Data Science and Service Research Discussion Paper

Transcription

1 Discussion Paper No. 69 A Network Analysis of International Financial Flows Hongwei Chuang and Navruzbek Karamatov Graduate School of Economics and Management, Tohoku University August, 2017 Data Science and Service Research Discussion Paper Center for Data Science and Service Research Graduate School of Economic and Management Tohoku University 27-1 Kawauchi, Aobaku Sendai , JAPAN

2 A Network Analysis of International Financial Flows Abstract Relying on the IMF Coordinated Portfolio Investment Survey, which reports countries bilateral investments in financial assets at end-2001 to end-2015, this paper shows that a country s stock market growth is not only spatially correlated with those of nearby countries, but also positively associated with the magnitude of connectedness of the country s international investments in debt within a dynamic financial investment flow network. The positive relation arises because debts have become an increasingly important source of capital for developing countries. Key words: financial network; international capital flows; spatial panel regression JEL classification: C13, C21, G11 1

3 1. Introduction Despite the advantages of international diversification of equity portfolios and the continuing existence of a strong domestic bias of investors in international equity investment, few fundamental questions about international financial flows are being discussed. What is the geography of international financial flows? How do international financial flows behave? Do international financial flows affect stock markets? This study provides answers to these questions from the perspective of network analysis. In the literature, Tesar and Werner (1995), Bohn and Tesar (1996), and Brennan and Cao (1997) have examined estimates of aggregate international portfolio flows and concluded that international financial inflows are positively and contemporaneously correlated with equity returns. In particular, Bohn and Tesar (1996) found evidence that international financial inflows were positively correlated with lagged flows, and with contemporaneous and lagged measures of expected equity returns. Moreover, Froot, O Connell, and Seasholes (2001) used daily custodian bank data and also found that international financial inflows have positive forecasting power for future equity returns, especially in emerging markets. However, most of existing studies are quite narrowly focused on a single source country. To provide a systemic analysis in understanding the process of such global financial integration, we investigate whether international financial flows in equities and debts have impacts on a country s stock market growth. We use the IMF Coordinated Portfolio Investment Survey (CPIS) data to construct dynamic international financial flow networks and examine the effects of networks on stock market growth for three categories of investment flows equity, long-term, and short-term debts. Our contribution is to measure a country s connectedness by calculating its centrality, i.e. its ability to capture another country s international attention. If a country is highly connected with others, its centrality measures will tend to be higher which means this country highly attracts international investors attention. We use a general spatial panel regression to test the relation between a country s stock market growth and its magnitude of connectedness in a network. We find that a country s stock market growth is not only spatially correlated with those of neighboring countries, but also positively associated with its centrality in a 2

4 dynamic international financial flow network. This positive relation remains even if we control for log GDP, log housing index, unemployment rate, and characteristics of the market itself such as local market excess return, earnings-to-price, and book-to-market ratios. 2. Dynamic International Portfolio Investment Networks To construct a dynamic international flow network, the ideal data would be a monthly world matrix of flows in which each element details the flow from one country to another. However, the task of constructing such a matrix is extraordinarily difficult. Fortunately, in 1997, the IMF started to do the survey work and constructed the CPIS by collecting information on annual international financial flows. Initially, the CPIS covered only 29 reporting countries, but this number has kept increasing and reached 87 reporting countries to cover more than 200 destination countries around the world in The CPIS reports the portfolio positions of international investors, excluding the official holdings of monetary authorities, disaggregated by regions and instruments. More specifically, the CPIS data set provides a geographical breakdown of international portfolio holdings disaggregated by three instruments equity, long-term, and short-term debts. Importantly, the CPIS includes virtually all major international investment, excluding foreign direct investment. The advantage of the CPIS data is their consistence, since all participants undertake a benchmark portfolio asset survey at the same time, follow the same definitions and classifications as in the fifth edition of the IMF Balance of Payments Manual, and provide a breakdown of their stock of portfolio investment assets by the country of residency of all non-resident issuers. Of course, we cannot deny that the CPIS data have their shortfalls, as argued by Lane and Milesi-Ferretti (2008); however, the CPIS data still provide a unique perspective on cross-country equity positions that warrants a detailed analysis. We assume that transactions occur uniformly over time and denote the international portfolio flows by p ckt, which represents the investment from country c to the receiving country k in year t. t runs from 2001 to In this study, we define three international flow networks: the first is based on international equity flows N e t =(V,E), the second uses international long-term debt flows N ld t for international short-term debt flows N sd t =(V,LD), and the final one is =(V,SD). Here V denotes a set of countries (all reporting and 3

