philippines Morocco (b) Gross Outflows Global factor loadings Peru Peru Argentina Uruguay Thailand Pakistan Kuwait Egypt Colombia Israel Poland Chile Brazil Cyprus Venezuela Brazil Korea Saudi Arabia Singapore Colombia Malaysia Jordan Romania Kuwait South Africa Singapore Turkey China Saudi Arabia Oman Oman Romania Israel Poland Chile South Africa India India China Turkey 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 -0.6 -0.4 -0.2 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0 -0.6 -0.4 -0.2 Australia Canada Norway Japan Japan Australia Finland Finland Canada Norway Netherlands Netherlands Denmark Ireland Sweden Sweden Austria United States Ireland New Zealand Germany Spain France Portugal New Zealand France Portugal Denmark Spain Switzerland United States Germany Switzerland Italy Italy Austria United Kingdom United Kingdom Kuwait Cyprus Oman Kuwait Saudi Arabia Jordan Turkey Oman 49 Jordan China Venezuela Poland Figure 5 Cyprus Turkey (a) Gross Inflows Romania Saudi Arabia (b) Gross Outflows Group factor loadings Morocco Romania South Africa India Pakistan Egypt Israel Venezuela Malaysia Pakistan China South Africa Singapore Peru Egypt Morocco India Singapore Uruguay Thailand Mexico Uruguay Argentina Israel Chile Argentina Colombia Malaysia Poland Colombia Peru Mexico Brazil Chile Korea Brazil Philippines Korea Thailand Philippines 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 -0.6 -0.4 -0.2 -0.6 -0.4 -0.2 Figure 6 Variance decomposition by country (a) Gross Inflows 0.00 0.20 0.40 0.60 0.80 1.00 New Zealand Switzerland Portugal Japan Canada Finland Denmark United Kingdom Germany Italy Austria Sweden Spain France Ireland Australia Norway United States Netherlands Morocco Peru Korea Uruguay Venezuela Pakistan Mexico Egypt Israel Argentina Malaysia Thailand Chile Cyprus Brazil Philippines Saudi Arabia Colombia Jordan Kuwait Singapore China Oman Romania Poland South Africa India Turkey 0.00 0.20 0.40 0.60 0.80 1.00 Global share Group share Country share 50 Figure 6 (continued) Variance decomposition by country (b) Gross outflows 0.00 0.20 0.40 0.60 0.80 1.00 New Zealand Portugal Switzerland United Kingdom Japan Italy Finland Spain United States Australia Austria France Denmark Ireland Canada Sweden Germany Netherlands Norway Pakistan Uruguay Egypt Jordan Mexico Cyprus Morocco Philippines Peru Argentina Thailand Kuwait Colombia Poland Brazil Venezuela Korea Singapore Malaysia Romania South Africa Turkey Saudi Arabia Oman Israel Chile India China 0.00 0.20 0.40 0.60 0.80 1.00 Global share Group share Country share 51 Figure 7 (a) Gross inflows: variance decomposition over time (i) Advanced countries 1.0 1.0 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share (ii) Emerging countries 1.0 1.0 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share 52 Figure 7 (continued) (b) Gross outflows: variance decomposition over time (i) Advanced countries 1.0 1.0 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share (ii) Emerging countries 1.0 1.0 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share 53 Figure 8 Variance decomposition over time, by degree of financial openness (a) Gross inflows Low financial openness High financial openness 1.0 1.0 1.0 1.0 0.9 0.9 0.9 0.9 0.8 0.8 0.8 0.8 0.7 0.7 0.7 0.7 0.6 0.6 0.6 0.6 0.5 0.5 0.5 0.5 0.4 0.4 0.4 0.4 0.3 0.3 0.3 0.3 0.2 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share Global share Group share Country share (b) Gross outflows Low financial openness High financial openness 1.0 1.0 1.0 1.0 0.9 0.9 0.9 0.9 0.8 0.8 0.8 0.8 0.7 0.7 0.7 0.7 0.6 0.6 0.6 0.6 0.5 0.5 0.5 0.5 0.4 0.4 0.4 0.4 0.3 0.3 0.3 0.3 0.2 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share Global share Group share Country share 54 Figure 9 Variance decomposition over time, by exchange rate regime (a) Gross inflows Peg regime Floating regime 1.0 1.0 1.0 1.0 0.9 0.9 0.9 0.9 0.8 0.8 0.8 0.8 0.7 0.7 0.7 0.7 0.6 0.6 0.6 0.6 0.5 0.5 0.5 0.5 0.4 0.4 0.4 0.4 0.3 0.3 0.3 0.3 0.2 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share Global share Group share Country share (b) Gross outflows Peg regime Floating regime 1.0 1.0 1.0 1.0 0.9 0.9 0.9 0.9 0.8 0.8 0.8 0.8 0.7 0.7 0.7 0.7 0.6 0.6 0.6 0.6 0.5 0.5 0.5 0.5 0.4 0.4 0.4 0.4 0.3 0.3 0.3 0.3 0.2 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share Global share Group share Country share 55 Figure 10 Variance decomposition over time, by degree of financial depth (a) Gross inflows (i) Low financial depth (ii) High financial depth 1.0 1.0 1.0 1.0 0.9 0.9 0.9 0.9 0.8 0.8 0.8 0.8 0.7 0.7 0.7 0.7 0.6 0.6 0.6 0.6 0.5 0.5 0.5 0.5 0.4 0.4 0.4 0.4 0.3 0.3 0.3 0.3 0.2 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share Global share Group share Country share (b) Gross outflows (i) Low financial depth (ii) High financial depth 1.0 1.0 1.0 1.0 0.9 0.9 0.9 0.9 0.8 0.8 0.8 0.8 0.7 0.7 0.7 0.7 0.6 0.6 0.6 0.6 0.5 0.5 0.5 0.5 0.4 0.4 0.4 0.4 0.3 0.3 0.3 0.3 0.2 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 1998 2000 2002 2004 2006 2008 2010 2012 2014 1998 2000 2002 2004 2006 2008 2010 2012 2014 Global share Group share Country share Global share Group share Country share 56 Figure 11 Developing-country group factors 4 4 CIF 2 COD 2 0 0 -2 -2 -4 -4 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 Year 57