The determinants of cross-border equity flows

Annual bilateral gross portfolio equity transaction flows for 14 countries over 1989-1996 from Cross Border Capital (London) – 1456 observations – whose distinctive feature is coverage of country pairs that exclude the United States (§3). “They are annual bilateral (source and destination) gross portfolio equity flows, 1989-1996. The data set includes the US, but unlike any other data on asset flows we know of, it also has observations on bilateral country pairs excluding the US. Because of this special feature of the data, we can analyze the determinants of international trade in equities in a general set up, controlling for the special status of the US as the biggest world economy, as one of the main financial centers and as issuer of the main international currency.” The countries are the US and Canada; Japan, Hong Kong and Singapore; the UK, Germany, France, the Netherlands, Spain, Italy and Scandinavia; Switzerland; and Australia, which together accounted for 86.6 percent of global equity market capitalisation in 1996. The non-US share of the transactions is substantial: “there are many transactions (42% of the total, in value), which do not involve the US either as the source or the recipient country. The US is therefore far from being a ‘hub’ for all world financial transactions,” and “[i]ntra-European trade accounts alone for 24% of the transactions.”

Because gross transactions reveal frictions and segmentation that net flows and holdings do not – at the cost of requiring a separate step to connect flows to stocks (§1, §3, §6). The case for transactions: “[s]tudying transactions in those securities is interesting (and complementary to studying holdings and net flows), because transactions tell us something about the nature of the frictions economic agents encounter when trading assets and the degree of market segmentation.” The authors also note why net flows are the wrong object in their data: “[i]n these annual data, the net flows are typically very small by comparison with gross purchases and sales – perhaps of the same order of magnitude as the measurement error in the data.” The cost is stated in a footnote about the diversification argument: “our data are for transactions, not asset holdings, so this argument is valid only if there is a positive relationship between flows and stocks. We do indeed find this to hold (see Section 6).”

From Martin and Rey’s general-equilibrium model of asset trade with endogenous market capitalisation, not from analogy with the goods-trade literature (§2.3). The three ingredients “required to generate such an equation are: (1) that assets are imperfect substitutes because they insure against different risks; (2) that cross-border asset trade entails some transaction and/or information costs; (3) that the supply of assets is endogenous.” Endogenous supply operates through a specific channel: “[h]igher aggregate demand from foreign countries implies a higher asset price, which in turn increases the incentives of agents to start new risky projects and list more financial assets.” The resulting equation makes log transactions linear in the log product of the two “economic masses” – here equity market capitalisations – and in log trading costs, with coefficients predicted to be positive on mass and negative on cost. The authors flag the model’s limitation plainly: “[t]he model briefly sketched here is a simple static model. Transactions in and holdings of foreign assets coincide. We are fully aware of this limitation.”

Unit elasticities on market capitalisation, a distance elasticity of -0.881, and an R-squared near 0.70 with five regressors (§4.1, Table 2). The first specification uses beginning-of-period market capitalisations and financial-market sophistication in source and destination: 0.987 (0.037) and 1.055 (0.035) on the capitalisations, 0.456 (0.038) and 0.094 (0.037) on sophistication, R-squared 0.555. “In column (2) we add distance. Distance is appropriately negatively signed and precisely estimated, and the R2 of the regression jumps from 0.555 to 0.693: with five independent variables, this straightforward, simple ‘gravity’ regression captures almost 70% of the variance in our 1456 observations.” The comparison the authors draw is with the goods-trade literature: the fit “compares very favorably in terms of precision of estimates and explanatory power with the gravity regressions run in the goods trade literature, which have been hailed as one of the strongest and most robust stylized facts in international economics.” A between-estimator on group means gives a distance elasticity of -0.890 (0.063) and “the R2 for this cross-section regression is remarkable: we explain 84% of the variance.” Because the mass elasticities are “never statistically different from one,” the rest of the paper normalises transactions by the product of the two capitalisations.