5 destination countries) and E represents a set of directed edges based on equity, while LD is for long-term debt and SD is for short-term debt. The geographies of the dynamics of international equity flows N2001, e N2008, e and N2015 e are shown in the top, middle, and bottom panels of Figure 1, respectively. From these, we can see a dramatic change in the pattern of world equity flows taking place in emerging markets over the past two decades. International investors pay attention to developing countries such as Brazil, China, and Russia. Moreover, the euro-area countries are a highly connected area because of the monetary union of European Union members. <Insert Figure 1 here> We next present the dynamics of international long-term debt flows N ld 2001, N ld 2008, and N ld 2015 and of international short-term debt flows N sd 2001, N sd 2008, and N sd 2015 in Figure 2 and Figure 3, respectively. The geographical patterns of international long- and short-term debt flows are quite similar to those of international equity flows, again showing a sizable change over the past 15 years. One thing worth noticing is that the scale of international long-term debt flows is larger than the scale of international short-term debt flows. <Insert Figure 2 and Figure 3 here> Having constructed the international financial flow networks, it is natural to examine a country s importance within these by using a node s centrality as a representation. In network theory, for a given graph G := (V,E) with V number of vertices and A =(a ij ) be the adjacency matrix, i.e., 1 vertex i is linked to vertex j a ij = 0 otherwise i, j, (1) three centrality measures are commonly used: degree centrality, betweenness centrality, and eigenvector centrality. Degree centrality is defined as the number of links incident upon node i, betweenness centrality considers the number of shortest paths from all vertices to all others that pass through i, and eigenvector centrality 4

6 calculates the relative centrality score of vertex i as x i = 1 x t = 1 a it x t, (2) κ κ t M(i) t G where M(i) is a set of neighbors of i and κ is a non-negative eigenvalue. Usually, eigenvectors of the adjacency matrices are used for the measure of centrality, but many different eigenvalues κ exist for an eigenvector solution. The multiple solutions problem can be solved by selecting the largest eigenvalue upon the power iteration eigenvalue algorithm to find the dominant eigenvector. The purpose of the three represented centrality measures is to capture different aspects of investor attention of a country for a network. For example, a country with high betweenness centrality has a large influence on the transfer of items through the network under the assumption that item transfers follow the shortest path. Eigenvector centrality assigns relative scores to all nodes in the network based on the concept that connections to high-scoring nodes contribute more to the score of the node in question than equal connections to low-scoring nodes. 3. Spatial Panel Regression and Empirical Results In this paper, we consider a general spatial panel regression that allows for an endogenous spatial lag of the dependent variable and the disturbances to be correlated over time and across spatial units. We assume that in each year t =1,...,T the data are generated according to the following model: N y it = λ w ij y jt + x it β + α + u it, (3) j=1 where index i = 1,...,N denotes the countries, y it is S&P Global index change for country i in year t, Nj=1 w ij y jt denotes the spatial lag of the dependent variable with w ij being observable non-stochastic spatial weights. x it is a 1 K 1 vector of exogenous variables, β is a K 1 1 parameter vector, λ is a scalar parameter and α refers to the constant. u it is the overall disturbance term. As discussed in Kapoor, Kelejian, and Prucha (2007), we allow the disturbances follow a Cliff and Ord 5

