The Geography of Mutual Funds: The Advantage of Distant Investors

Transcription

1 The Geography of Mutual Funds: The Advantage of Distant Investors Miguel A. Ferreira * NOVA School of Business and Economics Massimo Massa INSEAD Pedro Matos University of Virginia Darden School of Business This Version: October 2011 Abstract We test the hypothesis that funds with clients located far away from the assets in which the fund invests higher fund client-stock distance (CSD) have a lower correlation between flows and performance. Funds with higher CSD invest in assets when the other locally based players are divesting (and therefore prices are low) and sell assets when the other locally based players are investing (and therefore prices are high). CSD thus acts as a sort of insurance that allows funds to take more risk and enjoy better performance. Using a worldwide sample of equity mutual funds, we find that CSD is positively associated with fund risk-taking behavior and performance. A one-standard deviation increase in CSD is related to a 28 to 46 basis point improvement in fund performance (per quarter). Higher CSD allows funds to increase their stock selectivity and tilt their portfolios toward stocks that are less liquid. Our results suggest that the location of mutual fund clients affects fund risk-taking behavior and performance. JEL classification: G23, G30, G32 Keywords: Mutual funds, Performance, Proximity investment, Co-movement * NOVA School of Business and Economics, Faculdade de Economia, Campus de Campolide Lisboa, Portugal; miguel.ferreira@novasbe.pt. INSEAD, Boulevard de Constance, Fontainebleau Cedex, France; massimo.massa@insead.edu. University of Virginia Darden School of Businessm 100 Darden Blvd. Charlottesville, VA U.S.; matosp@darden.virginia.edu..

2 1. Introduction The academic literature has traditionally been skeptical about the ability of mutual funds to generate superior performance for their customers. Beginning with Jensen (1968), numerous authors find that average risk-adjusted mutual fund performance is zero if not negative. If markets are efficient, there is likely no systematic source of advantage for mutual funds. One source of competitive advantage that has received some attention is geographic proximity. Coval and Moskowitz (1999, 2001) link fund location with information, arguing that funds located closer to an asset i.e., nearer the company headquarters have better information about it. Internationally, geographic distance or national boundaries are important sources of information asymmetry. Domestic (closer) investors have better information than international (distant) ones. 1 For example, Choe, Kho, and Stulz (2005), Dvorak (2005), and Teo (2009) find that international funds perform more poorly than domestic funds. This suggests that proximity helps performance. One disadvantage of proximity that has received almost no attention in the literature is related to the comovement in shocks between fund investment and fund flows. Funds with clients located close to an asset are more subject to the shocks to its price, as investors may increase redemptions at the very time this asset is underperforming. Funds investing in an asset that is located farther away from its clients experience less correlation between their investor flows and the performance of the asset. This suggests that a negative correlation between flows and investment opportunities should positively affect performance. This effect is distinct from the proximity investment effect since being located away from the assets in which to invest improves performance, rather than hampering it. While proximity investment is based on information benefits, shock comovement is related to diversification benefits. Consider, for example, the Fidelity Europe Fund, which invests in European stocks but is sold mostly to U.S. investors. A lower correlation between the U.S. market and the European market implies that a bear market in Europe does not necessarily mean a bear market in the U.S. market. In this case, bullish U.S. investors will still invest in mutual funds when the more bearish European investors are pulling out. This means that the Fidelity Europe Fund will be in the 1 Brennan and Cao (1997) show that international funds react more to the release of new information about a stock, which is consistent with international funds being less informed than domestic funds. Local analysts also seem to have an informational advantage relative to foreign analysts (Bae, Stulz, and Tan (2008)). 1

3 position of investing in European stocks when Europeans are divesting and selling such stocks when Europeans invest. That is, the fact that Fidelity Europe Fund targets U.S. investors may allow it to buy when prices are low and sell when prices are high. This will let Fidelity Europe Fund outperform other funds investing in European stocks but located in Europe and selling mostly to European clients. For example, Fidelity Europe Fund will be able to outperform the AriDeka Fund, which invests in European stocks but is sold only to German, Belgian, and Austrian clients. This example suggests that fund location which defines its clients can be a source of strategic advantage that is not related to information but rather more directly to the ability to exploit investment opportunities created by limits to arbitrage (Shleifer and Vishny (1997)). This intuition is similar to the one proposed for hedge funds: Hedge funds take advantage of mutual fund investors fire sales. When mutual funds are forced to sell to meet redemption calls, hedge funds buy the assets liquidated at fire sale prices (Chen, Hanson, Hong, and Stein (2008)). The fact that mutual funds intermediate a way higher fraction of asset under management than hedge funds suggests that this effect represents a large-scale limits of arbitrage phenomenon, way more important for the economy. In this paper, we test whether this effect exists and its magnitude. We focus on a worldwide sample of equity mutual funds that includes both domestic and international funds. We examine the limit-to-arbitrage constraints as well as the benefits that location provides to mutual funds, testing whether the performance of domestic (nearby) funds is hurt by investor demand that is closely related to their investments. We propose a new hypothesis the correlation hypothesis that posits that geographic location matters in terms of the correlation between investor flows and the assets a fund invests in. If fund flows are very little or even negatively correlated with the performance of the assets the fund invests in, the fund can buy when an asset price is low and sell when the price increases. This eases the problem of limits of arbitrage by providing a sort of insurance for the fund. This hedge allows the fund to take more risk and invest in assets in good times, improving both fund investment timing and selectivity. To test this hypothesis, we construct several measures of distance between the location of investors in a fund and the location of the assets in which the fund invests the client-stock distance (CSD). The first CSD measure is the geographic distance between the markets in which 2