Because neither transport costs nor diversification can produce it, and diversification predicts the opposite sign (§1). “The very significant negative impact of distance on transactions is at first sight quite surprising and puzzling: unlike goods, assets are ‘weightless’, and distance cannot proxy transportation costs! Moreover, if investors seek to diversify their portfolios, they may want to buy equities in distant countries whose business cycles have a low or negative correlation with their own country’s cycle. If that were so, distance could have a positive effect on asset trade because of the diversification motive.” The empirical premise for that counterfactual is supplied by citation rather than assumed: trade between country pairs is positively related to business-cycle correlation and trade falls with distance, and Imbs gives direct evidence that business-cycle correlations decline with distance. The authors’ proposed resolution – “[t]he most natural explanation is that informational frictions are positively correlated with distance” – is what the rest of the paper tests.

Telephone traffic, foreign bank branches, trading-hours overlap and an insider-trading index; the first two are consistently strong, overlap has some power, and insider trading is the fragile one (§3, §4.1, Tables 2, 3, 5). Telephone call traffic is the paper’s own innovation – “we believe we are the first to introduce this variable” – normalised by the square root of the product of real GDPs, and defended as exogenous: “[b]ecause this variable measures total telephone call traffic between the two countries, it is not significantly endogenous to financial market activity.” In the normalised pooled specification the coefficients are 0.174 (0.027) on telephone traffic and 0.148 (0.034) on bank branches, with distance at -0.673 (0.040); the insider-trading index is essentially zero, -0.001 (0.044). Trading-hours overlap enters positively at 0.057 (0.033). The bank-branch variable is entered at its beginning-of-period value “[i]n order to avoid potential endogeneity problems,” with the additional argument that “bank branches are not set up primarily to deal with portfolio equity trade, but for a wide range of reasons.” On why several information proxies rather than one, the authors offer an explicit conjecture: telephone calls “might represent the information gathering of the broad population and the cross-country networks associated with migration, cultural ties, past colonial relationships,” while “[f]oreign bank branches might transmit information about specific companies directly to investment managers in the home country” – “[t]he argument is conjectural, but the heterogeneity of information sets might leave room for several distinct ‘information variables’.”

The authors address the collinearity concern directly and report the correlation (§4.1). “One might be concerned about multicollinearity between distance and telephone calls – indeed, a causal relation between them – but the (robust) standard errors on their coefficient estimates are low, these estimates are very stable across specifications, and the correlation between the two variables is also not disturbingly high (-0.32).” They also report that telephone traffic “[w]hen added on its own without distance … also performs very well,” and that “[w]hen both are included, the coefficient of each is significantly less than what we obtain in estimates with either alone” – which is what one would expect if both partly capture a common information channel, and is the basis for the paper’s reading of distance as an information proxy rather than as something else.

45 percent of the pooled variance and 65 percent of the cross-sectional variance with six regressors (§4.1, Table 2 columns 5-6). “With a total, then, of six explanatory variables, we capture 45% of the variance of bilateral cross-border equity flows (and 65% of the cross-sectional variance) for fourteen countries over 8 years.” The drop from the 0.693 of the unnormalised specification is mechanical – normalising by the product of capitalisations removes the two mass variables’ contribution from the left-hand side – and the authors are explicit that the informative variation is cross-sectional: “the interesting variation in our panel is virtually all cross-sectional; a ‘between’ estimator on the time-series means for the country pairs demonstrates this clearly.”

Because the variable of interest is time-invariant, so fixed effects would absorb it – an argument the paper makes explicitly, alongside a country-dummy robustness check (§4.1, Table 2 column 7). “We do not introduce country-pair fixed effects … because we have a strong prior that the distance variable should be a major determinant of the flows. By construction, the distance variable (which is constant over all observations for a given country pair) will pick up some of the fixed effects. Conversely, with fixed-effects panel data estimation, we cannot use any time-invariant variable, because any such variable is spanned by the individual dummies representing the fixed effects.” Random effects are ruled out on sampling grounds – the data “are not drawn randomly from a larger population” – though the authors report GLS estimates are “fairly similar.” What they do run is a full set of source- and destination-country dummies, which leaves distance at -0.646 (0.056) and bank branches at 0.236 (0.057), with telephone traffic falling to 0.078 (0.032) and insider trading turning significantly negative at -0.209 (0.105). The unavailability of formal poolability tests is also stated rather than glossed: a Wald test “fails because of the Behrens-Fisher problem,” a Chow test “fails because variances of the sub-samples are not equal over years,” and the generalized Chow test requires the within estimator that cannot handle time-invariant variables.