7 type spatial autocorrelation, that is, N u it = ρ m ij u jt + it, (4) j=1 it = µ i + v it, (5) where ρ is a scalar parameter and m ij are observable spatial weights (the same as the weights w ij of Equation (3) in this study). The innovations it follow the one-war error component structure in which the v it are independent innovations and µ i are individual effects, which can be either fixed or random. Our x it not only includes a country s centrality but also a number of control variables market excess return (MKT), earnings-to-price (E/P), and book-to-market (B/M) ratios 1 proposed by Fama and French (1996) and country-specific macroeconomic variables logarithm of GDP, logarithm of housing price index (HPI), and unemployment rate. 2 Due to the availability of Fama and French (1996) s multi-factor data, the regression can only consider these 21 countries Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Hong Kong, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, Singapore, Spain, Sweden, Switzerland, the UK, and the US. The corresponding vector of regression parameters is estimated by an instrumental variable estimation as proposed by Mutl and Pfaffermayr (2011). In running the spatial panel regression, the first task is to determine the spatial weights matrix, W = (w ij ). In this study, we consider three weights: spatial contiguity weight and k-nearest neighbor weights (k =1and 2). The spatial contiguity weight is defined as whether two countries share a boundary or not. If country i and country j share the same boundary, w ij =1. Otherwise, w ij =0. The k-nearest neighbor weights is to let centroid distances from each country i to all country j = i be ranked as follows: d ij(1) d ij(2)... d ij(n 1). Then for each k =1,...,N 1, the set N k = j(1),j(2),...,j(k) contains the k closest 1 We downloaded the international research returns data formed by country portfolios on E/P and B/M ratios from Kenneth R. French s data library. 2 The macroeconomic variables are mostly from the OECD databank ( but the HPIs for Singapore and Hong Kong are obtained from the Singapore statistics center ( id=9e4a9c73-0e9d-44a5-8d24-dc39feaa5730) and the Hong Kong University ( respectively. All the macroeconomic variables are for the base period of

8 countries to i. For each given k, the k-nearest neighbor weight matrix has spatial weights of the form of 1 j N k (i) w ij = 0 otherwise i, j. (6) To show their difference, we plot the geographies of the three kinds of weights in Figure 4. <Insert Figure 4 here> Our fitted results for international equity flows are in Table 1. We present the coefficient estimations for each spatial weights matrix and consider the different centrality measures of eigenvector, degree, and betweenness in Model I to Model III, respectively. <Insert Table 1 here> From the Table 1, we find that the coefficients of λ are all positively significant, which means the spatial effect is an important factor affecting a country s stock market growth. As the geographical closeness in w ij increases, the estimated coefficients increase as well. Local market excess return is also positively correlated with a country s stock market growth, but the housing price index and unemployment are always significantly negatively correlated. The parameters of interest in Table 1 are the coefficients of centrality. For international equity flows, a country s connectedness within the financial network only has a slightly positive impact on stock market growth for the betweenness centrality measure. We further examine the results for international long-term and short-term debt flow networks in Table 2 and Table 3, respectively. The empirical findings in these tables are similar to those in Table 1. In particular, they indicate the importance of a country s centrality within the networks of international long-term and short-term debt flows. We find that eigenvector centrality plays an important role in these networks and has a positive impact on stock market growth. When more and more investors pay attention to a country s debt market, the stock market will also be positively affected. We argue that international capital flows from debt 7

9 markets to stock markets cause stock market growth, presumably due to the spillover effect. <Insert Table 2 and Table 3 here> 4. Concluding Remarks This paper uses a network model to provide a systemic analysis of the factors driving stock market growth across countries. As Bekaert and Harvey (2000) have argued, a number of developing economies initiated reforms to liberalize their capital markets and these reforms made it easier for foreign investors to access the local market and for domestic investors to diversify their portfolios internationally. Over the last couple of decades, these financial reforms have increased an important source of capital for those developing countries. Our findings agree with and reinforce their arguments, as these countries became highly connected with other countries in terms of international financial flows in equities and debts. When we examine the centralities of international long-term and short-term debt flow networks, the positive association between stock market growth and the magnitude of a country s connectedness remains significant even if we control for log GDP, log housing price index, unemployment rate, and characteristics of the market itself such as local market excess return and earnings-to-price and book-to-market ratios. Our research sheds light on the study of international portfolio investment from the perspective of network analysis. 8

