WPS6866 Policy Research Working Paper 6866 International Asset Allocations and Capital Flows The Benchmark Effect Claudio Raddatz Sergio L. Schmukler Tomás Williams The World Bank Development Research Group Macroeconomics and Growth Team May 2014 Policy Research Working Paper 6866 Abstract This paper studies channels through which well-known specific, time-varying effects. Reverse causality does not benchmark indexes impact asset allocations and capital drive the results. Exogenous, pre-announced changes flows across countries. The study uses unique monthly in benchmarks result in movements in asset allocations micro-level data of benchmark compositions and mutual mostly when these changes are implemented (not fund investments during 1996–2012. Benchmarks have when announced). By impacting country allocations, important effects on equity and bond mutual fund benchmarks affect capital flows across countries through portfolios across funds with different degrees of activism. direct and indirect channels, including contagion. They Benchmarks explain, on average, around 70 percent of explain apparently counterintuitive movements in capital country allocations and have significant impact even flows, generating outflows from countries when upgraded on active funds. Benchmark effects are important after and with large market capitalization and better relative controlling for industry, macroeconomic, and country- performance. 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 craddatz@bcentral.cl; sschmukler@worldbank.org; tomas.williams@upf.edu. 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 International Asset Allocations and Capital Flows: The Benchmark Effect Claudio Raddatz Sergio L. Schmukler Tomás Williams* JEL Classification Codes: F32, F36, G11, G15, G23 Keywords: benchmark indexes, contagion, coordination mechanism, ETFs, international portfolio flows, mutual funds * We are grateful to Juan José Cortina, Sebastián Cubela, Julián Kozlowski, Matías Moretti, and Lucas Núñez for excellent research assistance. We received very useful comments from Matías Braun, Ganga Darbha, Gaston Gelos, Bernardo Guimaraes, Marcel Fratzscher, Pedro Matos, Guillermo Ordoñez, Luis Serven, Carlos Vegh, Stefan Zeume, and participants at presentations held at Adolfo Ibañez University (Santiago, Chile), the Annual Meeting of the Chilean Economic Society (Santiago, Chile), the Central Bank of Argentina (Buenos Aires, Argentina), the CEPR Annual Workshop on Macroeconomics of Global Interdependence (Barcelona, Spain), Columbia University (New York), the Darden School of Business (Charlottesville, VA), the LACEA Annual Meetings (Mexico City, Mexico), the Latin Finance Network (Mexico City, Mexico), the London Business School (London, U.K.), NIPFP (Rajasthan, India), the Sao Paulo School of Economics (Sao Paulo, Brazil), the Universidad de San Andres (Buenos Aires, Argentina), and the World Bank (Washington, DC). We thank the support of the World Bank Research Department, the Knowledge for Change Program, and the Latin American and the Caribbean Chief Economist Office. Raddatz is with the Central Bank of Chile. Schmukler is with the World Bank Research Department. Williams is with Universitat Pompeu Fabra. The views expressed here do not necessarily represent those of the Central Bank of Chile or the World Bank. Email addresses: craddatz@bcentral.cl; sschmukler@worldbank.org; tomas.williams@upf.edu. 1. Introduction Theories and empirical work abound about how capital is invested internationally, studying the behavior of both country portfolios (international asset and liability positions) and capital flows. A significant part of the literature has focused on the role that economic fundamentals play in international investment decisions.1 In this paper, we focus instead on another factor that, so far, has been mostly absent from the literature on international investments and that we call “the benchmark effect.”2 The benchmark effect refers to various channels through which prominent international equity and bond market indexes (such as, the MSCI Emerging Markets Index or the MSCI World Index) affect asset allocations and capital flows across countries. These indexes have become popular and are frequently used as benchmarks by international mutual funds, which manage a significant part of the international assets. By helping alleviate agency problems, benchmarks allow the underlying investors and the supervisors to evaluate and discipline the fund managers on a short-run basis using, for example, the tracking error of the fund (the deviation of its returns from the benchmark returns).3 To the extent that the investment strategy of these funds is pinned down by the composition of their benchmark indexes (“benchmarks”), changes in the weights that a popular benchmark gives to different countries can trigger a similar rebalancing among the funds that track it and result in sizeable movements in international portfolio allocations and capital flows. Furthermore, because a growing number of mutual funds follow benchmarks more passively as a way to cut costs, increase transparency, and provide simple investment vehicles (such as, index funds and exchange-traded funds or ETFs), the importance of the benchmark effect is likely to rise over time. Although the effect of benchmarks on international asset allocations and capital flows has not received much attention in the international finance literature, it is frequently mentioned in the broader discussions. For example, when Israel was moved from the MSCI Emerging Markets Index to the World Index (composed of developed markets) capital was expected to leave the country at the time of the switch due to the 1 Some examples of the many papers on the topic are Di Giovanni (2005), Kraay et al. (2005), Lane and Milesi-Ferreti (2007), Antràs and Caballero (2009), Martin and Taddei (2013), Reinhardt et al. (2013), and Gourinchas and Rey (2014). 2 Several papers study the importance of benchmarks, focusing primarily on the performance evaluation of mutual funds relative to their benchmarks, in particular, on whether active management pays (Lehmann and Modest, 1987; Sharpe, 1992; Wermers, 2000; Cremers and Petajisto, 2009; Sensoy, 2009; Cremers et al., 2013; Busse et al., 2014). A related literature focuses on how benchmark redefinitions affect stock returns and liquidity (Harris and Gurel, 1986; Shleifer, 1986; Chen et al., 2004; Barberis et al., 2005; Greenwood, 2005; Hau, 2011; Vayanos and Wooley, 2011). 3 See, for example, Khorana et al. (2005), Shiller (2008), Hellwig (2009), Mishkin (2011), Levy Yeyati and Williams (2012), and Gelos (2013). 1 behavior of funds following these indexes, even though the upgrade was announced in advance and it occurred because Israel’s fundamentals had improved (Business Week, 2010). Similar discussions emerged with the upgrades of Portugal (1997), Greece (2001), Qatar (2014), and the United Arab Emirates (2014), the potential upgrades of the Republic of Korea (2013) and Taiwan, China (2013), and the downgrades of Venezuela (2006), Argentina (2009), and Greece (2013) (Financial Times, 2013a,b,c). One reason for the effect on capital flows is that a country’s inclusion (exclusion) in a benchmark index should drive managers with index-tracking strategies to rebalance their portfolios and direct capital flows into (out of) that country (The Economist, 2012, 2014). In this paper, we systematically study different ways in which benchmarks affect the international asset allocations and capital flows of mutual funds. To understand their impact, we first focus on the cross-sectional and time-series determinants of the country composition of benchmark indexes (“benchmark weights”) as well as on characterizing their dispersion. Then, in the main section of the paper, we present thorough econometric evidence that movements in benchmark weights result in movements in the actual country weights (“weights”) of the funds that declare that benchmark, depending on their degree of activism. Last, we show the consequences that the relation between mutual fund weights and benchmark weights has for capital flows, and explain the various channels through which the benchmark effect impacts those flows. To conduct the research we compile a novel dataset of detailed portfolio allocations across countries by a large number of international mutual funds that we match with the allocations of the benchmarks they follow. The dataset covers the period from January 1996 to July 2012 and contains international mutual funds based in major financial centers around the world investing in at least two countries (i.e., it excludes country funds). A total of 2,837 equity and 838 bond funds are in the sample. These equity and bond funds collectively had 1,052 and 293 billion U.S. dollars in assets under management as of December 2011, respectively.4 One important advantage of our database is that it allows us to test how the use of benchmarks affects international capital allocation and capital flows. First, we measure the independent effect of these benchmarks on country allocations and capital flows after controlling for several factors often mentioned in the literature, most notably industry and macroeconomic effects. In particular, the fact that the weight of a given 4Mutual funds are offered to investors in different ways, for example, in different currencies and with different costs. These funds have the same portfolios but many times are counted as separate funds. In our data, we just count them once to avoid repeating the portfolios, but we report their aggregated assets. 2 country can move in different magnitudes, or even different directions across benchmarks at each point in time, allows us to isolate the effect of benchmarks on the mutual funds that follow them. Second, because benchmarks are adjusted frequently, and subject to significant exogenous revisions, we are able to test the causality from benchmarks changes to mutual fund portfolio changes. Third, the benchmark effect can lead to counterintuitive movements in capital flows by, for example, generating outflows when countries are upgraded, directing flows to countries with deteriorating fundamentals, and reducing outflows for countries with declining asset prices. Fourth, by linking different countries in the same portfolio, benchmarks can trigger contagion and other effects across countries in that portfolio, connecting countries that might otherwise be disconnected (like Brazil, Russia, India, and China in the BRIC index) through their relative performance and their underlying investors. Regarding how benchmarks work, the cross-sectional results show that benchmark weights are related to country fundamentals such as a country’s market capitalization, GDP, GDP per capita, and the quality of its institutions. These same fundamentals help explain whether a country is included in a benchmark index or not. Over time, changes in benchmark weights can be largely explained by changes in a country’s market capitalization relative to the other countries included in the benchmark. This can impinge a pro-cyclical bias in benchmark allocations because countries that do relatively well will tend to gain weight in a benchmark relative to the rest. Nonetheless, there is still a sizeable fraction of movements in benchmark weights that cannot be accounted for in this way. Importantly, at a given point in time, the benchmark weight of a given country varies significantly across benchmarks. While in some benchmarks a country might become more prevalent, in others it might lose importance. Benchmarks have statistically and economically significant effects on mutual fund allocations and capital flows across countries. Starting with allocations, our results indicate that mutual funds follow benchmarks rather closely. For example, a 1 percent increase in a country’s benchmark weight results on average in a 0.7 percent increase in the weight of that country for the typical mutual fund that follows that benchmark. There is also relevant heterogeneity across funds. Explicit indexing funds follow benchmarks almost one-for-one, generating some mechanical effects in allocations and capital flows. Although the most active funds in our sample are less connected to the benchmarks, they are still significantly influenced by their behavior, with about 50 percent of their allocations explained by the benchmark effect. These benchmark effects 3 on the country portfolios of mutual funds are relevant even after controlling for time- varying industry allocations and country-specific or fundamental factors, among others. Furthermore, the results are not driven by reverse causality. Exogenous events that modify benchmark indexes (such as, downgrades/upgrades of countries, changes in the assets covered for each country, and changes in the loading of each asset or country) affect benchmark weights. This effect is separate from any possible endogenous pressure that mutual funds could exert on allocations, returns, and eventually benchmark weights. By influencing the mutual fund asset allocations across countries, benchmarks also affect their capital flows across them. For given past allocations, realized returns, and net inflows to a fund, there is a direct relation between the fund’s allocation and its capital flows to various countries. This association decreases with the degree of activism as funds might reallocate their holdings across countries. Therefore, reallocations in the benchmarks directly impact capital flows through the reallocations in the fund weights. Furthermore, because the sensitivity of country flows to fund flows is partly mediated by the benchmark weight, the use of benchmarks might also generate amplification and contagion effects across countries. These effects arise from the impact that a shock to a country’s returns or to the returns of other countries in its benchmark has on its benchmark weight. We show algebraically the presence of these direct, sensitivity, amplification, and contagion effects and describe them through various examples derived from our data. The benchmark effects documented in this paper can help understand some of the discussions in the literature related to cross-country portfolio allocations. Theoretical work shows that benchmarks can matter for portfolio allocations because managers will optimally tilt their portfolio to the assets in the index used to track their performance (Basak and Pavlova, 2012). But this effect is not trivial and has not been tested empirically. In practice, as Appendix 1 discusses, the extent to which the portfolios of both passive and active funds are linked to their benchmarks depends on several factors, including the manager’s risk aversion and the correlation among the assets in the benchmark portfolio, among others things. Moreover, mutual funds declare prospectus benchmarks but they need not follow them, as deviations from benchmarks could bring greater profitability (Cremers and Petajisto, 2009). Furthermore, the number of assets in benchmark indexes is much larger than that held in international mutual fund portfolios (Didier et al., 2013), which suggests that some funds do not fully replicate these indexes. We contribute to these discussions by showing, for different types of mutual funds, how 4 closely related the country portfolios are to their benchmarks, and how shocks to the latter affect the former. Our findings on the benchmark effect also shed light on some of the numerous discussions on international capital flows. First, our findings show that the use of benchmarks to reduce principal-agent problems and the mechanics of benchmark construction have an independent effect on capital flows, aside from the role that fundamentals and industry factors play. Moreover, benchmarks seem to account for some of the shifts in capital flows and contagion effects that are sometimes difficult to explain. Second, through their effect on individual portfolios, benchmarks could act as a coordinating mechanism that leads mutual funds (and other asset managers following similar strategies) to move in tandem in given countries, having quantitatively significant systemic effects through herding-like behavior. This is important because individual funds tend to be relatively small compared to the size of capital flows to a country. While there is a large literature showing that mutual funds might imitate their peers and display herding behavior (Scharfstein and Stein, 1990; Froot et al., 1993; Hirshleifer et al., 1994; Hong et al., 2005), there exist only a handful of cases where coordination has been shown empirically (Chen et al., 2010; Hertzberg et al., 2011).5 Here we provide evidence consistent with another coordinating mechanism. Third, the existing literature shows that mutual funds tend to behave pro-cyclically and can have important effects on domestic markets (Kaminsky et al., 2004; Gelos and Wei, 2005; Broner et al., 2006; Jotikasthira et al., 2012; Forbes et al., 2012; Fratzscher, 2012; Raddatz and Schmukler, 2012; Stein, 2013; IMF, 2014). However, it has only started to show why and how these effects take place. Our results suggest that benchmarks might be a potential avenue through which these effects occur. Fourth, our findings provide a possible explanation for the momentum and feedback loop theories (Barberis et al., 1998; Daniel et al., 1998; Shiller, 2000; Gervais and Odean, 2001; Wurgler, 2011; Vayanos and Wooley, 2013). A shock to a country’s return might lead to a higher benchmark weight, a larger mutual fund allocation, and larger capital flows, perpetuating these loops. The rest of the paper is organized as follows. Section 2 describes the data. Section 3 analyzes how benchmarks behave. Section 4 studies the effect of benchmarks on mutual fund asset allocations. Section 5 analyzes the relation between asset allocations and capital flows and the effects of benchmarks on these flows. Section 6 concludes. 