4 the fund sells its shares and the markets in which it invests its assets (client-stock geographic distance). The second measure is the correlation between changes in demand (i.e., fund flows) in the markets in which the fund sells shares and in the markets in which the fund invests (clientstock flow correlation). The third measure is the correlation between the returns in the markets in which the fund sells shares and the returns in the markets in which the fund invests (client-stock return correlation). These measures effectively proxy for the fact that the location of the investor demand may give the fund the ability to exploit opportunities provided by limits of arbitrage. While we test this hypothesis using a sample of worldwide mutual funds, the same intuition is more general and applies to domestic funds as well (e.g., funds operating only in the U.S.). Yet a test based on U.S. domestic funds would be almost impossible given the lack of disaggregated data on the regional client composition of investors in U.S. funds. Moreover, stock markets are more likely to be integrated domestically in the U.S. than across different countries. For these two reasons, international fund data provide us with a unique experimental ground to test our hypothesis. We start by investigating if there is any link between CSD and the actions of the fund manager. We focus on portfolio risk, liquidity, and trading strategies. The correlation hypothesis suggests that fund managers, protected by a more advantageous flow-investment relationship, should be able (and willing) to take more risk and invest in more illiquid stocks. This will allow them to have trading profiles that diverge from the average ones of the styles they belong to. We find that funds characterized by higher CSD display both higher total return volatility and higher systematic volatility. This finding is not only statistically robust for both client-stock geographic distance and client-stock correlation of flows or returns, but it is also economically meaningful. A one-standard deviation reduction in client-stock geographic distance is associated with a 20 basis points increase in total standard deviation (per quarter) and a 0.07 increase in systematic risk (i.e., market beta). A one-standard deviation reduction in client-stock return correlation is associated with a 43 basis points increase in total standard deviation (per quarter) and a 0.17 increase in systematic risk. This finding provides evidence in favor of the correlation hypothesis. Funds with greater client-stock distance take more risk, as CSD provides a sort of insurance or hedge for them. We then relate CSD to the degree of liquidity of portfolio holdings of the fund. We find a strong negative relationship between portfolio liquidity and CSD. Funds with higher CSD 3

5 display portfolio stock holdings with lower turnover, higher (Amihud) illiquidity, and lower firm size. These findings further support the correlation hypothesis and show that CSD provides a sort of insurance for the fund that allows it to invest in more illiquid and therefore more profitable stocks. These results also imply that fund managers with higher CSD should be able to have trading profiles that diverge from the average ones of the styles they belong to. We therefore study the types of trading strategies that mutual funds pursue. If CSD makes a fund more able to take risk, then risk taking can take different forms. Funds may either load up more on market risk or increase their idiosyncratic exposure. In the latter case, the funds will either invest in more risky stocks than their more closely located peers do or implement a dynamic trading strategy. One way to investigate this issue is to look at measures of active management and stock selection. Following Wermers (2003) and Amihud and Goyenko (2009), we focus on the fund tracking error and the ability of the benchmark-factor model (R-square) to explain performance. The idea is that higher tracking error or lower explanatory power of the factor model signal that the fund follows a conventional or passive investment strategy. We find a strong positive correlation between CSD and tracking error, and a strong negative correlation between CSD and R-square. The effect is also economically significant. A onestandard deviation increase in client-stock geographic distance is related to a 61 basis points (per quarter) higher tracking error and 3.8 percentage points lower R-square. A one-standard deviation reduction in client-stock return correlation is related to a 103 basis points (per quarter) higher tracking error and a 7 percentage points lower R-square. These results suggest that CSD lets funds be more selective in choosing assets. Funds with lower CSD are more constrained in tracking their benchmark. These results show that CSD provides a source of strategic advantage that allows fund manager take more risk and invest in more illiquid assets. We therefore expect CSD to improve fund performance. We investigate this hypothesis by looking at how CSD relates to fund performance. We show that the higher the CSD, the better on average the performance of the fund. This result is strong, and holds for the different measures of client-stock distance. The result is also economically significant. A one-standard deviation increase in client-stock geographic distance is associated with a 46 basis points improvement (per quarter) in performance (four-factor Carhart (1997) alpha), while a one-standard deviation reduction in 4