10 References Bekaert, G., and Harvey, C. R. (2000). Capital flows and the behavior of emerging market equity returns. In S. Edwards (ed.), Capital Flows and the Emerging Economies: Theory, Evidence, and Controversies, University of Chicago Press, Bohn, H., and Tesar, L. L. (1996). US equity investment in foreign markets: portfolio rebalancing or return chasing?. American Economic Review, 86(2), Brennan, M. J., and Cao, H. H. (1997). International portfolio investment flows. Journal of Finance, 52(5), Fama, E. F., and French, K. R. (1996). Multifactor explanations of asset pricing anomalies. Journal of Finance, 51(1), Froot, K. A., O Connell, P. G., and Seasholes, M. S. (2001). The portfolio flows of international investors. Journal of Financial Economics, 59(2), Kapoor, M., Kelejian, H. H., and Prucha, I. R. (2007). Panel data models with spatially correlated error components. Journal of Econometrics, 140(1), Lane, P. R., and Milesi-Ferretti, G. M. (2008). International investment patterns. Review of Economics and Statistics, 90(3), Mutl J. and Pfaffermayr M. (2011). The Hausman test in a Cliff and Ord panel model. Econometrics Journal, 14, Tesar, L. L., and Werner, I. M. (1995). Home bias and high turnover. Journal of International Money and Finance, 14(4),

11 10 Fig. 1. Dynamics of International Equity Flows 2001 (Top), 2008 (Middle), and 2015 (Bottom)

12 11 Fig. 2. Dynamics of International Long-term Debt Flows 2001 (Top), 2008 (Middle), and 2015 (Bottom)

13 12

14 Fig. 4. Spatial Weights Matrix, W ij,n 13

15 Table 1 Results of Spatial Panel Regressions International Equity Flows This table reports the coefficient estimates of the regression model in Equation (3), which regresses the changes in the S&P Global index on the market excess return (MKT), the earnings-to-price ratio (E/P), the book-to-market ratio (B/M), the logarithm of GDP, the logarithm of the housing price index (HPI), and the unemployment rate. The centralities are calculated using the international equity flow networks. Model I to Model III represent the results for eigenvector, degree, and betweenness centralities, respectively. The standard errors are given in parentheses. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. 14 W (Contiguity) W (K=1 Nearest Neighbors) W (K=2 Nearest Neighbors) Model I Model II Model III Model I Model II Model III Model I Model II Model III λ ** ** *** *** *** *** *** *** *** (0.045) (0.045) (0.045) (0.036) (0.036) (0.036) (0.043) (0.042) (0.042) MKT *** *** *** *** *** *** *** *** *** (0.077) (0.077) (0.075) (0.082) (0.082) (0.080) (0.077) (0.077) (0.075) E/P (0.073) (0.072) (0.072) (0.073) (0.072) (0.072) (0.067) (0.067) (0.066) B/M ** ** *** (0.065) (0.065) (0.065) (0.066) (0.067) (0.066) (0.059) (0.060) (0.059) log(gdp) (0.798) (0.788) (0.793) (0.808) (0.796) (0.793) (0.754) (0.748) (0.750) log(hpi) ** *** *** *** *** *** *** *** *** (3.841) (3.848) (3.755) (3.816) (3.838) (3.711) (3.566) (3.609) (3.534) Unemployment *** *** *** *** *** *** *** *** *** (0.303) (0.303) (0.302) (0.303) (0.300) (0.297) (0.290) (0.288) (0.287) Eigenvector (91.209) (90.495) (89.181) Degree (0.038) (0.038) (0.036) Betweenness ** * (0.003) (0.003) (0.002)