5 Other possible mechanisms are the exposure to common funding shocks, pure herding, or the use of similar investment strategies unrelated to benchmarks. 5 2. Data Our database consists of: (i) country weights or weights, , which are the country portfolio allocations of international mutual funds (those investing in several countries); (ii) benchmark weights, , which are the country allocations in the relevant benchmarks; (iii) mutual fund-specific information, such as its assets, returns, and relevant benchmarks; (iv) country-specific information, such as stock and bond market index returns.6 Throughout the paper, the sub-index i refers to funds, c to countries, and the supra-index B to benchmarks. For the final database, we cleaned the raw data and merged data from several sources, some of which had not been previously used or matched in the literature. The final data cover the period from January 1996 to July 2012 and constitute an unbalanced panel. Our two main sources for country portfolio allocations of international mutual funds are EPFR (Emerging Portfolio Fund Research) and Morningstar Direct (MS). Both sources include dead and live mutual funds. The data from EPFR are at a monthly frequency, and include open-end equity and bond funds classified according to their geographical investment scope. Global funds invest anywhere in the world, global emerging funds only in emerging countries, and regional funds in groups of countries within a specific geographical region (e.g., developed Asia).7 Frontier market funds are usually classified as regional funds. The data also comprise portfolios of ETFs. We use only funds that have information for at least one year. For each fund i and each month t, the data contain information on the share of the fund’s assets invested in each of 124 countries and cash, as well as its total net assets (TNAs, ). We also have information on each fund’s static characteristics, such as the asset class, domicile, currency, whether it is an ETF, its strategy (passive or active), and, crucially, its declared benchmark. We complement these data with information on each fund’s net asset value (NAV), obtained from Datastream and MS. We match the funds from these different databases. We use similar data from MS to complement the EPFR data. That is, we use data on country weights, TNAs, NAVs, and static fund characteristics for additional international mutual funds not included in EPFR with at least one year of monthly data.8 6 Benchmark weights are fund-specific because each fund chooses its benchmark. We thus denote it with sub-index i. The same applies to other benchmark characteristics such as benchmark returns. 7 While global funds theoretically can invest anywhere in the world, a large proportion of them track the MSCI World Index, which only has developed countries as constituents. A minor proportion of these funds gauge their performance relative to the MSCI All Country World Index that contains both developed and emerging countries. 8 Although MS includes funds that report quarterly, almost 90 percent of the original MS sample reports allocations on a monthly frequency. 6 This increases importantly the cross-sectional coverage of our final dataset. MS reports country weights in only 52 countries and does not contain data on cash allocations. 9 The combination of the two databases provides us with an extensive cross-sectional and time- series coverage of funds. MS contains a large number of funds after 2007 but very few in earlier years, while EPFR has a more balanced number of funds dating back to 1996. 10 In addition, we use stock and bond market country indexes from J.P. Morgan and MSCI to compute the country returns, , which we impute to each fund’s investment in each country (we do not have information on the actual returns of each fund in each country). We obtained this information from Datastream and MSCI. In addition to our data on fund country weights, we also use data on the country benchmark weights and returns of several major benchmark indexes ( ). We obtain these data directly from FTSE, J.P. Morgan, and MSCI through bilateral agreements, and indirectly through MS for indexes produced by Dow Jones, Euro Stoxx, and S&P. For each of the benchmark indexes in MS and MSCI, we collect data on price returns, gross returns, and net returns. Appendix Table 1 presents a detailed list of the benchmarks indexes included in our data. We rely heavily on the MSCI benchmark indexes because 86 percent of our data on equity mutual funds declare to follow them.11 To match the data on international mutual funds with the benchmark indexes, we assign to each fund the index declared in its prospectus. For funds with no declared index, we impute the benchmark assigned to it by industry analysts, as reported by MS.12 We were able to match 88 percent of the equity funds and 18 percent of the bond funds in our database. The reduced matching of bond funds with their benchmarks is not because of matching problems but for lack of information on the detailed portfolio composition of their benchmark indexes.13,14 We do not use the rest of the funds because it is not clear whether the missing information is due to the fund not following a 9 In our estimations, we only use country allocations and, thus, do not include the residual category of other countries nor cash. 10 In our consolidated database we kept the country coverage of MS (52 countries) and adapted the EPFR database to this format, lumping countries outside these 52 in a residual category called “other equity” (also present in MS). We have also performed robustness tests for the impact of this change for the EPFR database. The results are qualitatively similar. 11 Some funds follow a linear combination of two or more indexes. We use that combination as their benchmark. 12 Results are qualitatively similar when excluding these funds. 13 Most bond funds follow J.P. Morgan bond indexes. However, within this family we could only get access to the detailed composition of the EMBI+, EMBI+ Global, and EMBI+ Global Diversified. 14 There is no agreement in the literature on how to assign benchmarks. To different degrees, papers use the declared benchmark, the one assigned by analysts, and the one that yields the smallest deviation from the fund portfolio (Cremers and Petajisto, 2009; Sensoy, 2009; Cremers et al., 2013; Jiang et al., 2013; Busse et al., 2014). 7 benchmark or following a benchmark unknown to us (for dead funds, this information was impossible to retrieve).15 Our final database consists of an unbalanced panel, where each observation is a country-fund-time observation containing the percentage of TNAs invested in a particular country by a mutual fund, the percentage allocation of that same country at the same time for the assigned benchmark, plus fund-specific information. Because we have much more matched data on equity funds than bond funds we rely more heavily on the former than the latter. However, despite the much lower data availability, the results on bond funds are broadly consistent with those on equity funds. We also classify funds according to their degree of activism, that is, the extent to which their country allocations deviate on average from those of their respective benchmarks. Following Cremers and Petajisto (2009) but using country weights instead of security weights, we classify funds as “explicit indexing,” “closet indexing,” “mildly active,” and “truly active” funds. Explicit indexing funds are either ETFs or passive funds. Closet indexing funds do not declare to be passive but behave similarly to explicit indexing funds. Mildly and truly active funds are those that deviate importantly from their self-declared benchmarks. Specifically, for each fund we first compute its active share each month and then take the average over time as a time-invariant measure of a fund’s deviation from its benchmark allocations. This measure gives the average percentage of a fund’s portfolio that deviates from its benchmark. 16 Because mutual funds in our sample have only long positions, this measure ranges from 0 to 100 percent. We then define closet indexing funds as those that on average have an active share within two standard deviations of the active share of explicit indexing funds. Funds not belonging to the explicit indexing or closet indexing groups are classified into mildly active (truly active) if they are in the lower part (upper) of the distribution of the active share measure (using the median active share).17 Our database of mutual funds, before matching with the benchmarks, contains 2,837 equity funds and 838 bond funds with three geographical investment scopes: global, global emerging, and regional funds (Table 1). Equity funds are domiciled around the entire world but most of the funds are located in Canada, France, Ireland, 15 Having access to the benchmarks makes the matching relatively straightforward given that funds have increasingly reported their benchmarks. For instance, among the funds covered by EPFR, 28 percent of equity funds did not report a benchmark in 1996, while 5 percent did not do so in July 2012. Our matching for equity funds is rather complete because only 9 percent of equity funds in our sample do not report (or are assigned) a benchmark. For bond funds, that number is 16 percent. 16 More formally, it is defined as ∑| |. 17 The results are robust to the selection of benchmarks, where we assign the minimum active share benchmark to each fund. 8 Luxembourg, the United States (U.S.), and the United Kingdom (U.K.). Most bond funds are domiciled in Denmark, Germany, Ireland, Israel, Italy, Luxembourg, the U.S., and the U.K. The TNAs of mutual funds increased significantly over time, reaching large values at the end of the sample (Figure 1). In 2011, the equity (bond) funds in our sample had 1.2 trillion (303 billion) U.S. dollars in TNAs. Moreover, funds in our combined dataset capture an important part of the assets held by the industry of international funds. For example, our sample of U.S.-domiciled equity funds had 442 billion dollars in TNAs, while the Investment Company Institute (ICI) reports that, during the same period, U.S. (non-domestic) international funds held 1.4 trillion dollars including the numerous country funds that we exclude due to our interest on country weights. Similar estimates for Europe from the European Fund Asset Management Association (EFAMA) show that our sample accounts for approximately 53 percent of the international funds in this region. Explicit indexing funds (mostly ETFs) represent a fast growing but still relatively small share of the industry. By also including closet indexing funds, both the level and growth rate of the funds that closely track benchmark indexes increases significantly.18 3. Benchmark construction and behavior Benchmark weights are assembled with the portfolio weights of individual securities included in a benchmark index, aggregated at the country level according to the market where the security was issued. That is, international benchmark indexes are typically constructed using a bottom-up approach and consist of composite stock (or bond) market indexes that include securities from many countries as constituents. The following example from MSCI, the provider of the most prevalent equity indexes, illustrates more details on how benchmarks are assembled (other companies use a similar approach).19 MSCI first defines the main scope of an index (such as, geography, industry, and type of firms) and in which category each country is classified at each point in time 18 The trends exhibited by the share of total assets of ETFs in our sample also appear in (unreported) data on U.S. mutual funds from the Investment Company Institute (ICI), which does not identify closet indexing funds. 19 While benchmark indexes have become very popular over the past years and there exist a high diversity of indexes, a large percentage of mutual funds declare a concentrated subset of these indexes. For instance, the MSCI Emerging Markets Index, the MSCI World Index, or the MSCI Europe Index are followed by 52 percent of our equity mutual fund sample. J.P. Morgan constructs many of the indexes bond funds follow. Appendix 2 provides more information on the benchmark index industry. 9 (developed, emerging, or frontier). Then, it selects a number of securities that fall within the scope and meet the size, market capitalization, liquidity, and other requirements. Each of these securities gets a loading (or inclusion factor) in the index portfolio assigned by the index producer according to how much it meets the index-construction criteria and how accessible it is to investors (given by the free-float market capitalization, restrictions to foreign investors, and so forth). The return of the index consists of the returns of its constituent securities, using various approaches to aggregate fluctuations in individual instruments (e.g., Laspeyres, chain-weighting). Namely, each index captures the market capitalization weighted returns of all constituents included in the index. The indexes are periodically rebalanced to ensure their continuity and representativeness (MSCI Barra, 2013a,b). We then study how benchmark weights are related to country characteristics by analyzing the factors that affect the probability of including a country in a benchmark (the extensive margin) and the weight it receives (the intensive margin). We start with a cross-sectional analysis and relate these two margins to common market and country characteristics, namely, market capitalization (MC), gross domestic product (GDP), GDP per capita (GDPPC), the composite ICRG index of country risk (RISK), the Polity index of quality of institutions (INST), and the Chinn-Ito index of capital account openness (KOPEN). We estimate the parameters of the following specifications for the extensive and intensive margins. - Extensive margin: ( ) ( ) (1) { - Intensive margin: (̅ ) (2) where ̅ is the time average of the weight of country c in benchmark B, the country level variables (indexed by c) correspond to the average over 1996-2012, and the errors are clustered at the country level. The extensive margin specification estimated by Probit and Ordered Probit considers only the composition of the MSCI benchmarks (whether a country is in an index or not and whether a country is in the developed or 10 emerging market category or not). 20 Thus, it does not use most of the upgrades and downgrades occurring within our sample, which we exploit later. The regressions for the extensive margin show that the inclusion of a country in an index is positively related to a country’s size, measured by market capitalization and GDP (Table 2, Panels A and B). Moreover, country risk and the quality of institutions are significant for the probability of including a country in an index. The regressions for the intensive margin (Table 2, Panel C) show that the country weight is positively correlated with the country’s market capitalization (with and without a free-float adjustment), GDP, GDP per capita, and the quality of institutions. Although at the security level the benchmark weights are directly related to market capitalization, at the more aggregate country level this does not need to be the case. For this to occur, it is necessary that the market capitalization of a country’s selected securities, corrected by their individual inclusion factors, be proportional to the overall market capitalization of the country. Our results confirm that the aggregation process preserves some of the characteristics of the individual security-level composition of the index, so that the selection of securities across countries is related to the size of different markets after controlling for other factors. Next, we explore the panel structure of our benchmark weight data to study the determinants of their time-series variation. We focus only on the intensive margin because we do not have many episodes of country upgrades and downgrades. Based on our cross-sectional results, we concentrate on the role of changes in a country’s market capitalization relative to other countries in the same benchmark. We estimate the following regression: ( ) ( ) (3) where ( ) is the change in the log-weight of country c in benchmark B between t-1 and t, and ( ) is the difference between the net return of country c and that of benchmark B at time t. 21 The coefficient of interest is , which captures the relation between relative returns and percentage changes in benchmark weights. The parameters , , and correspond to benchmark-country, benchmark-time, and country-time 20 In these estimations, we use only the MSCI indexes to analyze a consistent sample. For instance, FTSE classifies Korea as a developed country, while MSCI puts it under the emerging market category. We choose MSCI because it represents the majority of the benchmarks in our sample and it allows us to estimate well the extensive margin. 