6 client-stock return correlation is associated with a 28 basis points improvement (per quarter) in performance. This suggests that the natural hedge provided by a low correlation between flows and investment returns allows funds to pursue riskier and more profitable stock selection strategies. One potential concern is that our proxy of client-stock geographic distance may be related to fund-stock geographic distance. We control for the distance between the location of the fund and the location of the assets in which the fund invests (fund-stock geographic distance) throughout the analysis. This allows us to control for the information advantage of local investors. The fact that the results hold suggests that the CSD effect enhances performance above and beyond the traditional location-based effect. Our results are stronger for international funds and funds domiciled outside of the U.S., as these type of funds more frequently invest and source in several countries. The results are robust whether we control or not for fund-stock geographic distance. We contribute to three different strands of literature. First, we contribute to the literature on mutual fund performance (Grinblatt and Titman (1989, 1994), Elton, Gruber, Das, and Hlavka (1993), Brown and Goetzmann (1995, 1997), Elton, Gruber, and Blake (1996), Ferson and Schadt (1996), and Wermers (2000)). The literature has mostly focused on exogenous variations in fund characteristics that help to explain performance. For example, Edelen (1999) shows the negative effects of flow-induced trading but stops short of exploring the location of a mutual fund and its investors. We contribute to this literature by providing evidence on how an exogenous source of variation client location directly impacts performance, potentially providing the fund with a source of strategic advantage. Given that the geographic area in which a mutual fund sells its shares affects fund performance, marketing activities and the degree of market penetration become key factors in affecting the ability of a fund to outperform. This effect is not related to just making outflows more stable, but extends to coordinating outflows and portfolio holdings returns. Second, our findings add to the growing literature on the importance of geography in portfolio investment. Some authors have analyzed the location of an investment firm (relative to fund location) as a source of performance, arguing that closely related managers have access to better information and therefore deliver better performance (Coval and Moskowitz (1999, 2001)). We complement this literature by analyzing a different channel: client location. Our 5

7 results show that distant mutual funds may perform better, contrary to the finding in the standard literature on proximity investing. 2 Finally, our results contribute to the literature on the limits of arbitrage (e.g., Shleifer and Vishny (1997)) and to the literature on mutual fund demand (e.g., Sirri and Tufano (1998), Zheng (1999)). We investigate how the location of investor demand may affect fund performance. Moreover, we provide a new test of the limits of arbitrage. Location and customer demand provide an ideal source of segmentation that may allow some funds to gain an edge over others. Our results quantify the importance of limits of arbitrage across markets internationally and on classes of assets that are liquid. The remainder of the paper is organized as follows. Section 2 describes our hypotheses and the empirical approach. Section 3 discusses the data. Section 4 examines how client-stock affects the flow-performance relationship. Sections 5 and 6 present the empirical results on the effect of client-stock distance on fund risk-taking behavior, performance, and trading strategies. Section 7 presents robustness checks. Section 8 concludes. 2. Hypotheses Investors can be constrained in their ability to hold assets for a long period. The shorter the investment horizon i.e., the less able an investor is to hold assets for a long period the less able a fund is to exploit profitable but more risky and illiquid investment opportunities. We consider what happens if a fund, all else equal and given the same level of flows, receives inflows at a time it has better investment opportunities, while outflows induce it to sell when market conditions allow it to reap a higher capital gain. We define this correlation between flows into and out of a fund and the performance of the stocks in which the fund invests as CSD. The more negative the correlation between the performance of the markets in which the fund invests and the markets where fund shares are sold (or sourced), the higher is the degree of CSD. CSD effectively equips the fund with a sort of insurance or natural hedge that allows it to take more risky and illiquid positions. This suggests our first hypothesis: 2 In fact, our results are consistent with the findings showing that foreign investors do actually overperform the local ones (Grinblatt and Keloharju (2000), Froot and Ramadorai (2008)).Some other studies find no difference between the performance of local and foreign investors (Kang and Stulz (1997), Froot, O Connell, and Seasholes (2001)). 6

8 H1: Funds with higher client-stock distance (CSD) are able to take more risks and invest in less liquid stock. If funds are protected by the CSD hedge, they can invest in assets that the average fund of its style will not dare to venture into, and be more aggressive in stock selection. This will also change exposure to the benchmark. Indeed, funds with a higher degree of CSD should be better able to diverge from the benchmark successfully. This suggests our third hypothesis: H2: Funds with higher client-stock distance (CSD) are more active - i.e., more able to deviate from their investment benchmarks. What are the implications in terms of performance? If CSD allows a fund to invest in better investment opportunities, this should translate into better performance. Given that this is not a source of performance that can be matched by funds that do not enjoy the same CSD, this should generate abnormal performance that cannot be easily arbitraged away. This suggests our second testable hypothesis: H3: Funds with higher client-stock distance (CSD) have better abnormal performance. We test these hypotheses using a worldwide sample of equity mutual funds that includes information about funds investor location. 3. Data and Variables Construction We draw data on open-end equity mutual funds from the Lipper Hindsight database over We exclude index funds, exchange-traded funds, funds-of-funds, and off-shore funds from the sample. The database is survivorship bias-free, as it includes data on both active and defunct funds. Although Lipper lists multiple share classes as separate funds, classes have the same holdings, the same manager, and the same returns before expenses. We therefore eliminate multiple classes of the same fund to avoid multiple counting of returns. We keep the fund share class that Lipper identifies as the primary one. The initial sample includes 37,910 primary equity funds (both active and defunct funds). 3 We compare the coverage of funds by Lipper with aggregate statistics on mutual funds from from the Investment Company Institute (ICI). The total numbers of equity funds reported by 3 The primary fund is in general the class with the highest total net assets (TNA). The primary class typically represents more than 80% of the total assets across all share classes. 7