16 Table 2 Results of Spatial Panel Regressions International Long-term Debt Flows This table reports the coefficient estimates of the regression model in Equation (3), which regresses the changes in the S&P Global index on the market excess return (MKT), the earnings-to-price ratio (E/P), the book-to-market ratio (B/M), the logarithm of GDP, the logarithm of the housing price index (HPI), and the unemployment rate. The centralities are calculated using the international long-term debt flow networks. Model I to Model III represent the results for eigenvector, degree, and betweenness centralities, respectively. The standard errors are given in parentheses. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. 15 W (Contiguity) W (K=1 Nearest Neighbors) W (K=2 Nearest Neighbors) Model I Model II Model III Model I Model II Model III Model I Model II Model III λ ** *** *** *** *** *** *** *** *** (0.044) (0.045) (0.045) (0.035) (0.037) (0.036) (0.042) (0.042) (0.042) MKT *** *** *** *** *** *** *** *** *** (0.074) (0.077) (0.076) (0.079) (0.082) (0.080) (0.074) (0.077) (0.075) E/P (0.071) (0.073) (0.072) (0.072) (0.073) (0.072) (0.066) (0.067) (0.067) B/M ** ** ** (0.064) (0.066) (0.065) (0.066) (0.066) (0.066) (0.059) (0.060) (0.060) log(gdp) (0.779) (0.794) (0.792) (0.796) (0.798) (0.796) (0.742) (0.750) (0.752) log(hpi) *** *** *** *** *** *** *** *** *** (3.707) (3.868) (3.775) (3.711) (3.847) (3.754) (3.502) (3.621) (3.595) Unemployment *** *** *** *** *** *** *** *** *** (0.298) (0.307) (0.303) (0.298) (0.300) (0.299) (0.285) (0.289) (0.288) Eigenvector *** * *** (31.140) (31.084) (30.432) Degree (0.032) (0.032) (0.031) Betweenness (0.003) (0.004) (0.003)

17 Table 3 Results of Spatial Panel Regressions International Short-term Debt Flows This table reports the coefficient estimates of the regression model in Equation (3), which regresses the changes in the S&P Global index on the market excess return (MKT), the earnings-to-price ratio (E/P), the book-to-market ratio (B/M), the logarithm of GDP, the logarithm of the housing price index (HPI), and the unemployment rate. The centralities are calculated using the international short-term debt flow networks. Model I to Model III represent the results for eigenvector, degree, and betweenness centralities, respectively. The standard errors are given in parentheses. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. 16 W (Contiguity) W (K=1 Nearest Neighbors) W (K=2 Nearest Neighbors) Model I Model II Model III Model I Model II Model III Model I Model II Model III λ ** ** ** *** *** *** *** *** *** (0.045) (0.045) (0.045) (0.036) (0.036) (0.036) (0.042) (0.042) (0.042) MKT *** *** *** *** *** *** *** *** *** (0.075) (0.076) (0.076) (0.080) (0.081) (0.080) (0.076) (0.075) (0.075) E/P (0.072) (0.072) (0.072) (0.072) (0.072) (0.072) (0.066) (0.067) (0.067) B/M ** ** ** (0.065) (0.065) (0.065) (0.066) (0.066) (0.066) (0.059) (0.059) (0.059) log(gdp) (0.785) (0.787) (0.788) (0.791) (0.796) (0.796) (0.743) (0.746) (0.748) log(hpi) *** *** *** *** *** *** *** *** *** (3.761) (3.832) (3.770) (3.731) (3.805) (3.741) (3.567) (3.590) (3.553) Unemployment *** *** *** *** *** *** *** *** *** (0.301) (0.302) (0.302) (0.298) (0.300) (0.299) (0.287) (0.288) (0.288) Eigenvector * * * (35.023) (36.304) (32.995) Degree (0.048) (0.049) (0.045) Betweenness (0.002) (0.002) (0.002)