21 This derivation uses that the change in the market capitalization of a country relative to other countries in the same benchmark equals its relative gross return (assuming a constant number of securities in an index). It also uses that the log gross relative return, ( ), is approximately equal to the difference in net returns ( ). 11 fixed effects that absorb non-parametrically those dimensions of the data. The inclusion of these fixed effects means that the identification comes exclusively from the time variation of the data (within a benchmark-country). The results show that benchmark weights move almost one-to-one with relative returns (Table 3). Namely, changes in benchmark weights are almost completely driven by changes in a country’s relative market capitalization and, as such, exhibit almost complete pass-through from relative returns at the monthly frequency. The results are robust to the inclusion of the different fixed effects capturing different types of shocks. Although the return shocks get transmitted entirely to the benchmark weights in the short run, the importance of pass-through slightly declines at lower frequencies for equity benchmarks, which is consistent with other factors affecting these benchmarks (e.g., changes in free floats, inclusion factors, and so forth). Including more lags of log changes in benchmark weights or relative returns do not have much effect on the relative return coefficients, and the economic and statistical significance of the other lags diminish rapidly. The results for equity funds and bond funds are fairly similar across specifications. The results also show that, while relative returns are important determinants of changes in benchmark weights, they are far from being the whole story. In fact, the of the various regressions are between 0.3 and 0.6 at the monthly level. There are two reasons for this. First, because we do not know the return of a country within each benchmark and instead use a common country return imputed to all benchmarks that include that country, the residual term could capture these differences. Nonetheless, this residual is probably small due to the bottom-up approach; that is, benchmarks in the same country category (developed, emerging, frontier) will tend to have the same stocks for each constituent country and then the country returns will be similar across them. Second, there is time variation in the inclusion factor that leads to re-weighting of indexes that are not captured by movements in market capitalization. These types of fluctuations are exogenous to the performance of a country and will prove very useful for identification in our empirical analysis. A second important source of identification stems from our data. Because the benchmark weight for each country moves according to its relative return, it can move in opposite directions in two different benchmarks. This dispersion of benchmark weight movements across benchmarks within countries plays a key role in our identification strategy below. The reason is that it would be difficult to disentangle country-specific 12 determinants from benchmark effects if a country’s benchmark weight moves similarly across all benchmarks that include that country. Figure 2 shows various measures of benchmark weight dispersion. Panel A shows for each country and point in time a histogram of the distance (in absolute terms) between the log change in benchmark weight and its average log change across all the benchmarks that include that country. It shows that around 10 percent of the mass of observations lies beyond 10 percentage points of weight differential. Panel B displays the ellipse capturing 90 percent of the scatter plot of the log change in benchmark weight for each country and point in time for different pairs of benchmarks pooling all observations. The figure shows that while these changes are positively correlated, there is important dispersion. In fact, there is a significant amount of observations (20 percent) in the second and fourth quadrants, which correspond to cases where the log change displays contemporaneous opposite movements in the two benchmarks. 4. Benchmarks and asset allocations In this section we estimate how benchmark weights affect mutual fund weights. Before starting with the systematic analysis, we provide the example of Israel, which illustrates the impact of benchmarks on the allocations of different types of funds. This change is part of the often-large restructurings that index-producing companies announce about the calculation of their indexes. The most important changes entail upgrades/downgrades of countries between the categories developed, emerging, and frontier markets and changes related to the index construction methodology. In June 2009, MSCI announced its decision to upgrade Israel from emerging to developed market status. In May 2010, the benchmark weight of Israel in the MSCI Emerging Markets Index turned zero and its weight in the MSCI World Index became positive. Figure 3 shows the behavior of the average weight of Israel among the explicit indexing and truly active funds that declare to follow the MSCI Emerging Markets Index and the MSCI World Index. Explicit indexing funds track the benchmark very closely. At the time the upgrade became effective, the funds that tightly follow the MSCI Emerging Markets Index instantly dropped Israel’s weight to zero, while those following the MSCI World Index incorporated Israel to their portfolios. However, when MSCI announced the upgrade decision, these funds did not significantly change their allocation in Israel; instead, they waited until the actual upgrade materialized. Truly active funds did not react 13 so mechanically to the upgrade, but they still gradually adjusted their portfolio in a manner that is consistent with movements in the benchmark weights. While the Israel example involved large reallocations and a complete removal and incorporation into two different indexes, there are many more frequent but smaller changes in the indexes. However, this example serves to make the point that there is a very tight connection between benchmarks and passive funds and a looser connection between benchmarks and active funds. It also shows that these exogenous events to the composition of benchmarks matter for mutual fund allocations. Namely, the reclassification of countries across benchmarks can trigger asset liquidation to reduce the country exposure, which is not necessarily driven by price effects. 4.1 Basic specifications To study more systematically how mutual fund weights respond to benchmark weights, we start by estimating a set of simple panel regressions that relate a fund’s country weight (log country weight) to its benchmark weights (log benchmark weight), including different sets of fixed effects to capture various types of shocks. More specifically, we estimate the parameters of the following specifications: , (4) ( ) ( ) , where, is the weight for fund i, in country c, and at time t; is the respective benchmark weight that fund i follows; and are fund-country and fund-time fixed effects that account for persistent differences in the weight that each fund holds in each country and for the shocks that funds receive at each point in time (such as, redemptions and injections or changes in the cash or other equity positions). The errors, , are clustered at the benchmark-time level, which allows for unobserved correlation among all funds that declare a common benchmark. The results are robust to alternative clustering structures.22 We run these regressions pooling all funds together and separating them by their degree of activism. To help interpret the estimates of and in Equation (4), Appendix 1 discusses a possible portfolio decision framework following Roll (1992) and Brennan (1993), among others. The estimates show, on average, how much the weight of a 22The errors in our specification are correlated at the fund-time level because at each point in time an increase in the weight of a country in a fund’s portfolio requires the decline of other countries. Part of this mechanical correlation is removed by excluding residual countries and cash, but it is still likely to be present. 14 country in a fund increases when its weight on the benchmark increases. Holding other things constant, a manager that is more risk averse or has a smaller tracking error will allocate its weights more closely to the benchmark weights. For managers that want to have a tracking error equal to zero, and will be equal to one. For managers that do not follow the benchmark, and will equal zero. For managers that partially follow the benchmark, and will be between zero and one. The results using all equity funds (Table 4 and 5, Panel A) show that, for a given group of funds, the estimated coefficients for benchmark weights ( and in Equation (4)) vary little across specifications. Using weights (Table 4) or log weights (Table 5) makes very little difference in the point estimate, even though log weight regressions use 900,000 fewer observations (out of 2.5 million). For instance, for the group of all funds and no fixed effects the coefficient obtained in the weight regressions is 0.823, and in the log-weight regressions is 0.821, both statistically significant. The fit of the weight regressions tends to be better than for the log weight ones because of the persistence of zero allocations. The interpretation of the coefficient is different of course; is a slope and an elasticity. Controlling for the two sets of fixed effects reduces the estimated coefficients for the total sample to 0.77 and 0.75 depending on whether we use weights or log weights, respectively. This suggests that part of the relation estimated without fixed effects is driven by the average levels of country and benchmark weights. For instance, larger countries get higher weights across benchmarks and are also more prominent across funds. Controlling for this through the inclusion of fund-country fixed effects shifts the identification to variation in country weights relative to that average and is close to using only time-series variation (we also use fund-time fixed effects to control for shocks over time at the fund level). Nonetheless, both coefficients are statistically significant and not very different from those without including fixed effects. In general, the overall fit of the regressions is good, both including and excluding fixed effects, ranging from about 0.6 to 0.9 for the total sample. Where we find meaningful variation in the estimates is when comparing funds with different degrees of activism. The coefficients of benchmark weights are much larger for more passive funds. For instance, for the regressions in weights including fixed effects, the coefficient for explicit index funds is 0.92 while the coefficient for truly active funds is 0.5 (Table 4, Panel A). In fact, the estimated coefficients decline monotonically with the degree of activism. Nonetheless, the results indicate that benchmark weights are 15 significantly associated with the mutual fund portfolio allocations even for the most active funds in the sample. This is observed not only in the statistical and economic significance of the estimated coefficients, but also in the degree of variance of the data that can be explained by the benchmark. The lowest reaches about 40 percent for truly active funds. A closer look at the results of Panel A in Tables 4 and 5 reveals some interesting patterns for the level regressions. First, explicit indexing funds move almost one-to-one with benchmarks, both in terms of slopes and elasticities. The point estimate fluctuates between 0.92 and 0.97, depending on whether we use fixed effects or not. The percentage of the variance explained is also higher relative to all funds. It ranges from 95 to 99 percent. Second, estimates for closet indexing funds are close to those of explicit indexing ones, with an estimated coefficient that ranges between 0.9 and 0.97, and similar R-squared estimates. In fact, they are much closer to explicit indexing than to mildly active funds, whose estimated coefficients range between 0.82 and 0.87. Third, in the case of truly active funds, benchmark weights explain 40 percent of the variation in their country weights without including fixed effects. But once fund-country and fund-time fixed effects are included this explained variation increases to almost 85 percent. This indicates that an important part of their “activism” comes from persistent deviations from their benchmarks. The results for bond funds are qualitatively similar (Table 4 and 5, Panel B). However, explicit indexing funds do not move one-to-one with benchmarks, although the variation explained by the benchmarks is still 99 percent when including the fixed effects. This might be due to a small sample problem given that we have few explicit indexing bond funds in our sample. Moreover, fund managers might invest differently in bonds than in equities due to the different nature of these markets, which might explain the somewhat smaller coefficients for bond funds in general. For example, Raddatz and Schmukler (2012) show that bond funds hold more cash as a buffer against shocks, which could explain a smaller reaction to benchmarks. 4.2 Is it benchmark weights? Our baseline results show a tight relation between country weights and benchmark weights that goes beyond persistent country-fund allocations and fund-time shocks. Nonetheless, there are multiple caveats with these results that relate to technical issues, omitted variables, and causality. The results presented in the rest of this section address 16 these issues and further establish that our main finding largely comes from the causal impact that benchmark weights have on portfolio allocations. A first technical concern comes from the persistence of country and benchmark weights. Although weights do vary over time, they are very persistent. Because benchmark weights also move slowly, it is possible that the relation between them comes mainly from the cross-sectional dimension of the data. This concern was partially addressed by the inclusion of country-fund fixed effects in the baseline regressions, but a more stringent test of the hypothesis of a time-series relation between these two series can be run by estimating our main specifications in differences. That is: , (5) ( ) ( ) . Obtaining a significant estimate for and would ensure that, beyond a long-run, persistent relation between country and benchmark weights, changes in the latter are also associated with changes in the former in the short run. The results show that this short-run relation also exists (Tables 4 and 5). The coefficients estimated for and are a bit smaller but similar to those estimated in levels. They are also positive and statistically and economically significant. Moreover, they follow a similar pattern for different degrees of activism. Panel B shows that, with the standard caveat of having significantly fewer observations, the results tend to be similar for bond funds.23 An alternative explanation for the result that funds follow benchmark weights is that they follow the industry, and benchmark weights are capturing the impact of an industry-level omitted variable. Indeed, the literature on managerial incentives highlights the use of relative performance to evaluate managers against the industry (thus accounting for common shocks). To control for the possibility that the mutual funds in our sample are following the industry, we add the median weight across a specified segment of mutual funds to our regressions (Table 6). As in the previous case, this is a complex test because if all funds in an industry are following similar benchmarks, the median weights will be very similar to those of the benchmark. We thus rely on the dispersion of the industry behavior relative to each of the benchmarks to separate the two effects. We obtain coefficients for benchmark weights of similar size and significance as those of our baseline results (Table 4). Industry weights are positive and 23In unreported robustness exercises, we estimated other dynamic specifications with several lags and an error correction term. The economic significance of those additional terms tends to be small relative to the contemporaneous change in benchmark weights, not changing our conclusions. 