9 Lipper and ICI are, respectively, 26,800 and 26,950 as of December The total net assets of the equity funds (sum of all share classes) reported by Lipper and ICI are, respectively, $10.9 trillion and $12.5 trillion as of December Thus, our initial sample of equity funds covers 87% of the total net assets of worldwide equity funds. There is, however, some variation in coverage across countries. While Canada, Germany, Sweden, the U.K., and the U.S. have coverage above 90%, coverage in Australia and France is roughly 60% and in Japan only 40%. 4 We exclude off-shore funds because the location of their clients is not well defined. That is, the investors in off-shore funds are mostly international investors, so it is hard to define their actual location. Excluding off-shore funds e.g., funds domiciled in Luxembourg or Dublin closed-end funds, index-tracking funds, exchange-traded funds, and funds-of-funds reduces the sample to 25,110 open-end actively managed equity funds from 34 countries. We require mutual funds to have data on total net assets (TNA), age, total expense ratios, front-end and back-end loads, management team, countries where a fund is sold, and monthly total returns. We also require a fund to have at least 24 months of reported returns because we need to estimate fund factor loadings using past fund returns. The final sample includes 15,683 funds in 27 countries (11,944 active funds and 3,739 defunct funds as of December 2007). Table 1 presents the number and TNA of the sample of mutual funds by country as of December When there are multiple share classes, the TNA is the sum of the TNAs of all the share classes. The 11,944 equity funds manage $6.3 trillion of assets. U.S. funds represent 66% of the sample in terms of TNA, but only 19% of the total number of funds. The Lipper database provides information on a fund s country of domicile and geographic investment focus. We use these data to classify funds in terms of their geographic investment style: domestic funds i.e., funds that invest in their domicile country and international funds i.e., funds that invest in countries or regions different from the one where they are domiciled (foreign country funds and regional funds), and funds that invest worldwide (global funds). Domestic funds represent about half of the sample in terms of the number of funds and 62% in terms of TNA. The U.S. mutual fund industry is heavily tilted toward domestic funds, which account for more than 70% of the number of the funds and TNA in the U.S. International funds, however, are dominant in other countries like Australia, Canada, France, Germany, and the U.K. 4 There are 24,050 equity funds with TNA of $10.2 trillion in Lipper if we exclude closed-end funds and funds-offunds. In this case our initial sample covers 82% of the TNA of equity funds worldwide. The ICI statistics are not entirely consistent across countries on whether they include these types of funds. 8

10 3.1. Measuring (CSD) We use the information on the countries in which a fund is distributed to construct several proxies for the client-stock distance (CSD). The first measure is based on the geographic distance between the country in which the fund sells and the country in which the fund invests (client-stock geographic distance). For each fund, Lipper provides a list of countries the fund is legally authorized to sell in ( countries notified for sale ). The database also provides the countries of the stocks in which the fund invests ( geographic focus ). In the case of multiple countries or regions of investment, we weight the countries by their degree of stock market capitalization. For example, Fidelity Europe Fund, which invests in European stocks but is mostly sold to U.S. investors, has a high client-stock geographic distance according to the valueweighted average distance between the U.S capital (Washington, D.C.) and European capitals. Similarly, if a fund sells in multiple countries, we weight the countries by their market capitalization. For example, for the AriDeka Fund which invests in European stocks but is sold to German, Belgian, and Austrian clients, we measure the client-stock geographic distance as the value-weighted average distance between Berlin, Brussels, and Vienna and the European capitals. 5 The distance d i,j between fund client i and stock j is given by: where: deg latlon d i, j cos( lat ) cos( lon ) cos( lat i cos( lat ) sin( lon ) cos( lat i i i 2 r arccos(deglatlon) (1) 360 j j ) cos( lon ) sin( lon j j ) ) sin( lat ) sin( lat i j ) and lat and lon are the latitude and longitude of the capital city of the country, and r is the radius of the earth (approximately 6,378 km). We use the logarithm of one plus the client-stock geographic distance as explanatory variable in the tests since the client-stock geographic distance can take the value of zero. The second CSD measure is based on the correlation of returns (client-stock return correlation). More specifically, it is the correlation between the stock market returns of the countries where the fund invests (value-weighted average) and the stock market returns of the 5 We obtain consistent results if we use equal weights or if we weight the countries where a fund is sold by the population of the country or by its GDP. 9