17 statistically different from zero, but their point estimate is much smaller than those of benchmark weights. For example, for all equity funds at the monthly frequency the coefficient of benchmark weights is 0.67 while that of industry weights is 0.36. Moreover, the R-squared coefficients increase only marginally by adding the industry weights. Furthermore, in unreported regressions we find that, when including only industry weights, their coefficients are much larger. They only diminish when we include benchmark weights. That is, instead of benchmark weights mistakenly capturing the role of industry weights, these tests favor the explanation that the industry effects on mutual fund allocations might be driven by all funds following similar benchmarks. Another potentially important omitted factor emphasized by the literature is related to macroeconomic effects that affect both mutual fund and benchmark weights. Macroeconomic variables might affect allocations if, for example, mutual fund managers and benchmarks invest in countries with better fundamentals. Also, common shocks to returns can affect both the benchmark and mutual fund weights. We control for this potentially omitted factor in two different ways: parametrically and non-parametrically. First, we include common macroeconomic variables available at monthly frequency such as industrial production, inflation, exchange rate changes, and stock market returns. 24 Second, we simply add a set of country-time fixed effects, absorbing non-parametrically all possible time-varying, country-specific shocks. This second approach identifies exclusively from how the within-time variation in a country’s weight across benchmarks relates to the within-time variation in its weight across funds that follow those different benchmarks. This is a very clean form of identification of an effect that comes exclusively from benchmark weights and not from the country-time variation. The results show that after including standard high-frequency measures of macroeconomic variables as controls the coefficient from the benchmark index remains almost unaltered (Table 7). 25 Moreover, when including country-time fixed effects the results are similar to the ones previously reported. If anything, the coefficients for the more active funds diminish in size but those of the more passive ones remain high. These results show that, although benchmarks might be correlated with macroeconomic 24 In additional tests, we also included the expected values of the macroeconomic variables and obtained similar results. 25 Explicit indexing bond funds are few and do not allow us to perform estimations with country-time fixed effects. 18 factors, there is an independent benchmark effect on mutual fund weights beyond these factors. Figure 4 illustrates the relation between benchmark weights and mutual fund weights based on the regressions with country-time fixed effects. Let ̃ ( ⁄ )∑ be the average weight of country c among all funds following benchmark B at each point in time, and ̂ ( ⁄ )∑ ̃ the average weight across all benchmarks including country c. Figure 4 plots ̂ on the horizontal axis against ̃ on the vertical axis, thus isolating the average fund weight and benchmark weight of a country at each point in time and showing only how the variation in a country’s benchmark weight maps into the variation in its actual weight across mutual funds. There is a clear relation between these two variables that is very tight for explicit indexing funds and somewhat looser as the degree of activism increases. A final concern with our findings is that they could be due to reverse causality. This would be the case if benchmark weights responded to movements in fund allocations instead of the other way around. We believe this is unlikely. Benchmark indexes are built and adjusted using pre-determined criteria that do not take into account actions by fund managers. This mechanical evolution of benchmark weights reduces considerably the scope for reverse causality. Nonetheless, this problem could arise if the aggregate actions of mutual fund managers had a price impact that, in turn, changed relative market capitalizations and benchmark weights. The backward looking nature of benchmark weight calculation at each point in time should reduce this possibility, but the time aggregation at a monthly frequency could still give rise to it. To address this reverse causality possibility, we use regular exogenous changes to the benchmark indexes and study how these changes affect mutual fund weights. These exogenous changes might originate from upgrades/downgrades of countries, additions/deletions of securities from certain indexes, and changes in the inclusion factors of various securities, which have some measurable impact in a country’s benchmark weight. We exploit these changes in three ways that expand the type of evidence already presented for the case of Israel. First, we decompose benchmark weights into a buy-and-hold component that moves with relative returns and an exogenous component driven by changes introduced by the benchmark provider. In the absence of exogenous reallocations, benchmark weights would just follow a buy-and-hold pattern, ( ⁄ ), where and 19 are the return of the country and the return of the benchmark, respectively. With exogenous changes, , benchmark weights follow: ( ⁄ ) (6) We then substitute the benchmark weight in Equation (4) for its two components and estimate the parameters of the following specification: [ ( ⁄ )] . (7) We test whether the coefficient for the exogenous shocks is significantly different from zero. This approach exploits all the variation in benchmark weights that is unrelated to the buy-and-hold component to identify their causal impact. The results show that the exogenous component has a significantly positive effect on mutual fund weights (Table 8). As expected, the relation is decreasing in the degree of activism, but even active fund allocations are positively correlated to this component of benchmark weights. Second, we focus on large events, expanding the type of analysis conducted for Israel. Because these large events are usually pre-announced, finding evidence of an impact on allocations when they take place provides evidence that actual, contemporaneous benchmark weights matter for international mutual funds. However, we face the problem that there are few events of whole country upgrades/downgrades to exploit, so we include episodes of large changes in the intensive margin to increase our statistical power. We identify these “exogenous event times/episodes” using the fact that changes in MSCI indexes are released in the months of February, May, August, and November. 26 We compute the exogenous component during these months as in Equation (6) and assume that finding a large exogenous component (below the 25 th and above the 75th percentile of the sample distribution) in any of these months is likely due to the announcement of an exogenous change in the calculation of the index.27 We then test whether the mutual fund weights respond to benchmark weights differently in days with exogenous events relative to other days by estimating two separate regressions: during no-event times, during exogenous event times, (8) 26For this estimation we exclude non-MSCI indexes. 27We also performed estimations with the 10th and 90th percentiles (instead of the 25th and 75th percentiles) and found qualitatively similar results. Alternatively, we computed estimations with dummies on these months of exogenous changes without finding evidence of changes in the relationship between benchmark and country weights. 20 where “no-event times” are those outside the tails of the distribution mentioned above.28 Finding that ( ) would mean that the relation between benchmark weights and country weights weakens (strengthens) in months when benchmark weights are largely driven by exogenous episodes. Alternatively, not being able to reject the hypothesis that means that the exogenous movements in benchmark weights matter for country weights as much as those driven by relative returns. The results show that the latter is the case and that the link between mutual fund weights and benchmark weights does not change during exogenous episodes (Table 8). That is, funds do not tend to respond differently to exogenous events or other changes in the benchmark weights.29 Third, we exploit another exogenous episode for the MSCI indexes (also used in Hau, 2011) that involved overall index redefinitions. In December 2000, MSCI announced that they would change all their indexes to adjust the market capitalization by the free-float rate (that is, by the proportion of the stocks publicly available). These changes were effective in two steps, at the end of November 2001 and at the end of May 2002. In fact, the changes in (the residual component in equation (6)) at those times were indeed much larger (there was more dispersion in exogenous changes to benchmarks) than during the other months (Table 9, Panel A). To test how mutual funds responded to these events, we regress the changes in mutual fund weights against the changes in the buy-and-hold component and the changes in the exogenous component for the months were MSCI made the change effective. With the exception of the truly active funds, mutual funds responded almost one-to-one to the exogenous changes at the time the indexes were readjusted (Table 9, Panel B).30 From all these exercises we conclude that it is unlikely that our results on the benchmark effect are mainly driven by reverse causality. There is evidence consistent with a causal link between changes in benchmark weights and changes in fund weights. 28 In alternative and related exercises, we found that the link between benchmark weights and relative returns weakens significantly during all the months when there are exogenous revisions to the benchmarks. This suggests that these revisions introduce a wedge in the connection between benchmark weights and relative returns. Thus, not just relative returns matter for the evolution of benchmark weights. 29 Formal tests of equality of the coefficients are not presented for brevity. For the total sample, funds slightly under-react during exogenous times. The same happens for closet indexing and mildly active funds. There is no statistical difference for explicit indexing and truly active equity funds. Still, in all of the cases the difference in point estimates is of little economic significance. 30 Explicit indexing funds are excluded in these estimations due to the low number of observations. 21 5. Benchmarks and capital flows In this section, we explore the consequences of our findings that benchmark weights affect mutual fund weights for the capital flows of these funds, and we study the various channels through which the benchmark effect impacts country flows.31 To capture the relation between benchmark weights and capital flows, we start from the following identity: ̃ ( ), (9) where is the net flow (in dollars) from fund i in country c at time t. is the portfolio weight the fund decides to have in that country at time t, ̃ is the value of the fund’s assets at the beginning of time t, and is the fund’s buy-and-hold weight in that country resulting from movements in total and relative returns. is the net flow (in dollars) to fund i at time t, also known as injections or redemptions. The two terms in the equation above relate to the two forces driving a fund’s flows to a country: net inflows and reallocation. Net inflows to countries occur as net flows to the fund ( ) are allocated across countries in proportion to the fund’s desired country weight at that moment ( ).32 The flows due to reallocation of existing assets, ̃ ( ), arise from the difference between a fund’s desired country weight and the buy-and-hold weight that mechanically results from the fund’s previous allocation and movements in relative returns. Equation (9) shows a direct connection between weights and country flows. Fund managers’ decisions about country weights have a direct impact on country flows. For instance, an increase in the desired weight in a given country induces both a reallocation of existing assets to that country and more inflows to that country when the fund itself has injections. To describe and quantify the various mechanisms through which the benchmark effect operates on flows, it is useful to normalize Equation (9) by lagged fund assets ( ), obtaining: 31 By capital flows we mean the flows of the funds we analyze into countries in which they invest. Because we do not know who is on the other side of the transactions, we cannot determine to what extent these flows are reflected in the balance of payments statistics at the country level. However, according to some estimates, the EPFR funds alone account for around 25 percent of total foreign portfolio investments (from all sources) at the country level (Puy, 2013) and there is a significant correlation between the EPFR flows and those obtained from the balance of payments (Fratzscher, 2012; Miao and Pant, 2012). 32 We use the term “desired country weight” to refer to the weight the fund decides to have in that country considering all the possible constraints it faces. It does not mean to imply that it is the optimal weight that the fund would choose in an unconstrained or partially constrained scenario. For example, if the fund cannot change positions in a country to move to the portfolio suggested by its view of the country fundamentals because doing so would be too costly, we consider the desired outcome of this trade-off. Thus, this is a constrained optimal decision of the portfolio manager. 22 ( ) , (10) where ̃ ⁄ , ⁄ , and . Starting from Equation (10) along with the use of Equation (4) linking and , we can derive the response of flows to changes in several variables, and the role that the link between funds and benchmarks has on these responses. The derivations below summarize the responses of country flows to shocks to benchmark weights, fund flows, own-country returns, and third-country returns, respectively. All of them assume that variables as of t-1 are kept constant. ( ) (11) (12) ( ) ( ) (13) ( ) ( )( ) (14) ̃ ̃ Using Equations (11)-(12), we discuss the different effects of benchmarks on capital flows. While Equation (11) directly shows the response of flows to changes in benchmark weights, the other benchmark effects on flows appear in the first terms of Equations (12)-(14), which include current and lagged benchmark weights.33 Equation (11) captures the “direct benchmark effect,” or the direct impact of changes in benchmark weights. The impact on flows of an exogenous change in benchmark weights (i.e., a change not driven by returns) is proportional to the gross growth in fund assets, or . The proportionality depends on how closely fund weights track benchmark weights, as captured by the estimated in Section 4. Equation (12) shows the “sensitivity effect” in its first term, which captures that an increase (decrease) in a fund’s inflows will increase (decrease) the fund’s capital flows to a country proportionally to the country’s benchmark weight. Thus, benchmark weights determine the sensitivity of country flows to fund flows. The last term in this equation corresponds to the response of the active part of a fund portfolio to the shock. Equation (13) shows the response of country flows to own-country returns. The first term measures the “amplification effect,” according to which an increase in a country’s return has a positive impact on its flows. In this case, the link to a benchmark 33 The derivations take as given and use the following expressions: , ∑ , and ∑ . 23 induces inflows (outflows) into countries experiencing positive (negative) return shocks when a fund expands. The second term captures the extent to which the increase in returns increases the value of the fund’s existing assets and, if fund flows respond to returns, also its injections. The third, negative term in this expression comes from the direct effect of country returns on buy-and-hold weights and, for a given benchmark weight, reallocations. Equation (14) displays the response of country flows to third-country returns. The first term shows the “contagion effect.”34 This effect is qualitatively similar to that in Equation (13), but in this case the effect is negative because an increase in every other country’s returns reduces a country’s relative market capitalization (and thus its benchmark weight). Therefore, it brings home shocks to returns occurring to other countries that share the benchmark. This form of contagion could be benign because negative shocks to other countries bring inflows to the unaffected one (although positive shocks to other countries bring outflows to the unaffected one). However, even under negative shocks to other countries it is possible to have outflows in the unaffected country if the effect on the second term is large enough, namely, if flows to the fund decline strongly enough in response to a shock to its returns. Notice that, when this happens and is small, the second term in Equation (14) dominates and the contagion is no longer benign. Although the contagion effect analyzed here focuses on the impact of third- country returns, shocks to other countries can affect flows to the own country through the direct benchmark effect. For example, the potential upgrade of Korea or Taiwan, China (that account for 25 percent of the MSCI Emerging Markets benchmark) from emerging to developed status would also trigger positive (negative) direct benchmark effects to other countries that share the portfolio with these countries. Even though these effects are all considered as direct benchmark effects in our equations, conceptually they could be related to contagion in the sense of bringing home shocks from other countries in the benchmark. The effects described here affect different types of funds differently. For closed- end, explicit indexing funds, the country flows are different from zero only when there is a direct benchmark effect. For open-end, index funds all the channels operate because of the flows that funds receive. For non-explicit indexing funds, the total country flows 34This contagion effect is different from the “margin call” and other effects described in the literature and occur in the absence of leverage (Calvo and Mendoza, 2000; Kodres and Pritsker, 2002; Hau and Lai, 2012; Manconi et al., 2012). 