11 country where the fund is sold to investors. The correlations are estimated using a 36-month rolling window of daily returns of the weighted average of stocks held in the portfolio of the fund and the return of the stock market index of the country where the fund is sold. The returns are denominated in local currencies. 6 If there is more than one country of sale, we calculate the return for this group as the market capitalization-weighted average of all the countries where the fund is sold. Alternatively, we calculate the client-stock return correlation using the actual portfolio holdings rather than using country-level weights for the countries of investment (client-stock return correlation holdings). The fund portfolio holdings are drawn from the Factset/Lionshares database. For more details on this database, see Ferreira and Matos (2008). These client-stock return correlation measures based on actual portfolio holdings are available for a smaller number of mutual funds and are potentially more subject to endogeneity as the holdings are chosen by the funds themselves. Another set of CSD measures is based on investor flows. In particular, we focus on the flows into the funds that invest in the same countries and sell to the same investors (client-stock flows correlation). We define flows as the percentage growth in total assets under management (in local currency) between the beginning and the end of quarter t, net of internal growth (assuming reinvestment of dividends and distributions): TNAi, t TNAi, t 1 (1 Ri, t ) Flow i, t (2) TNA i, t 1 where TNA i,t is total net assets in local currency of fund i, and R i,t is fund i return in the local currency. For each country, we aggregate the flows of all the funds selling in the country in the quarter. Then, we weigh these flows by the market capitalization of the segment in which the fund operates. This provides a segment-based measure of flows. Finally, we estimate the correlation between the measure of flows and returns using a 36-month rolling window Measuring Risk-Adjusted Performance The systematic risk component is estimated on the basis of both the market model and the Carhart (1997) four-factor model. We estimate four-factor alphas using regional factors (Asia, 6 We obtain similar findings using returns denominated in U.S. dollars. 10

12 Europe, and North America) based on a fund s investment region ( geographic focus ) in the case of domestic country funds, foreign country funds and regional funds or world factors in the case of global funds in the manner of Bekaert, Hodrick, and Zhang (2009). For each fund-month, we estimate the monthly factor loadings by running the regression: R MKT SMB HML MOM, (3) i, t i 1, i t 2, i t 3, i t 4, i t i, t where R i,t is the return in U.S. dollars of fund i in excess of the one-month U.S. Treasury bill rate in month t; MKT t is the excess return in U.S. dollars on the fund s investment region in month t; SMB t (small minus big) is the average return on the small-capitalization portfolio minus the average return on the large-capitalization portfolio on the fund s investment region; HML t (high minus low) is the difference in return between the portfolio with high book-to-market stocks and the portfolio with low book-to-market stocks on the fund s investment region; MOM t (momentum) is the difference in return between the portfolio with the past 12-month winners and the portfolio with the past 12-month losers on the fund s investment region. We first construct country-level factors MKT, SMB, HML and MOM using individual stock returns in U.S. dollars obtained from Datastream, following closely the method of Fama and French (1992). The regional and world factors are value-weighted averages of countries factors. 7 We use monthly fund returns (net of expenses) denominated in U.S. dollars from July 1997 through December 2007 to estimate the factor models. We estimate the time series regression equation (3) of the monthly fund excess returns on the risk factors using the previous 36 months of data at quarterly frequency. In particular, we proceed as follows. First, we estimate the factor loadings using the four-factor model in equation (3). To do this, we require a minimum of 24 months of return data. Since we use 36 months of return data to estimate the factor model, our first estimates of a fund s performance and risk starts at the first quarter of We then subtract the expected return from the realized fund return to estimate the fund s abnormal return in each quarter, or alpha, which is measured as a sum of an intercept of the model and the residual. Alpha measures the manager s contribution to performance due to stock selection or market timing. A positive alpha indicates that the fund overperforms the benchmark, a negative that it underperforms. 7 See Ferreira, Miguel, and Ramos (2010) for details about the construction of the factors. 11