24 depend on the level of activism and how the manager allocates the active part of the portfolio. However, the effects described above illustrate how their country flows respond to different shocks to the extent that they follow benchmark indexes. Several of these benchmark effects can be observed in our flows data. The direct benchmark effect helps explain the counterintuitive outflows when Israel was upgraded from the MSCI Emerging Markets Index to the MSCI World Index. To show the effect of the exogenous change in benchmark weights we compare the explicit indexing funds tracking these two indexes (Figure 5). The direct benchmark effect captures almost all the variation in country flows for both types of funds, which occur due to all the reallocations right at the time of the switch. To understand the total effect on country flows, it is important to consider both the size of the two groups of funds and the size of Israel in each index. At that time, Israel’s weight in the Emerging Markets Index was 3.17 percent and that in the World Index was 0.37 percent, and the assets in the funds following these two indexes were not very different. Emerging market funds withdrew 2 billion U.S. dollars from Israel while developed market funds injected 160 million. This outflow due to the direct benchmark effect was significant given that, according to its balance of payments, Israel witnessed a 1.4 billion U.S. dollar portfolio equity outflow during the second quarter of 2010 and average quarterly equity inflows of about 700 million between the first quarter of 2005 and the first quarter of 2010. They might also be behind the negative returns in Israel’s stock market during the period of the upgrade. The sensitivity effect shows that countries with higher weights in a benchmark are prone to more inflows (outflows) when the funds receive injections (redemptions). This can explain why large countries might be subject to large changes in capital flows despite their fundamentals. Figure 6 illustrates this effect by showing the flows to Brazil and India from explicit indexing funds tracking the MSCI Emerging Markets Index against the flows into each of these equity funds. In each case, the relation of country and fund flows is depicted by two points in time, when each country had different benchmark weights. The relation becomes steeper as each country’s benchmark weight increases, as shown in Equation (12). For a more systematic analysis, we regress country flows against benchmark weights multiplied by fund flows (Table 10). There is a positive and significant relationship between the two variables, which monotonically decreases with the degree of activism. For example, on average across all equity funds, an injection of one dollar to a fund is associated with country flows of 0.74 dollars times the benchmark weight. Every 25 dollar an explicit fund receives is associated with 84 cents allocated proportionally to the benchmark weight. This number declines for funds that are more active, being 0.69, 0.55, and 0.41 for closet indexing, mildly active, and truly active funds, respectively. The relationship is also maintained when we control for different sets of fixed effects. Under this estimation, a change in the benchmark weight changes the sensitivity of country flows to fund flows as indicated above. A corollary of Equation (12) and the estimates in Figure 6 and Table 10 is that changes in benchmark weights change the sensitivity of country flows to fund flows (which can be derived by taking a second derivative of Equation (12) with respect to benchmark weights). This leads to interesting dynamic interactions between various effects. For instance, a decline in the returns of the rest of the countries sharing a benchmark with country A will induce a higher benchmark weight for country A. But the same increase in benchmark weights makes country A more vulnerable to future movements in fund flows. If in reaction to the initial shock there are large withdrawals of funds, country A would be more affected even though it was the country that performed relatively well. Namely, during good times (when funds are receiving injections), a country that does relatively well gets more country flows. During bad times, a country that does relatively poorly (its weight decreases) is less affected by the outflows. Some of these benchmark effects can be illustrated by the evolution of country flows to China and Russia from explicit indexing funds following the MSCI Emerging Markets Index, before the global financial crisis and during the stronger phase of the European crisis (Figure 7). We also show a similar analysis for Spain and Ireland in the explicit indexing funds tracking the MSCI Europe, Australasia, and, Far East index. Before the global financial crisis China and Russia had similar benchmark weights and flows. However, during the global financial crisis China did relatively well compared to Russia, which increased its benchmark weight importantly. During the peak of the European crisis, emerging market funds had net withdrawals, which translated into much larger outflows in China than in Russia (proportionally to their weights). That is, China was penalized by its stronger performance. A similar pattern is observed for developed countries. Spain and Ireland received inflows during the pre-crisis, with the former receiving four times more flows than Ireland according to its benchmark weight. Still, Ireland received around 80 million U.S. dollars in that period. Immediately after the crisis, Ireland did relatively worse than Spain, and the subsequent outflows were smaller in Ireland than in Spain. 26 We conclude this section by simulating the quantitative importance of the various manifestations of the benchmark effect. To do so, we impute values to the different parameters involved in Equations (11)-(14) using the medians and interquartile ranges of the actual data. Table 11 yields order-of-magnitude estimates for the four effects described above, where a shock entails a move from the 25th to the 75th percentile for each variable in our sample. The different manifestations of the benchmark effect result in non-trivial variations in country flows. The simulation shows that the direct benchmark effect has the highest potential to induce inflows (or outflows). For instance, a 1.5 percentage points increase in a country’s benchmark weight (from 4 to 5.5 percent in this case) results in an inflow corresponding to approximately 30 percent of a fund’s total assets allocated to that country.35 On the other extreme, the sensitivity effect has the lowest impact (a 3.2 percent increase in response to a 4 percentage point increase in fund flows). This is reasonable because, as its name puts it, the direct benchmark effect has a direct impact on flows. An exogenous, independent change in a country’s benchmark weight induces net inflows and reallocation effects to that country in detriment of all other countries. In contrast, an increase in fund flows is shared across all countries where a fund invests, more or less proportionally to the usually small country weights. The sizes of the amplification and contagion effects are identical in our baseline parameterization. They both lie between the direct benchmark and sensitivity effects. The reason is that these effects work indirectly through the response of benchmark weights to each of the changes. These responses depend on the initial level of returns and benchmark weights, but it is usually less than one-for-one. To assess the potential of the overall benchmark effect to induce flows it is crucial to take into account that the various effects described above can interact and build up. For instance, a shock to a country’s returns increases its benchmark weight and induces inflows through the amplification effect. If these inflows are important enough to have an impact on returns, a feedback loop might be established. Also, a current increase in benchmark weights, either through the direct benchmark effect or other channels will increase the future response of that country’s flows to injections through the sensitivity effect. Moreover, with the exception of the direct benchmark effect, other effects could be present for funds that do not follow a benchmark ( ), as shown by the presence of the response of the non-benchmark component to each of the shocks. 35This is an approximation because we divide by , and thus take it as a percentage of a fund’s total assets in a country if it perfectly followed the benchmark. 27 What is particular of the benchmark effect is that the form benchmarks are calculated guarantees that the response of flows to an own-country shock through benchmarks is positive, and it is negative for shocks to the returns to other countries. For the non- benchmark component, the sign of these responses is indeterminate. 6. Conclusions This paper shows how benchmarks affect asset allocations and capital flows across countries using a novel dataset of well-known benchmark indexes and mutual funds from around the world investing in equities and bonds. We find that benchmarks have important effects on both asset allocations and capital flows, not only because funds explicitly declare a benchmark to compare their performance, but also because funds with different degrees of activism tend to follow their benchmark asset allocation closely, though to different extents. Given that benchmarks are based on market capitalization, they instantaneously absorb any return shock to the countries in the index. Benchmark weights also receive frequent, exogenous revisions by the companies that construct them. These benchmark changes affect the mutual fund portfolios, their reallocations, and their sensitivity to injections or redemptions. The effects of benchmarks on mutual fund allocations are significant even after controlling for industry effects, country-time effects, macroeconomic fundamentals, potential reverse causality, and other important micro and macro factors that drive country portfolios. The decisions about allocations impact capital flows through different channels. These results can explain some of the findings documented in the literature, as well as counterintuitive and unexpected movements in cross-country investments and capital markets. First, the reclassification of countries across benchmarks is likely to have significant direct reallocation effects on capital flows, given that assets under management and country weights differ across benchmarks and types of funds. For example, advanced emerging countries tend to have larger weights in emerging market indexes than in developed market ones. Therefore, countries might face capital outflows when upgraded and capital inflows when downgraded. Second, sensitivity, amplification, and contagion effects can occur even when fundamentals or the absolute returns of a country do not warrant them. For example, during global crises, some countries might suffer the curse of being large or having done relatively well. That is, during large retrenchments, countries with larger weights will suffer more withdrawals, although in some cases their larger market capitalization might 28 help them withstand the shock. During generalized declines in asset prices, countries whose prices fall less than other countries in the same benchmark will see their benchmark weight increase and, thus, will be more exposed to subsequent withdrawals by the underlying investors of the funds that follow that benchmark. During good times, when funds receive injections, countries that do relatively well will receive more inflows, witnessing an amplification of the shock that increased its relative return. More generally, as a country becomes more relevant in a benchmark, it becomes more sensitive to shocks because injections and redemptions have stronger effects on the capital flows to this country. While this effect might be entirely driven by fundamentals (e.g., by the country growing relatively fast), it can also be driven by non-fundamental factors such as bubbles, self-fulfilling expectations, shocks to other countries sharing the same benchmark, or exogenous decisions made by the company constructing the benchmark. For example, if investors suddenly favor a country and drive its asset valuations upward, the subsequent injections that the relevant mutual funds receive will be more tilted toward this country. This, in turn, might generate more upward pressure on prices, reinforcing the effect. This positive-feedback loop increases as more funds follow benchmark indexes more closely over time, generating pro-cyclicality and possibly explaining some of the widely documented momentum, whereby investment reallocations are related to past returns. Third, while the underlying investors drive part of the observed behavior of mutual funds, a significant portion is explained by manager behavior. The findings in this paper suggest that a non-trivial part of the manager behavior is driven by the fact that managers follow standard benchmark indexes. Therefore, the use of benchmarks as a disciplining mechanism coordinates the manager decisions across institutions, which might also generate herding, information cascades, and aggregate or systemically important effects. Fourth, the benchmark effects might pose difficulties to investors and policymakers in countries with a limited number of assets in the short run. For example, countries that improve their standing by conducting a restrictive fiscal policy or a better monetary policy will increase the probability of being included in more indexes, but this makes it difficult for bond investors to invest in these countries because the better fiscal policy reduces the volume of instruments available for investment. It might also complicate the conduct of monetary policy as investors purchase the central bank instruments in lieu of the disappearing treasury ones. 29 Although this paper presents several new findings, the research on the effects of benchmarks is likely to expand. First, the evidence suggests that funds worldwide are becoming less active (Cremers et al., 2013) and the number of benchmarks are increasing rapidly. Therefore, the types of benchmark effects documented here are expected to grow. Second, models of international asset allocation and capital flows that use macroeconomic fundamentals and other important factors might start incorporating the type of mechanisms described in this paper. Third, benchmarks offer several advantages for researchers. Among other things, they help compare individual portfolios against some well-known specific portfolio allocations and allow for the identification of various effects. Fourth, although benchmark effects shed light on the behavior of heterogeneous investors, the general equilibrium effects still need to be understood. For example, given that some funds try to replicate their benchmark index almost mechanically, do other funds or sophisticated investors anticipate or compensate for their reaction? Are there wealth transfers? Or do they also follow these benchmarks? How do funds manage their active portfolio? What are the effects of benchmarks on asset prices? Are the price effects documented in the literature (e.g., Hau et al., 2010; Bartram, Griffin, and Ng, 2012; Faias et al., 2012; Jotikasthira et al., 2012) driven by passive investment behavior? What are the effects of benchmarks on capital market financing, the returns to retail investors, the balance of payments, and the real economy? 