13 3.3. Control Variables We now describe the control variables. Following Coval and Moskowitz (2001), we control for the geographic distance between the country in which a fund is located and the country in which it invests (fund-stock geographic distance), which proxies for the information asymmetry between the stock and the fund. The fund-stock distance is estimated in the same way as the stock-client distance, as defined above. We use the logarithm of one plus the fund-stock geographic distance as control variable in the tests since the fund-stock geographic distance can take the value of zero (e.g., in the case of a domestic fund). Additional control variables are the size of the fund, defined as the logarithm of assets under management or total net assets (TNA); the size of the family the fund belongs to, defined as the logarithm of assets under management of all the equity funds managed by the same management company; the age of the fund, defined as the number of years the fund has been active; the fund expense ratio and loads; and flows into the fund. We measure all fund characteristics quarterly. In the regression tests, we also control for time fixed effects (quarter dummies), fund country fixed effects, and investment region fixed effects (Africa, Asia-Pacific, Eastern Europe, Europe, Global, Latin America, and North America). We obtain consistent findings if we use investment region returns, rather than a fixed effect. We also obtain consistent findings if we include geographical investment style fixed effects (domestic country funds, foreign country funds, regional funds, and global funds) Summary Statistics Table 2 reports the means (time series averages of quarterly cross-sectional averages) and t- statistics for the variables of interest for all funds, and for funds in the bottom versus the top half of the client-stock geographic distance distribution. Table 2 also reports similarly constructed means for client-stock return correlation halves. Given that CSD is highly correlated with the fund-stock geographic distance, we use the residual of a regression of CSD on the fund-stock distance (i.e., the component of CSD orthogonal to the fund-stock distance) to classify funds based on CSD. In the multivariate tests, we use directly CSD as an explanatory variable since we also include fund-stock geographic distance as an explanatory variable. We obtain consistent findings if the use the CSD residual in the multivariate tests. 12

14 The results show that funds in the top half of client-stock geographic distance have better risk-adjusted performance and higher risk (total or systematic risk) than funds in the bottom half of client-stock geographic distance. We also find that funds in the bottom half of client-stock return correlation have better risk-adjusted performance and higher risk (total or systematic risk) than funds in the top half of client-stock geographic distance. These findings are consistent with the correlation hypothesis: Funds with high CSD both geographic distance-based and correlation-based take on more risk and have better performance. This is preliminary univariate evidence. In the next sections, we will control for the other fund characteristics. 4. Flow-Performance Relationship Results An underlying assumption of the correlation hypothesis is that funds investing in stocks that are located farther away should experience a lower correlation between their investor flows and the performance of the assets in which they invest. In particular, we expect a poor relative performance to be accompanied by cash inflows/outflows that are less sensitive to fund performance in the case of funds with high client-stock distance (CSD) than for funds with low client-stock distance. To test this hypothesis, we estimate the flow-performance sensitivity (e.g., Sirri and Tufano, 1998) and examine the cross-sectional relationship between fund flows and the fund s performance rank at the end of the previous quarter. We use a piecewise-linear specification, which allows for different flow-performance sensitivities at different levels of performance. We allow slopes to differ for the lowest quintile, middle three quintiles, and the top quintile. The slopes are estimated separately for the bottom quintile (Low), the three middle quintiles (Mid), and the top quintile (High) of the fractional fund performance ranks. In each quarter and for each country fractional fund performance, ranks ranging from zero (poorest performance) to one (best performance) are assigned to funds according to their performance in the prior year as measured by raw returns. The coefficients on these piecewise decompositions of fractional ranks represent the marginal fund-flow response to performance. We define the performance ranking variables for each of the three performance measures as: 13

15 Low min( 0.2, Ranki, 1) i, t 1 t Mid i, t 1 min( 0.6, Rank Lowi, t 1 ) (4) High Rank ( Low Mid 1) i, t 1 i, t 1 i, t To test whether the sensitivity of flows to past performance is statistically different for funds with high and low CDS, we interact Low, Mid and High with a dummy variable that takes the value of one if a fund has CSD above the median, and zero otherwise (CSD dummy). As a robustness check, we also report the results using the four-factor model alphas as a measure of performance to construct the ranks. We estimate panel regressions of quarterly fund flows on the piecewise past performance interacted with CSD as well as control variables. We use two alternative measures of CSD in this test: client-stock geographic distance and client-stock return correlation. All the explanatory variables are lagged one quarter. All the regressions include time fixed effects (quarter dummies) to account for cross-sectional dependence, fund country fixed effects, and investment region fixed effects. Standard errors are clustered at the fund level to account for autocorrelation in fund risk. We expect the flow-csd interaction variable coefficient to be negative in the case of the client-stock geographic distance and positive in the case of the client-stock return correlation. The results are reported in Table 3. We find a positive and significant relationship between fund flows and relative performance for the top performance quintile in line with Sirri and Tufano (1998). While the flow-performance relationship is strongest in the top performance quintile, there is also a positive and significant relationship in the mid performance quintiles, and especially, in the low performance quintiles. More importantly, as shown by the Low x CSD dummy coefficient, for the lowest performance quintile the sensitivity of the flow-performance relationship is significantly lower for funds with high client-stock geographic distance. In other words, the funds with a clientele farther away from the stocks in which they invest are less affected by cash outflows when the fund performs poorly. There is also evidence of lower sensitivity in the mid performance quintiles, but the decrease in sensitivity is much lower than in the low performance quintiles. There is no evidence that the sensitivity is lower in the top performance quintile as shown by the insignificant High x CSD dummy coefficient. The results using client-stock correlation as a measure of CSD are consistent with those using client-stock geographic distance. In particular, we find that the Low x CSD 14