30 References Antràs, P., Caballero, R., 2009. Trade and Capital Flows: A Financial Frictions Perspective. Journal of Political Economy, 117 (4), 701-744. Barberis, N., Shleifer, A., Vishny, R., 1998. A Model of Investor Sentiment. Journal of Financial Economics, 49 (3), 307-343. Barberis, N., Shleifer, A., Wurgler, J., 2005. Comovement. Journal of Financial Economics, 75 (2), 283-317. Bartram, S., Griffin, J., Ng, D., 2012. How Important are Foreign Ownership Linkages for International Stock Returns? Working Paper 122012, Hong Kong Institute for Monetary Research. Basak, S., Pavlova, A., 2012. Asset Prices and Institutional Investors. American Economic Review, 103 (5), 1728-1758. Brennan, M., 1993. Agency and Asset Pricing. UCLA Working Paper. Broner, F., Gelos, G., Reinhart, C., 2006. When in Peril, Retrench: Testing the Portfolio Channel of Contagion. Journal of International Economics, 69 (1), 203-230. Busse, J., Goyal, A., Wahal, S., 2014. Investing in a Global World. Review of Finance, 18 (2), 561-590. Business Week, 2010. Israel Gets an Upgrade, and Investors Depart. September 9th. Calvo, G., Mendoza, E., 2000. Rational Contagion and the Globalization of Securities Markets. Journal of International Economics, 51 (1), 79-113. Chen, H., Noronha, G., Singal, V., 2004. The Price Response to S&P 500 Index Additions and Deletions: Evidence of Asymmetry and a New Explanation. Journal of Finance, 59 (4), 1901-1929. Chen, Q., Goldstein, I., Jiang, W., 2010. Payoff Complementarities and Financial Fragility—Evidence from Mutual Fund Outflows. Journal of Financial Economics, 97 (2), 239-262. Cremers, M., Ferreira, M., Matos, P., Starks, L., 2013. The Mutual Fund Industry Worldwide: Explicit and Closet Indexing, Fees, and Performance. Working Paper. Cremers, M., Petajisto, A., 2009. How Active Is Your Fund Manager? A New Measure that Predicts Performance. Review of Financial Studies, 22 (9), 3329-3365. Dahlquist, M., Pinkowitz, L., Stulz, R., Williamson, R., 2003. Corporate Governance and the Home Bias. Journal of Financial and Quantitative Analysis, 38, 87-110. Daniel, K., Hirshleifer, D., Subrahmanyam, A., 1998. Investor Psychology and Security Market Under - and Over- Reactions. Journal of Finance, 53 (5), 1839-1886. Di Giovanni, J., 2005. What Drives Capital Flows? The Case of Cross-Border M&A Activity and Financial Deepening. Journal of International Economics, 65 (1), 127-149. Didier, T., Rigobon, R. and Schmukler, S., 2013. Unexploited Gains from International Diversification: Patterns of Portfolio Holdings around the World. Review of Economics and Statistics, 95 (5), 1562-1583. Faias, J., Ferreira, M., Matos, P., Santa-Clara, P., 2012. Does Institutional Ownership Matter for International Stock Return Comovement? Darden School of Business. Mimeo. Financial Times c, 2013a. Greece Really Is a Submerging Market. June 12. Financial Times a, 2013b. UAE and Qatar Gain Upgrade to EM Status. June 12. Financial Times b, 2013c. South Korea: MSCI Upgrade Time? June 20. Forbes, K., Fratzscher, M., Kostka, T., and Straub, R., 2012, Bubble Thy Neighbor: Portfolio Effects and Externalities from Capital Controls. NBER Working Paper No. 18052. Fratzscher, M., 2012. Capital Flows, Push versus Pull Factors and the Global Financial Crisis. Journal of International Economics, 88 (2), 341-356. 31 Froot, K., Scharfstein, D., Stein, J., 1993. Risk Management: Coordinating Corporate Investment and Financing Policies. Journal of Finance, 48 (5), 1629-1658. Gelos, G., 2013. International Mutual Funds, Capital Flow Volatility, and Contagion–A Survey. In The Evidence and Impact of Financial Globalization. Edited by G. Caprio, T. Beck, S. Claessens, and S. Schmukler. Elsevier, 131-143. Gelos, G., Wei, S.-J., 2005. Transparency and International Portfolio Holdings. Journal of Finance, 60 (6), 2987-3020. Gervais, S., Odean, T., 2001. Learning to be Overconfident. Review of Financial Studies, 14 (1), 1-27. Greenwood, R., 2005. Short- and Long-Term Demand Curves for Stocks: Theory and Evidence on the Dynamics of Arbitrage. Journal of Financial Economics, 75 (3), 607-649. Gourinchas, P., Rey, H., 2014. External Adjustment, Global Imbalances and Valuation Effects. In Handbook of International Economics, Volume 4. Edited by G. Gopinath, H. Helpman, and K. Rogoff. Elsevier, 585-645. Harris, L., Gurel, E., 1986. Price and Volume Effects Associated with changes in the S&P 500 List: New Evidence for the Existence of Price Pressures. Journal of Finance, 41 (4), 815-829. Hau, H., 2011. Global versus Local Asset Pricing: A New Test of Market Integration. Review of Financial Studies, 24 (12), 3891-3940. Hau, H., Lai, S., 2012. The Role of Equity Funds in the Financial Crisis Propagation. CEPR Discussion Papers 8819. Hau, H., Massa, M., Peress, J., 2010. Do Demand Curves for Currency Slope Down? Evidence from the MSCI Global Index Change. Review of Financial Studies, 23 (4), 1681-1717. Hellwig, M., 2009. Systemic Risk in the Financial Sector: An Analysis of the Subprime- Mortgage Financial Crisis. De Economist, 157 (2), 129-207. Hertzberg, A., Liberti, J.M., Paravisini, D., 2011. Public Information and Coordination: Evidence from a Credit Registry Expansion. Journal of Finance, 66 (2), 379-412. Hirshleifer, D., Subrahmanyam, A., Titman, S., 1994. Security Analysis and Trading Patterns When Some Investors Receive Information before Others. Journal of Finance, 49 (5), 1665-98. Hong, H., Kubik, J., Stein, J., 2005. The Neighbors’ Portfolio: Word of Mouth Effects in the Holdings and Trades of Money Managers. Journal of Finance, 60 (6), 2801-2824. IMF, 2014. Global Financial Stability Report. International Monetary Fund. Jiang, H., Verbeek, M., Wang, Y., 2013. Information Content when Mutual Funds Deviate from Benchmarks. Management Science, forthcoming. Jotikasthira, C., Lundblad, C., Ramadorai, T., 2012. Asset Fire Sales and Purchases and the International Transmission of Financial Shocks. Journal of Finance, 67 (6), 2015- 2050. Kaminsky, G., Lyons, R., and Schmukler, S., 2004. Managers, Investors, and Crises: Mutual Fund Strategies in Emerging Markets. Journal of International Economics, 64 (1), 113-134. Khorana, A., Servaes, H., Tufano, P., 2005. Explaining the Size of the Mutual Fund Industry Around the World. Journal of Financial Economics, 78 (1), 145-185. Kraay, A., Norman, L., Servén, L., Ventura, J., 2005. Country Portfolios. Journal of the European Economic Association, 3 (4), 914-945. Kodres, L., Pritsker, M., 2002. A Rational Expectations Model of Financial Contagion. Journal of Finance, 57 (2), 769-799. Lane, P, Milesi-Ferretti, G. M., 2007. The External Wealth of Nations Mark II: Revised and Extended Estimates of Foreign Assets and Liabilities, 1970–2004. Journal of International Economics, 73 (2), 223-250. 32 Lehmann, B., Modest, D., 1987. Mutual Fund Performance Evaluation: A Comparison of Benchmarks and Benchmark Comparison. Journal of Finance, 42 (2), 233-265. Levy Yeyati, E., Williams, T., 2012. Emerging Economies in the 2000s: Real Decoupling and Financial Recoupling. Journal of International Money and Finance, 31 (8), 2102-2126. Manconi, A., Massa, M., Yasuda, A., 2012. The Role of Institutional Investors in Propagating the Financial Crisis of 2007-2008. Journal of Financial Economics, 104 (3), 491-518. Martin, A., Taddei, F., 2013. International Capital Flows and Credit Market Imperfections: A Tale of Two Frictions. Journal of International Economics, 89 (2), 441- 452. Miao, Y., Pant, M., 2012. Coincident Indicators of Capital Flows. IMF Working Paper 12/55. Mishkin, F. S., 2011. Over the Cliff: From the Subprime to the Global Financial Crisis. Journal of Economic Perspectives, 25 (1), 49-70. MSCI Barra, 2013a. MSCI Global Market Accessibility Review. MSCI Barra, 2013b. MSCI Index Calculation Methodology. Puy, D., 2013. Institutional Investors Flows and the Geography of Contagion. European University Institute Working Paper 2013/06. Raddatz, C., Schmukler, S., 2012. On the International Transmission of Shocks: Micro- Evidence from Mutual Fund Portfolios. Journal of International Economics, 88 (2), 357- 374. Reinhardt, D., Ricci, L., Tressel, T., 2013. International Capital Flows and Development: Financial Openness Matters. Journal of International Economics, 91 (2), 235-251. Roll, R., 1992. A Mean-Variance Analysis of Tracking Error. Journal of Portfolio Management, 18 (4), 13-22. Scharfstein, D., Stein, J., 1990. Herd Behavior and Investment. American Economic Review, 80 (3), 465-479. Sensoy, B., 2009. Performance Evaluation and Self-Designated Benchmarks Indexes in the Mutual Fund Industry. Journal of Financial Economics, 92 (9), 25-39. Sharpe, W., 1992. Asset Allocation: Management Style and Performance Measurement. Journal of Portfolio Management, 18 (2), 7-19. Shiller, R., 2000. Irrational Exuberance. Princeton University Press. Princeton, NJ. Shiller, R., 2008. The Subprime Solution: How Today’s Global Financial Crisis Happened, and What to Do About It. Princeton University Press. Shleifer, A., 1986. Do Demand Curves for Stocks Slope Down? Journal of Finance, 41 (3), 579-590. Stein, J., 2013. Overheating in Credit Markets: Origins, Measurements, and Policy Responses. Speech at the Federal Reserve Bank of St. Louis, February 7. The Economist, 2012. Nigerian Debt Gets the ‘Emerging Market’ Seal of Approval. October 27. The Economist, 2014. Wedge beyond the Edge. April 5. Vayanos, D., Woolley, P., 2011. Fund Flows and Asset Prices: A Baseline Model. FMG Discussion Papers dp667, Financial Markets Group. Vayanos, D., Woolley, P., 2013. An Institutional Theory of Momentum and Reversal. Review of Financial Studies, 26 (5), 1087-1145. Wermers, R., 2000. Mutual Fund Performance: An Empirical Decomposition into Stock- Picking Talent, Style, Transaction Costs, and Expenses. Journal of Finance, 55 (4), 1655-1695. Wurgler, J., 2011. On the Economic Consequences of Index-Linked Investing. In Challenges to Business in the Twenty-First Century. Edited by G. Rosenfeld, J. Lorsch, and R. Khurana. American Academy of Arts and Sciences, 20-34. 33 Appendix 1. Benchmark and portfolio choice This appendix briefly summarizes the determinants of the relation between benchmarks and actual portfolios, starting from the literature on portfolio allocations under benchmark or tracking error constraints. This framework might help understand and interpret the results presented in the paper. Consider the problem faced by the manager of fund that is deciding his portfolio allocation across a set of assets, in our case different countries. The manager’s performance is measured against that of a benchmark index, whose portfolio allocation across the countries is given by , such that [ ] { }, ∑ . The subscript i is used to indicate that the benchmark index corresponds to that tracked by fund i and the subscript c denotes the elements of that index different countries. These properties mean that the benchmark is a long-only portfolio (no short-long strategies are allowed) and that the allocations exhaust all the resources available. Roll (1992) and Brennan (1993), among others, have shown that a manager with mean-variance preferences relative to the benchmark will choose a portfolio allocation that can be expressed as , (A1) where is a hedge portfolio that is proportional to the difference between the minimum variance portfolio and the portfolio where a line through the minimum variance portfolio intersects the efficient portfolio frontier. For a manager that is constrained to follow long-only portfolios, the hedge portfolio must hold { } (A2) ∑ . (A3) The relative importance of the hedge portfolio depends on the manager’s risk aversion, or alternatively on the amount of tracking error (maximum difference between the return of the manager’s portfolio and that of the benchmark) that he is allowed (Roll, 1992). Intuitively, the less risk averse the manager, or the larger the tracking error, the more relevant is the hedge portfolio and the less relevant is the benchmark for the manager’s portfolio. Assume now that we fit a linear regression by OLS to the relation between the manager’s and the benchmark portfolio allocations . (A4) As it is well known, the estimated coefficient ̂ is given by 34 ̂ ( ) ( ) (A5) . ( ) ( ) So, the coefficient will be larger or smaller than one depending on whether the covariance between the benchmark and the hedge portfolio is positive or negative. For instance, if the manager tends to overweight (underweight) the countries with the highest benchmark weights the covariance will be positive (negative) and the coefficient will be larger (smaller) than one. The long-only constraint imposed on the hedge portfolio biases this covariance to be negative. In fact, assume that the manager chooses the hedge portfolio randomly from a distribution that is symmetric around zero (so that the extent of under or overweighting of a country is unrelated to the benchmark weight), but keeps the draw only if it satisfies the feasibility constraints described in Equation (A3). The higher (lower) the benchmark weight of a country, the higher the probability that a random draw that overweighs (underweights) the country will hit the upper (lower) constraint and has to be replaced. This random selection process will result in draws that make more likely to underweight (overweight) countries with higher (lower) benchmark weights. The extent of the negative bias depends on the degree of activism. Following Cremer and Petajisto (2009), we define the degree of activism as the sum of the absolute value of the portfolio deviations from the benchmark: ∑ | | ∑ | |. (A6) Equation (A6) can be interpreted as the source of a constraint imposed on the manager or as a result of his willingness to deviate from the benchmark, determined by his degree of risk aversion or tracking error constraint. A less risk adverse manager or one allowed more tracking error will deviate more from the benchmark and have a higher measured activism. The more the manager tries to deviate from the benchmark, the more likely he will hit one side of the constraints, forcing him to tilt his behavior and inducing a more negative bias. On the contrary, one could always draw the hedge portfolio of a distribution with a variance that is small enough such that the probability of hitting a constraint is negligible, resulting in an estimated coefficient close to one. Such manager will have a very small degree of activism and behave as an index fund. The coefficient will be zero only when the covariance between the hedge and benchmark portfolios equals minus the variance of the benchmark weight. This means that the linear projection of the hedge portfolio on the benchmark portfolio has a slope equals to negative one. In this sense, the hedge portfolio undoes what the benchmark 35 portfolio does and is a situation akin to having an allocation that does not follow the benchmark. In the paper, we estimate a series of regressions similar to that presented in Equation (A4), albeit in a panel setting and controlling for many other determinants in a parametric and non-parametric fashion. The coefficient of that regression tells us, on average, how much the weight of a country in a fund portfolio increases when its weight on the benchmark increases, taking into account the correlation between the benchmark and hedge portfolios present in the data. This is the relation of interest from a forecasting perspective, despite the fundamental relation given by Equation (A1). 36 Appendix 2. The proliferation of benchmark indexes As of May of 2012, there were 267,415 active equity indexes and 63,616 active bond indexes in Datastream, including the many indexes focused on single markets and different industrial sectors. While the number is high, most mutual funds are benchmarked against few and very popular indexes. For instance, the S&P 500 is the most popular index for U.S. funds, while the MSCI World or MSCI EAFE (Europe, Australasia, and, Far East) are the most prevalent indexes among international funds investing in developed markets. While there are approximately 18 companies producing bond indexes, many more companies are involved in the production of equity indexes, including the large international indexing companies (such as, FTSE, MSCI, and S&P), and the national producers of indexes and national stock exchanges. As of December 2012, the largest producer of equity indexes was MSCI with 126,821 indexes, then FTSE with 39,738 indexes, Russell with 27,826 indexes, S&P 17,723 with indexes, and, Dow Jones with 14,771 indexes. The largest producer of bond indexes was J.P. Morgan with 20,390 indexes, followed by Merrill Lynch with 18,897 indexes, Citigroup with 10,281 indexes, and Barclays Capital with 3,963 indexes. Despite the apparent diversity of indexes, the popularity of indexes is highly skewed, with a few indexes being followed by many investors. While there are broad indexes such as those focused in world markets, advanced (or developed) markets, emerging markets, frontier markets, or country specific, these are further subdivided by different characteristics. For instance, MSCI has different indexes according to the currency of denomination (e.g., U.S. dollar, euro, local), returns (e.g., net returns, gross returns, total returns), industry, size (e.g., large cap, medium cap, small cap), and style (e.g., value, growth). This generates a wide diversity among indexes, which has been increasing over time. 37 Figure 1 Total Net Assets This figure shows the average total net assets (TNAs) per year in the database and how these total net assets (TNAs) are distributed among funds with different degree of activism. Panel A shows these figures for equity funds and Panel B for bond funds. Although our data on bond funds start in 1997, for this figure we exclude the years up to 2001 due to the few observations available. A. Equity Funds B. Bond Funds 1,600 350 1,400 300 1,200 Billions of U.S. Dollars 250 1,000 200 800 150 600 100 400 200 50 0 0 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 1996 1997 1998 1999 2000 2001 2002 2003 2006 2007 2008 2009 2010 2011 2004 2005 100% 100% 90% 90% 80% 80% 70% 70% Share of TNAs 60% 60% 50% 50% 40% 40% 30% 30% 20% 20% 10% 10% 0% 0% 1996 1998 1999 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 1997 2000 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Figure 2 Contemporaneous Changes in Benchmark Weights This figure shows the country weight variation across benchmarks. Panel A shows, for each country and point in time, the histogram of the distance (in absolute terms) between the log changes in benchmark weight and the average change across all the benchmarks that include that country. Panel B shows an ellipse containing 90 percent of the observations of all the pairwise combinations for two different benchmarks for the same country at the same time. Both figures show annual changes in benchmark weights. A. Distance from the Average Change in Benchmark Weights 16 14 12 10 Mean= 5.7% Std. Dev.