16 dummy coefficient is positive and significant. This indicates that funds with low (high) clientstock correlation have flows that are less (more) sensitive to poor performance. These results provide supporting evidence for the assumption behind our main hypothesis: the distance between country of investment and country of sale directly affects the flowperformance sensitivity and therefore the ability of the fund to hold out for a longer period and be less exposed to the limit of arbitrage constraints. We now move on to test our main hypotheses on the implications for the policy of the fund manager. 5. Risk Taking Results In this section we investigate the link between CSD and fund manager actions. We focus on portfolio risk, illiquidity, and trading strategies. The correlation hypothesis posits that fund managers, protected by the more advantageous flow-investment relationship, can take more risk or afford to invest in more illiquid stocks. This will allow them to have trading profiles that diverge from the average ones of the styles they belong to Portfolio Risk We start focusing on the risk-taking behavior of the fund and we relate it to the CSD. For each fund and quarter, we construct two measures of risk taking. The first is based on the total risk of the fund, and the second just on systematic risk. Total risk is defined as the standard deviation of the fund returns in the prior 36 months. The systematic component of risk is the loading on the market factor. In the interest of brevity, we present only the results based on the measures of systematic risk estimated using the four-factor model in equation (3). We next regress the two measures of risk on our proxies for CSD and a set of control variables. All the explanatory variables are lagged one quarter. We estimate a panel specification with fund-quarter observations. All the regressions include time fixed effects (quarter dummies) to account for cross-sectional dependence, fund country fixed effects, and investment region fixed effects. Standard errors are clustered at the fund level to account for autocorrelation in fund risk. The set of control variables also includes the distance between the fund and the stock. This variable helps to control for the fact that foreign and distant fund managers have access to less information and this negatively impacts their performance (Brenann and Cao (1997), and Coval and Moskowitz (1999, 2001)). To address the issue of potential collinearity between fund-stock 15

17 geographic distance and CSD, the first specification does not include fund-stock geographic distance as a regressor. The results are reported in Table 4 for total risk and Table 5 for systematic risk. We report both the results based on client-stock geographic distance and those based on client-stock return or flow correlation. The results display a strong positive correlation between CSD and risk taking. This holds across the different specifications and for both types of proxies of CSD (i.e., correlation-based and geographic distance-based) as well as for the type of aggregation used (value-weighted, equally weighted, or weighed by market capitalization, GDP, and population). The effect is also economically significant. A one-standard deviation greater geographic distance between where the fund sources and where it invests is related to 20 basis points (per quarter) higher total standard deviation and a 0.07 higher systematic risk (i.e., market beta). Similarly, a one-standard deviation lower return correlation between the market in which the fund sources and that in which it invests is related to a 43 basis points (per quarter) higher total standard deviation and a 0.17 systematic risk (i.e., market beta). These results suggest that funds do indeed take advantage of their client-stock distance by taking on more risk when they enjoy a natural hedge provided by the location of their investor demand. If we look at the control variables, we notice that the distance between a fund and the assets in which it invests is always negatively related to the fund s total standard deviation. This is consistent with the idea that proximity-investing provides a fund with better information, and therefore allows it to pursue riskier stock selection strategies Portfolio Liquidity We now focus on portfolio liquidity. We relate the degree of liquidity of the portfolio holdings of the fund to our proxies of CSD. If the correlation hypothesis is true, we expect that funds with higher CSD will hold more illiquid assets. The insurance provided by the lack of correlation between performance and outflows will attenuate the liquidity needs of the funds. Indeed, there will be a lower probability that the funds will be forced to sell when asset prices are low. We regress measures of portfolio liquidity on CSD and a set of control variables as before. We consider alternative measures of liquidity based on the value-weighted average of portfolio stock holdings market capitalization, share turnover, or Amihud illiquidity measure. The results 16

18 are reported in Table 6. We report the results based on both client-stock geographic distance and client-stock return or flow correlation. The results display a strong negative relationship between portfolio illiquidity and CSD across the different specifications. The results are not only statistically significant but also economically relevant. Funds characterized by a one-standard deviation higher client-stock geographic distance display a $2.6 billion lower market capitalization, 24 percentage points lower share turnover, and 0.3 basis points higher Amihud illiquidity of its portfolio stock holdings. These results support our hypothesis that CSD provides a sort of insurance for the fund that allows it to invest in more illiquid and therefore more profitable stocks Trading Strategies We have argued that CSD provides a sort of natural hedge for a fund and allows it to take on more risk and invest in more illiquid assets. We now investigate the type of trading strategy that this risk taking involves. Funds may simply load up more on market risk. The increase in systematic risk taking provides evidence of this. Or a fund may also load up more on idiosyncratic risk. This will happen if the fund either invests in more risky stocks than the market or implements some dynamic trading strategy. To investigate this issue, we look at measures of selectivity or active management and show that they are related to the fund s CSD. A positive relation will suggest that CSD allows the fund to pursue stock-picking activities that funds less well positioned in terms of demand will not be able to follow. We focus on two measures of active management based on returns. The first is the fund s tracking error (e.g., Wermers (2003)). This measure is based on the standard deviation of the fund s risk-adjusted returns according to the four-factor model. The second measure is the fund s R-square (R 2 ) from the regression of fund s returns on the four-factor factor portfolios returns. Amihud and Goyenko (2009) propose that fund performance can be predicted by the R 2 of the regression of the fund s return on the four-factor benchmark portfolios. The lower the tracking error and the higher the R 2, the closer a fund mimics the benchmark portfolio and the lower its selectivity. The higher the tracking error and the lower the R 2, the less a fund follows a conventional investment and the higher its selectivity. Given that active management and selectivity enhances mutual fund performance (Cremers and Petajisto (2009)), a higher tracking error and a lower R 2 should proxy for the performance attributable to stock picking and 17