= 6.6% in % 8 25th Percentile= 1.7% 50th Percentile= 3.7% 6 75th Percentile= 7.4% 4 2 0 0 5 10 15 20 25 30 Percentage Points B. Changes vs. Changes: Ellipse Encircling 90 Percent of the Observations 60 % Change in Benchmark A's Weight 40 20 0 -60 -40 -20 0 20 40 60 -20 % of observations in -40 quadrants II and IV: 20% -60 % Change in Benchmark B's Weight Figure 3 The Upgrade of Israel from Emerging to Developed Market This figure shows the mean mutual fund and the benchmark weight around the upgrade of Israel from emerging to developed market in the MSCI indexes in May 2010. The mean weight in Israel is the weighted (by TNAs) average across funds for each type of fund. The left panels show the funds following the MSCI Emerging Markets index. The right panels show the funds following the MSCI World index. In each case we include the correspondent benchmark weight (MSCI Emerging Markets or MSCI World). The first grey bar indicates the month of the announcement and the second grey bar indicates the month the upgrade took place. A. Global Emerging Funds and MSCI Emerging Markets Index B. Global Funds and MSCI World Index Explicit Indexing Explicit Indexing 5 0.6 0.5 4 Effective Date Weight in % 0.4 3 0.3 2 0.2 Announcement 1 0.1 0 0.0 C. Global Emerging Funds and MSCI Emerging Markets Index D. Global Funds and MSCI World Index Truly Active Truly Active 5 0.6 4 0.5 Weight in % 0.4 3 0.3 2 0.2 1 0.1 0 0.0 Figure 4 Deviations in Mutual Fund and Benchmark Weights This figure shows scatter plots of the relation between mutual fund weights and benchmark weights for each country at each point in time. The panels show the scatter plots for explicit indexing (Panel A), closet indexing (Panel B), mildly active (Panel C), and truly active funds (Panel D). The vertical axis shows the mutual fund country weight for a certain benchmark minus the mutual fund average country weight across all the funds that invest in that country. The horizontal axis shows the benchmark weight of a country in a certain benchmark minus the average benchmark weight for the same country across all the benchmarks where the country is included. A. Explicit Indexing B. Closet Indexing C. Mildly Active D. Truly Active Figure 5 Direct Benchmark Effect: The Case of Israel This figure shows the direct benchmark effect using the example of the upgrade of Israel from emerging to developed market in the MSCI indexes. Panel A shows the capital flows over initial assets of explicit indexing funds during the month of the upgrade, May 2010. The direct benchmark effect is the response of country flows to an exogenous change in the benchmark weight. Total is the sum across the MSCI World Index ETF and the MSCI Emerging Markets Index ETF. Panel B illustrates, for the same month, the total capital flows from emerging market funds and developed market funds in levels. They include explicit indexing, closet indexing, mildly active and, truly active funds. A. Capital Flows over Initial Assets: Explicit Indexing Funds 0.5 0.35 0.0 -0.5 -1.0 in % -1.5 -2.0 -2.18 -2.5 -2.53 -3.0 MSCI Emerging Markets Index MSCI World Index ETF Total ETF ETF B. Capital Flows in Levels: All Types 500 159 0 Millions of U.S. Dollars -500 -1,000 -1,500 -2,000 -1,913 -2,073 -2,500 Emerging Markets Funds Developed Markets Funds Total Country Flows Figure 6 Sensitivity Effect of Country Flows This figure shows the sensitivity effect for Brazil (Panel A) and India (Panel B) in the MSCI Emerging Markets Index. Each panel displays the scatter plot of country flows against fund flows in millions of U.S. dollars for different equity funds for an initial date (in grey) and a later date (in black). The figure also shows the benchmark weight at each point in time. A. Brazil 30 Country Flows in Millions of U.S. Dollars 20 10 0 y = 0.06x + 1.29 R² = 0.35 -10 May 2006 Benchmark Weight= 10.5% y = 0.25x - 0.31 -20 R² = 0.34 May 2008 Benchmark Weight= 16.9% -30 -40 -80 -60 -40 -20 0 20 40 60 80 May 2006 May 2008 B. India 100 Country Flows in Millions of U.S. Dollars 80 y = 0.08x + 1.98 60 R² = 0.25 Jun 2008 40 Benchmark Weight= 5.7% 20 0 -20 -40 -60 y = 0.17x + 0.96 R² = 0.64 -80 Jun 2010 Benchmark Weight= 8.4% -100 -600 -400 -200 0 200 400 600 Jun 2008 Jun 2010 Fund Flows in Millions of U.S. Dollars Figure 7 Capital Flows and Benchmark Weights This figure shows the relation between mutual fund country flows and benchmark weights during periods of injections to mutual funds (pre-crisis) and redemptions from mutual funds (crisis and post-crisis). Panel A shows the case of the MSCI Emerging Markets Index ETF and Panel B the case of the MSCI Europe, Australasia, and Far East Index ETF. Periods are defined to be the peak (left) or trough (right) of flows over total assets for each fund. For the MSCI Emerging Markets Index ETF these periods are May 2003-May 2004 and September 2010-September 2011, respectively. For the MSCI Europe, Australasia, and Far East Index ETF these periods are February 2006-February 2007 and December 2008-December 2009, respectively. On the left axis the figure shows the inflows (outflows) to the respective countries during the corresponding period and in the right axis the benchmark weight at the end of the period. A. MSCI Emerging Markets Index ETF Pre-Crisis Period Global Financial Crisis and Post-Crisis Period 160 8 1,800 16.8 20 Flows in Millions of U.S. Dollars 143.5 7.1 140 7 15 Benchmark Weight in % 1,200 120 6 6.4 10 96.3 4.9 600 100 5 5 80 4 0 0 60 3 -5 -600 40 2 -537.3 -10 -1,200 20 1 -15 -1,352.3 0 0 -1,800 -20 China China Russia Russia China China Russia Russia B. MSCI Europe, Australasia, and Far East Index ETF Pre-Crisis Period Global Financial Crisis and Post-Crisis Period 400 368.3 4.5 500 10 Flows in Millions of U.S. Dollars 4.0 350 4 400 8 Benchmark Weight in % 3.5 300 4.6 6 300 3 200 4 250 100 0.3 2 2.5 200 0 0 2 150 -100 -25.2 -2 1.5 -200 -4 100 81.2 0.9 1 -300 -6 50 0.5 -400 -340.0 -8 0 0 -500 -10 Ireland Ireland Spain Spain Ireland Ireland Spain Spain Table 1 Mutual Fund Summary Statistics This table shows summary statistics of equity and bond mutual funds from the joint Morningstar Direct/EPFR database. Funds are divided by degree of activism, type of fund, and according to the country in which the fund is based (domicile). When divided by domicile the category Others includes Andorra, Australia, Austria, Bahrain, Bermuda, British Virgin Islands, Cayman Islands, Estonia, Finland, Germany, Greece, Guernsey, Hong Kong, India, Isle of Man, Israel, Italy, Japan, Jersey, Liechtenstein, Lithuania, Mauritius, Netherlands, Netherlands Antilles, Norway, Portugal, Singapore, Slovenia, South Africa, South Korea, Spain, Sweden, Switzerland, the United Arab Emirates, and funds with unassigned domicile. A. Summary Statistics Number of Funds Number of Observations First Available Last Available Median Observations Type of Fund (Fund-Month) Date Date per Fund (Months) Equity 2,837 156,253 January 1996 July 2012 70 Bond 838 35,219 March 1997 June 2012 54 B. Number of Funds and Observations by Different Attributes Number of Funds Number of Observations Type of Fund Number of Funds Number of Observations Degree of Activism (Fund-Month) (Fund-Month) Equity Funds Explicit Indexing 85 3,420 Global 569 29,037 Closet Indexing 939 50,906 Global Emerging 594 32,950 Mildly Active 994 58,960 Regional 1,674 94,266 Truly Active 819 42,967 Bond Funds Explicit Indexing 21 588 Global 554 22,958 Closet Indexing 54 2,851 Global Emerging 220 8,568 Mildly Active 714 29,768 Regional 64 3,693 Truly Active 49 2,012 C. Number of Funds and Observations by Domicile Number of Funds Number of Observations Domicile Number of Funds Number of Observations Domicile (Fund-Month) (Fund-Month) Equity Funds Belgium 51 2,495 Luxembourg 348 22,360 Canada 349 22,225 United Kingdom 225 16,615 Denmark 85 4,995 United States 495 25,887 France 158 6,206 Others 917 44,588 Ireland 209 10,882 Bond Funds Denmark 40 2,002 Luxembourg 31 1,700 Germany 35 1,421 United Kingdom 36 2,008 Ireland 56 2,314 United States 85 4,725 Israel 43 1,367 Others 479 18,720 Italy 33 953 Table 2 Benchmarks: Extensive and Intensive Margins This table shows the results of the determinants of the extensive and intensive margin for equity benchmarks. Panel A presents the marginal effects of Probit regressions of the extensive margin for equity benchmarks on different variables. Column 1 reports results of countries that are indexed by MSCI (category equal to one) versus countries that are not indexed (category equal to zero). Column 2 displays the same estimation but for non-index plus frontier countries versus emerging and developed countries for MSCI. Panel B reports Ordered Probit estimations for a categorical dependent variable containing four groups: non-index countries, frontier markets (FM), emerging markets (EM), and developed markets (DM). Both Panel A and B present cross-sectional regressions. Weights are computed as the average of the December weights across years for each country-benchmark combination. Panel C presents the results of ordinary least squares (OLS) regressions of the log country weights for equity and bond benchmarks on different variables. Only the intensive margin is considered for each benchmark (zero weights are not considered). Log free float market-capitalization is the log of the World Bank WDI equity market capitalization weighted by the free-float variable in Dahlquist et al. (2003). GDP PPP is the nominal GDP PPP from the WEO IMF Statistics. Country Risk is the country risk composite from ICRG, quality of institutions is the variable polity2 from Polity Database, and capital account openness is the Chinn-Ito de jure index for capital account openness. Standard errors are in parenthesis and clustered at the country level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. A. Probit Estimations B. Ordered Probit C. OLS Estimations Non- Estimations All Categories Non-Index vs. Index+FM vs. (Non-Index, FM, Dependent Variable: Log Country Weights Index Explanatory Variables EM+DM EM, DM) Equity Benchmarks Bond Benchmarks Log Market Cap. 0.019 0.094 *** 0.330 * 0.901 *** (0.035) (0.033) (0.170) (0.090) Log Free-Float Market Cap. 0.736 *** (0.099) Log GDP PPP 0.152 *** 0.095 *** 0.796 *** 0.958 *** 0.945 *** 0.709 *** (0.018) (0.026) (0.139) (0.080) (0.093) (0.091) Log GDP PPP per Capita -0.026 0.078 * 0.324 0.490 *** 0.525 0.55 ** (0.039) (0.047) (0.229) (0.129) (0.156) (0.256) Country Risk 0.018 *** -0.003 0.053 * 0.023 0.010 -0.047 (0.005) (0.006) (0.028) (0.016) (0.021) (0.031) Quality of Institutions 0.019 *** 0.007 0.087 *** 0.046 *** 0.024 * 0.025 (0.004) (0.007) (0.026) (0.012) (0.014) (0.032) Capital Account Openness -0.031 -0.038 0.050 0.055 0.090 0.168 (0.036) (0.029) (0.132) (0.082) (0.089) (0.111) Benchmark Fixed Effects No No No Yes Yes Yes Number of Observations 85 85 85 713 660 94 Pseudo R-Squared/R-Squared 0.784 0.505 0.447 0.880 0.876 0.611 Table 3 Log Difference Country Benchmark Weights This table shows the results of ordinary least squares regressions of the log difference of country benchmark weights on relative returns. Panel A shows results for equity benchmarks and Panel B for bond benchmarks. Relative returns are the difference between country net returns and benchmark net returns, expressed as decimals. Estimations are performed at different frequencies and include different combinations of fixed effects. Only countries in the benchmark are considered for each estimation. Standard errors are in parenthesis and clustered at the benchmark-time level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. Dependent Variable: Log Difference Country Benchmark Weights Explanatory Variables Monthly Semiannual Annual Biannual A. Equity Benchmarks Relative Returns 0.959 *** 0.960 *** 0.960 *** 0.961 *** 0.932 *** 0.865 *** 0.830 *** 0.760 *** (0.006) (0.006) (0.006) (0.006) (0.020) (0.017) (0.014) (0.013) Benchmark-Time Fixed Effects No Yes No Yes No No No No Benchmark-Country Fixed Effects No No Yes Yes Yes Yes Yes Yes Country-Time Fixed Effects No No No No Yes Yes Yes Yes Number of Observations 98,549 98,549 98,549 98,549 98,549 93,704 88,751 79,687 R-Squared 0.307 0.366 0.321 0.379 0.600 0.665 0.766 0.900 B. Bond Benchmarks Relative Returns 1.024 *** 1.022 *** 1.028 *** 1.027 *** 0.731 *** 1.065 *** 1.444 *** 1.778 *** (0.035) (0.032) (0.034) (0.031) (0.020) (0.160) (0.143) (0.126) Benchmark-Time Fixed Effects No Yes No Yes No No No No Benchmark-Country Fixed Effects No No Yes Yes Yes Yes Yes Yes Country-Time Fixed Effects No No No No Yes Yes Yes Yes Number of Observations 10,076 10,076 10,076 10,076 10,076 9,430 8,689 7,331 R-Squared 0.184 0.204 0.204 0.224 0.915 0.941 0.958 0.970 Table 4 Weights vs. Benchmark Weights This table presents OLS regressions of mutual fund country weights against benchmark country weights with different sets of fixed effects. Panel A displays results for equity funds and Panel B for bond funds. Funds are divided by degree of activism. Results are presented in levels and in differences. Estimations in levels do not contain observations where both weights and benchmark weights are zero. Standard errors are in parenthesis and clustered at the benchmark-time level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. Degree of Activism Total Sample Explicit Closet Mildly Truly Active Explanatory Variables Indexing Indexing Active A. Equity Funds Dependent Variable: Weights Benchmark Weights 0.823 *** 0.971 *** 0.961 *** 0.867 *** 0.598 *** (0.002) (0.003) (0.002) (0.002) (0.004) Fund-Country Fixed Effects No No No No No Fund-Time Fixed Effects No No No No No Number of Observations 2,524,798 42,029 577,241 988,198 917,330 R-Squared 0.725 0.975 0.941 0.817 0.401 Dependent Variable: Weights Benchmark Weights 0.773 *** 0.921 *** 0.919 *** 0.819 *** 0.499 *** (0.008) (0.013) (0.011) (0.010) (0.009) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 2,524,798 42,029 577,241 988,198 917,330 R-Squared 0.912 0.989 0.966 0.907 0.842 Dependent Variable: Changes in Weights Changes in Benchmark Weights 0.679 *** 0.792 *** 0.787 *** 0.726 *** 0.522 *** (0.011) (0.016) (0.015) (0.014) (0.011) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 2,166,004 35,647 483,721 858,626 788,010 R-Squared 0.113 0.481 0.162 0.108 0.089 B. Bond Funds Dependent Variable: Weights Benchmark Weights 0.714 *** 0.671 *** 0.876 *** 0.750 *** 0.392 *** (0.006) (0.008) (0.006) (0.007) (0.013) Fund-Country Fixed Effects No No No No No Fund-Time Fixed Effects No No No No No Number of Observations 153,402 723 57,338 57,335 38,006 R-Squared 0.360 0.863 0.679 0.419 0.077 Dependent Variable: Weights Benchmark Weights 0.697 *** 0.424 *** 0.935 *** 0.843 *** 0.223 *** (0.022) (0.032) (0.015) (0.023) (0.040) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 153,402 723 57,338 57,335 38,006 R-Squared 0.750 0.991 0.834 0.741 0.689 Dependent Variable: Changes in Weights Changes in Benchmark Weights 0.517 *** 0.347 *** 0.576 *** 0.499 *** 0.421 *** (0.038) (0.054) (0.053) (0.047) (0.102) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 77,386 635 32,409 29,076 15,266 R-Squared 0.156 0.241 0.116 0.142 0.196 Table 5 Log Weights vs. Log Benchmark Weights This table presents OLS regressions of log mutual fund country weights against log benchmark country weights with different sets of fixed effects. Panel A displays results for equity funds and Panel B for bond funds. Funds are divided by degree of activism. Results are presented in logs and in log differences. Standard errors are in parenthesis and clustered at the benchmark-time level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. Degree of Activism Total Sample Explicit Closet Mildly Truly Active Explanatory Variables Indexing Indexing Active A. Equity Funds Dependent Variable: Log Weights