19 selectivity. We thus expect CSD to be positively related to tracking error and negatively related to R 2. 8 We regress either the tracking error or R 2 on CSD and the same set of control variables as before. The results are reported in Tables 7 and 8 for tracking error and R 2. We report the results based on both client-stock geographic distance and client-stock return or flow correlation. We find a strong positive correlation between CSD and tracking error and a strong negative correlation between CSD and R 2. The effect is also economically significant. A one-standard deviation greater geographic distance between where the fund sources and where it invests is related to a 0.61 percentage points (per quarter) higher tracking error and a 3.8 percentage points lower R 2. A one-standard deviation reduction in return correlation between where the fund is sold and where it invests is related to a 1.03 percentage points (per quarter) higher tracking error and a 7.0 percentage points lower R 2 by 7.0 percentage points. In untabulated results, we obtain consistent findings using the market model to construct both measures. These results support the correlation hypothesis. The natural hedge provided by the scarce correlation between flows and investment returns allows funds to pursue riskier and more profitable stock selection strategies. This gives fund managers more selectivity. More closely or more highly correlated funds are restricted to tracking their benchmarks more closely. 6. Performance Results We now look at fund performance. We have argued that CSD provides a source of strategic advantage that allows funds to take more risk and invest in more illiquid assets. Both elements higher risk and higher illiquidity implies higher raw and style adjusted return, while the latter higher illiquidity should allow for higher net-of-risk return (performance). We therefore expect CSD to improve fund performance. To test this hypothesis, we regress the fund s abnormal performance (alpha) on our proxies for CSD and the same set of control variables defined above. All the explanatory variables are lagged one quarter. We estimate a panel specification with fund-quarter observations. All the regressions include time fixed effects (quarter dummies), fund 8 There are other measures of active management. For example, Kacperczyk, Sialm, and Zheng (2005) exploit the degree of concentration of the fund holdings in a specific industry. Brands, Brown, and Gallagher (2005) create a divergence index based on the sum of squared deviations of the fund portfolio s stock weights from the market portfolio. Cremers and Petajisto (2009) create a measure of active share based on the share of portfolio holdings that differ from the fund s benchmark index holdings. All these measures are appealing, but as they are holdingbased, they significantly reduce our sample size. 18

20 country fixed effects, and investment region fixed effects. Standard errors are clustered at the fund level to account for autocorrelation in fund performance. The fund abnormal performance is defined as above. To control for the correlation with the fund-stock geographic distance, we include the distance between the fund and the stocks among the control variables. To address the issue of potential collinearity between fund-stock geographic distance and CSD, the first specification does not include fund-stock geographic distance as a regressor. The results are reported in Table 9. We report the results based on both client-stock geographic distance and client-stock return or flow correlation. The results show a strong positive correlation between CSD and performance. This holds across the different specifications and for both types of proxies of CSD (correlation-based and geographic distance-based) as well as for the different types of aggregation we have used i.e., value-weighted, equally weighted, or weighed by market capitalization, GDP, and population. The effect is also economically significant. A one-standard deviation greater geographic distance between where the fund sources and where it invests is related to a 46 basis point (per quarter) higher performance. Similarly, a one-standard deviation lower return correlation between the market in which the fund sources and that in which it invests is related to a 28 basis point (per quarter) higher performance. It is worth stressing that we obtain qualitatively similar findings using the market model. An interesting observation is that the variable that proxies for the distance between the fund and the stocks in which the fund invests is consistently negative across all specifications. This is also economically significant. A one-standard deviation greater geographic distance between the fund and the stocks in which it invests is related to a 34 basis point (per quarter) lower performance. This is consistent with the existing evidence in the literature showing that proximity provides a fund with better information, and that distant or foreign investors do suffer an information disadvantage if compared to closer or domestic investors. Overall, the results show that CSD does indeed help funds to provide higher performance. It is also positively related to risk taking. Funds do take more risk and this translates into better risk-adjusted performance. CSD also induces higher beta, however. Does this imply that the funds just load up more on market risk or that they also take on more idiosyncratic risk? This is the question we will address in the next section. 19