Log Benchmark Weights 0.821 *** 0.974 *** 0.969 *** 0.823 *** 0.662 *** (0.002) (0.002) (0.002) (0.003) (0.002) Fund-Country Fixed Effects No No No No No Fund-Time Fixed Effects No No No No No Number of Observations 1,597,666 36,562 447,185 631,546 482,373 R-Squared 0.601 0.950 0.785 0.617 0.399 Dependent Variable: Log Weights Log Benchmark Weights 0.747 *** 0.970 *** 0.904 *** 0.754 *** 0.575 *** (0.007) (0.011) (0.009) (0.008) (0.007) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 1,597,666 36,562 447,185 631,546 482,373 R-Squared 0.851 0.982 0.899 0.823 0.825 Dependent Variable: Changes in Log Weights Changes in Log Benchmark Weights 0.829 *** 0.831 *** 0.898 *** 0.873 *** 0.707 *** (0.011) (0.013) (0.014) (0.015) (0.013) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 1,370,079 31,633 381,539 543,969 412,938 R-Squared 0.155 0.447 0.165 0.144 0.162 B. Bond Funds Dependent Variable: Log Weights Log Benchmark Weights 0.761 *** 0.789 *** 0.900 *** 0.772 *** 0.423 *** (0.007) (0.005) (0.007) (0.007) (0.011) Fund-Country Fixed Effects No No No No No Fund-Time Fixed Effects No No No No No Number of Observations 87,687 676 36,611 33,034 17,366 R-Squared 0.394 0.838 0.610 0.421 0.093 Dependent Variable: Log Weights Log Benchmark Weights 0.585 *** 0.640 *** 0.753 *** 0.590 *** 0.229 *** (0.019) (0.043) (0.025) (0.022) (0.033) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 87,687 676 36,611 33,034 17,366 R-Squared 0.777 0.990 0.809 0.763 0.747 Dependent Variable: Changes in Log Weights Changes in Log Benchmark Weights 0.480 *** 0.415 *** 0.558 *** 0.456 *** 0.339 *** (0.032) (0.122) (0.047) (0.041) (0.077) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 77,326 625 32,434 29,024 15,243 R-Squared 0.142 0.231 0.113 0.134 0.185 Table 6 Weights vs. Benchmark Weights, Controlling for Industry Weights This table presents OLS regressions of mutual fund country weights against benchmark country weights and industry weights at different frequencies with different sets of fixed effects. The industry weights are the median weight in a certain country at a certain point in time for different segments of the mutual funds industry. Panel A displays results for equity funds and Panel B for bond funds. Funds are divided by fund type and degree of activism. Standard errors are in parenthesis and clustered at the benchmark-time level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. Degree of Activism Total Sample Explicit Closet Mildly Truly Active Explanatory Variables Indexing Indexing Active A. Equity Funds Dependent Variable: Weights, Monthly Benchmark Weights 0.673 *** 0.846 *** 0.890 *** 0.648 *** 0.347 *** (0.011) (0.018) (0.012) (0.014) (0.011) Industry Weights 0.358 *** 0.196 *** 0.168 *** 0.444 *** 0.497 *** (0.011) (0.023) (0.017) (0.013) (0.011) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 2,524,798 42,029 577,241 988,198 917,330 R-Squared 0.914 0.989 0.967 0.910 0.845 Dependent Variable: Weights, Semiannual Benchmark Weights 0.711 *** 0.880 *** 0.906 *** 0.721 *** 0.389 *** (0.028) (0.039) (0.029) (0.036) (0.027) Industry Weights 0.365 *** 0.129 *** 0.239 *** 0.421 *** 0.499 *** (0.026) (0.045) (0.043) (0.033) (0.030) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 468,574 8,040 109,674 180,632 170,228 R-Squared 0.912 0.985 0.965 0.907 0.846 B. Bond Funds Dependent Variable: Weights, Monthly Benchmark Weights 0.369 *** 0.466 *** 0.552 *** 0.328 *** 0.027 (0.025) (0.075) (0.034) (0.035) (0.053) Industry Weights 0.378 *** 0.133 *** 0.349 *** 0.430 *** 0.348 *** (0.018) (0.030) (0.022) (0.026) (0.039) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 76,405 606 31,835 28,851 15,113 R-Squared 0.752 0.983 0.778 0.732 0.745 Dependent Variable: Weights, Semiannual Benchmark Weights 0.362 *** 0.445 *** 0.584 *** 0.224 *** 0.153 (0.055) (0.051) (0.079) (0.068) (0.142) Industry Weights 0.379 *** 0.167 *** 0.329 *** 0.500 *** 0.263 ** (0.046) (0.047) (0.057) (0.052) (0.113) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 13,075 96 5,419 4,975 2,585 R-Squared 0.766 0.984 0.791 0.729 0.781 Table 7 Weights vs. Benchmark Weights, Controlling for Macro Factors This table presents OLS regressions of mutual fund country weights against benchmark country weights and different macroeconomic control variables and sets of fixed effects. Panel A displays results for equity funds and Panel B for bond funds. Funds are divided by fund type and degree of activism. Macro variables include four-month lagged industrial production growth, two-month lagged inflation, exchange rate growth, and stock market returns. Standard errors are in parenthesis and clustered at the benchmark-time level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. Degree of Activism Total Sample Explicit Closet Mildly Truly Active Explanatory Variables Indexing Indexing Active A. Equity Funds Dependent Variable: Weights Benchmark Weights 0.792 *** 0.928 *** 0.902 *** 0.766 *** 0.582 *** (0.007) (0.010) (0.009) (0.010) (0.010) Macro Variables as Controls Yes Yes Yes Yes Yes Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Country-Time Fixed Effects No No No No No Number of Observations 1,164,590 26,558 321,412 464,292 352,328 R-Squared 0.943 0.997 0.976 0.929 0.898 Dependent Variable: Weights Benchmark Weights 0.743 *** 0.981 *** 0.928 *** 0.680 *** 0.423 *** (0.010) (0.018) (0.009) (0.017) (0.014) Macro Variables as Controls No No No No No Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects No No No No No Country-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 1,665,785 37,764 458,745 657,672 511,604 R-Squared 0.929 0.997 0.976 0.922 0.864 B. Bond Funds Dependent Variable: Weights Benchmark Weights 0.779 *** 0.529 *** 0.921 *** 0.804 *** 0.385 *** (0.021) (0.041) (0.021) (0.031) (0.047) Macro Variables as Controls Yes Yes Yes Yes Yes Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Country-Time Fixed Effects No No No No No Number of Observations 62,266 532 26,103 23,415 12,216 R-Squared 0.822 0.991 0.867 0.810 0.772 Dependent Variable: Weights Benchmark Weights 0.412 *** - 0.737 *** 0.053 0.718 *** (0.038) - (0.052) (0.050) (0.085) Macro Variables as Controls No - No No No Fund-Country Fixed Effects Yes - Yes Yes Yes Fund-Time Fixed Effects No - No No No Country-Time Fixed Effects Yes - Yes Yes Yes Number of Observations 88,918 - 37,132 33,577 17,533 R-Squared 0.770 - 0.849 0.780 0.726 Table 8 Weights vs. Benchmark Weights: Exogenous Events The top part of each panel in this table presents OLS regressions of mutual fund country weights against benchmark country weights and the residual between benchmark weights and buy-and-hold benchmark weights (exogenous component), with different sets of fixed effects. The bottom part shows regressions dividing the sample between no- event and exogenous event times. Exogenous event times are those beyond the 25th and 75th tails of the distribution of the sample during the months that MSCI revises the indexes. No-event times are observations within those tails plus all the months with no revisions. Panel A displays results for equity funds and Panel B for bond funds. Funds are divided by degree of activism. Results are presented in levels and in differences. Estimations in levels do not contain observations where both weights and benchmark weights are zero. Standard errors are in parenthesis and clustered at the benchmark-time level.*, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. Degree of Activism Total Sample Explicit Closet Mildly Truly Active Explanatory Variables Indexing Indexing Active A. Equity Funds Dependent Variable: Weights Buy-and-Hold Benchmark Weight 0.813 *** 0.924 *** 0.906 *** 0.799 *** 0.628 *** (0.007) (0.012) (0.009) (0.009) (0.010) Exogenous Component 0.494 *** 0.531 *** 0.587 *** 0.499 *** 0.326 *** (0.041) (0.085) (0.046) (0.054) (0.066) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 1,377,388 31,620 383,187 549,793 412,788 R-Squared 0.932 0.996 0.972 0.918 0.880 Dependent Variable: Weights, No-Event Times Benchmark Weights 0.872 *** 0.947 *** 0.988 *** 0.911 *** 0.658 *** (0.015) (0.020) (0.021) (0.020) (0.017) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 894,742 22,946 242,590 355,509 273,697 R-Squared 0.920 0.995 0.966 0.905 0.878 Dependent Variable: Weights, Exogenous Event Times Benchmark Weights 0.819 *** 0.917 *** 0.907 *** 0.799 *** 0.634 *** (0.008) (0.017) (0.010) (0.011) (0.012) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 688,513 13,557 201,128 271,583 202,245 R-Squared 0.935 0.995 0.970 0.918 0.888 B. Bond Funds Dependent Variable: Weights Buy-and-Hold Benchmark Weight 0.800 *** - 0.950 *** 0.837 *** 0.375 *** (0.019) - (0.018) (0.029) (0.043) Exogenous Component 0.731 *** - 0.927 *** 0.742 *** 0.240 ** (0.052) - (0.052) (0.074) (0.117) Fund-Country Fixed Effects Yes - Yes Yes Yes Fund-Time Fixed Effects Yes - Yes Yes Yes Number of Observations 76,964 - 32,043 28,861 15,420 R-Squared 0.799 - 0.850 0.781 0.754 Dependent Variable: Weights, No-Event Times Benchmark Weights 0.780 *** - 0.948 *** 0.821 *** 0.343 *** (0.026) - (0.031) (0.036) (0.062) Fund-Country Fixed Effects Yes - Yes Yes Yes Fund-Time Fixed Effects Yes - Yes Yes Yes Number of Observations 48,833 - 18,369 19,588 10,362 R-Squared 0.804 - 0.864 0.795 0.753 Dependent Variable: Weights, Exogenous Event Times Benchmark Weights 0.847 *** - 0.952 *** 0.895 *** 0.457 *** (0.030) - (0.028) (0.047) (0.077) Fund-Country Fixed Effects Yes - Yes Yes Yes Fund-Time Fixed Effects Yes - Yes Yes Yes Number of Observations 38,469 - 18,131 13,275 6,901 R-Squared 0.801 - 0.843 0.776 0.773 Table 9 Weights vs. Benchmark Weights: Exogenous MSCI Free-Float Event This table presents in Panel A the distribution of changes in the exogenous component for December 2001 and June 2002 (exogenous MSCE free-float event) versus the rest of the sample during 2000-2002 (normal times). Panel B reports OLS regressions for equity funds of the changes in mutual fund country weights against the changes in buy-and- hold benchmark weights and the changes in the exogenous component, with different sets of fixed effects. The estimations are only for December 2001-June 2002, when MSCI conducted changes in the construction of its equity indexes. Funds are divided by fund type and degree of activism. Standard errors are in parenthesis and clustered at the benchmark-time level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. A. Exogenous MSCI Event vs. Normal Times 30 Std. Dev.= 0.7% 25 Interquartile Range= 0.4% Kurtosis= 6.8% 20 15 10 5 0 -5 -4 -3 -2 -1 0 1 2 3 4 5 Normal Times 30 Std. Dev.= 1.6% Interquartile Range= 1.3% 25 Kurtosis= 8.01% 20 15 10 5 0 -5 -4 -3 -2 -1 0 1 2 3 4 5 Percentage points Exogenous Event MSCI Degree of Activism Total Sample Explicit Closet Mildly Truly Explanatory Variables Indexing Indexing Active Active B. Equity Funds Dependent Variable: Changes in Weights Changes in Buy-and-Hold Benchmark Weight 0.707 *** - 0.837 *** 0.709 *** 0.644 *** (0.093) - (0.116) (0.217) (0.182) Changes in Exogenous Component 0.904 *** - 1.081 *** 1.022 *** 0.483 *** (0.248) - (0.303) (0.367) (0.118) Fund-Country Fixed Effects Yes - Yes Yes Yes Fund-Time Fixed Effects Yes - Yes Yes Yes Number of Observations 3,387 - 934 1,553 885 R-Squared 0.701 - 0.665 0.717 0.739 Table 10 Country Flows vs. Benchmark Flows This table presents OLS regressions of country flows in billions of U.S. dollars against benchmark weights multiplied fund flows with different set of fixed effects. Panel A displays results for equity funds and Panel B for bond funds. Funds are divided by fund type and degree of activism. Explicit indexing bond funds are not included due to the low number of observations. Standard errors are in parentheses and clustered at the benchmark-time level. *, **, and *** denote statistical significance at 10, 5, and 1 percent, respectively. Degree of Activism Total Sample Explicit Closet Mildly Truly Active Explanatory Variables Indexing Indexing Active A. Equity Funds Dependent Variable: Country Flows Benchmark Weight*Fund Flows 0.744 *** 0.839 *** 0.690 *** 0.547 *** 0.407 *** (0.028) (0.036) (0.014) (0.014) (0.017) Fund-Country Fixed Effects No No No No No Fund-Time Fixed Effects No No No No No Country-Time Fixed Effects No No No No No Number of Observations 962,344 12,895 286,890 378,626 283,933 R-Squared 0.296 0.627 0.177 0.081 0.045 Dependent Variable: Country Flows Benchmark Weight*Fund Flows 0.700 *** 0.794 *** 0.644 *** 0.468 *** 0.254 *** (0.035) (0.043) (0.018) (0.018) (0.018) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects Yes Yes Yes Yes Yes Country-Time Fixed Effects No No No No No Number of Observations 962,344 12,895 286,890 378,626 283,933 R-Squared 0.410 0.700 0.299 0.192 0.214 Dependent Variable: Country Flows Benchmark Weight*Fund Flows 0.739 *** 0.854 *** 0.676 *** 0.532 *** 0.381 *** (0.031) (0.045) (0.013) (0.015) (0.016) Fund-Country Fixed Effects Yes Yes Yes Yes Yes Fund-Time Fixed Effects No No No No No Country-Time Fixed Effects Yes Yes Yes Yes Yes Number of Observations 960,928 12,895 285,897 378,101 284,035 R-Squared 0.331 0.770 0.213 0.132 0.130 B. Bond Funds Dependent Variable: Country Flows Benchmark Weight*Fund Flows 0.634 *** - 0.730 *** 0.610 *** 0.615 *** (0.036) - (0.036) (0.043) (0.082) Fund-Country Fixed Effects No - No No No Fund-Time Fixed Effects No - No No No Country-Time Fixed Effects No - No No No Number of Observations 59,415 - 25,327 23,440 10,648 R-Squared 0.066 - 0.099 0.068 0.049 Dependent Variable: Country Flows Benchmark Weight*Fund Flows 0.369 *** - 0.683 *** 0.371 *** 0.120 (0.051) - (0.053) (0.065) (0.113) Fund-Country Fixed Effects Yes - Yes Yes Yes Fund-Time Fixed Effects Yes - Yes Yes Yes Country-Time Fixed Effects No - No No No Number of Observations 59,415 - 25,327 23,440 10,648 R-Squared 0.251 - 0.236 0.236 0.274 Dependent Variable: Country Flows Benchmark Weight*Fund Flows 0.551 *** - 0.748 *** 0.586 *** 0.585 *** (0.045) - (0.035) (0.050) (0.101) Fund-Country Fixed Effects Yes - Yes Yes Yes Fund-Time Fixed Effects No - No No No Country-Time Fixed Effects Yes - Yes Yes Yes Number of Observations 59,773 - 25,327 23,440 10,648 R-Squared 0.147 - 0.242 0.186 0.230 Table 11 Quantitative Benchmark Effects on Capital Flows This table presents the calibration of each of the effects presented in Section 5. Parameters are calibrated according to the median values in our sample. Panel A presents the calibration for each parameter and Panel B displays the quantitative benchmark effects for shocks on different variables. A. Calibration Parameters α 0.8 γit 1.0 B wict 4.0 B wict-1 4.0 Rct 1.01 B Rit 1.01 B. Quantitative Effects Value Δfict Δ(fict/wict-1B) Shock (percentage points) (in %) Direct Benchmark Effect ΔwictB 1.5 1.212 30.3 Sensitivity Effect Δfit 4.0 0.128 3.2 Amplification Effect ΔRct 10.0 0.307 7.7 Contagion Effect ΔRc't 10.0 -0.307 -7.7 Appendix Table 1 List of Benchmarks Used This table presents the complete list of equity and bond benchmarks in our database using the following abbreviations: AC (All Country), EM (Emerging Markets), EAFE (Europe, Australasia, and Far East), EMU (European Monetary Union), and EMEA (Emerging Markets Europe, Middle East, and Africa). EMBI+, EMBI Global, and EMBI Global Diversified are bond benchmarks. Equity and Bond Benchmarks 25% MSCI Brazil+25% MSCI Russia+25% MSCI India+25% MSCI China MSCI AC ASIA Pacific MSCI Europe Small Cap 50% MSCI AC Far East + 50% MSCI AC Far East ex-Japan MSCI AC Asia Pacific Ex-Japan MSCI Frontier Markets 50% MSCI Japan + 50% MSCI AC Asia-Pacific Free ex-Japan MSCI AC Europe MSCI GCC Ex Saudi Arabia 60% MSCI AC Asia Pacific ex-Japan + 40% MSCI Japan MSCI AC Far East MSCI Pacific 75% MSCI AC Far East Free ex-Japan + 25% MSCI Japan MSCI AC Far East Ex-Japan MSCI Pacific Ex-Japan 75% MSCI Arabian Markets ex Saudi Arabia + 25% MSCI Saudi Arabian Domestic MSCI AC Pacific MSCI World 87% MSCI Eastern Europe + 13% MSCI Russia MSCI AC Pacific Ex-Japan MSCI World Small Cap Citigroup World Ex-US Extended MSCI AC World S&P Asia 50 TR DJ Asia Pac Select Dividend 30 MSCI AC World Ex-US S&P BRIC 40 DJ Asia Pacific Selected Div 30 MSCI AC World Investable Mkt S&P Citi BMI Emerging Markets DJ Asian Titans MSCI Arabian Markets Ex-Saudi Arabia S&P Citi BMI European Em Capped DJ Global Titans 50 MSCI BRIC S&P Citi EM EPAC Euro Stoxx MSCI EAFE S&P Citi EMI Global Euro STOXX 50 MSCI EAFE Small Cap S&P Citi PMI Eurozone Growth FTSE AW Eastern Europe MSCI EM Asia S&P Citi PMI World Value FTSE RAFI Emerging Markets MSCI EM Eastern Europe S&P Europe 350 FTSE World MSCI EM Eastern Europe ex Russia S&P Global 100 FTSE World Asia Pacific MSCI EM EMEA S&P IFC Investable FTSE World Eurobloc MSCI EM Europe S&P IFC Investable Composite FTSE World Europe MSCI EM Far East S&P IFCG Asia FTSE World Europe ex-UK MSCI EM Latin America S&P IFCG Latin America FTSE World Pacific ex-Japan MSCI Emerging Markets S&P IFCG Middle East & Africa J.P. Morgan EMBI Global MSCI Emerging Markets Europe+Middle East S&P IFCI Composite J.P. Morgan EMBI Global Diversified MSCI EMU S&P IFC Investable Latin America J.P. Morgan EMBI+ MSCI Europe S&P IFCI Latin America MSCI AC Asia Ex-Japan MSCI Europe Ex-UK S&P Latin America 40