The distance elasticity is stable, and the insider-trading variable comes to life exactly where its variation is (§4.2, Table 3). Dropping the US gives -0.721 (0.047) on distance with 1248 observations; dropping the US and the UK gives -0.856 (0.056) with 1056. Restricting to intra-European flows (448 observations), “the basic specification works for all our information variables. Insider trading is correctly signed and significant” at -0.398 (0.117), and “[t]he elasticity on distance is very close to the one we found for the whole sample” at -0.727 (0.139). The authors explain the change: “inspection of the data shows there is much more variation across Europe in the perceived extent of insider trading (with Spain, Italy and France at the ‘bad’ end of the spectrum) than there is among the non-European countries in our sample.” Excluding intra-European flows instead leaves 1008 observations with distance at -0.632 (0.087), telephone traffic at 0.182 (0.033) and bank branches at 0.192 (0.039). The intra-European finding is the one the authors emphasise interpretively: “[e]ven in an arguably very integrated economic area (but before currency unification), the evidence points toward significant informational segmentation.”

Adjacency, common language, regional and currency blocs, legal-system quality, direct transaction-cost measures, and private credit – reported as a list of things that did not overturn the result (§4.2). Adjacency “is strongly collinear with the regional bloc dummies and brings no improvement”; the common-language dummy “is significant with the expected sign for some specifications” but “the coefficients on the initial explanatory variables were very stable in all specifications.” On exchange rates, using Frankel and Wei’s continuous volatility measure and a constructed stability dummy, “this variable took on a (insignificant) negative coefficient. The continuous volatility measures did not perform well either. Again, exchange rate stability does not seem to have a positive influence on cross-border equity transactions” – with an immediate qualification: “(this does not imply that currency union would have no such effect).” Two legal-effectiveness indices were “[n]either … consistently significant,” which the authors attribute to limited variation: “[m]ost of the countries in our sample rank so highly on this criterion that there is relatively little variation.” Direct Elkins/McSherry transaction-cost estimates “perform almost as well” as the sophistication index with correct signs, but slightly worse fit, so the survey-based index was retained – a choice the authors state rather than bury.

That the covariance term takes its predicted negative sign only after information is controlled for, and that the evidence for diversification is weak and less robust than the rest (§4.3, Table 5). For this exercise the dependent variable becomes normalised net purchases, since “we are now investigating the motive for acquiring foreign equities.” The prediction is a negative coefficient on the covariance of the two countries’ stock-market returns: “the greater the comovements between financial assets of two countries, the lower the benefit of diversification.” But the friction may dominate, precisely because “the correlations of different countries’ assets tend to be negatively correlated with distance,” which the authors illustrate concretely: “[i]f the diversification motive were powerful, French people, say, should invest a lot in Australian equities (controlling for size and transaction costs), since the French and Australian stock markets are not highly correlated. But if French people know very little about Australia, they may not want to invest there much anyway.” The data behave accordingly: “the covariance variable enters with a positive sign in our baseline regression when we do not control for the information friction” – 0.346 with a standard error of 0.136 – “[w]e are just picking up here the fact that people prefer to invest in markets ‘close’ to them.” With distance and the explicit information variables included, the covariance term turns negative as predicted, and interacting comovement with distance also gives the expected sign. The authors’ own verdict is the calibration to carry: “[t]hese results however are somewhat unstable across specifications. On balance, we conclude that there is weak evidence for a diversification motive for asset trade in our annual data, but only when we control for the informational friction. We view these results as less robust than our results on the informational friction itself.”

The distance elasticity roughly halves, which the authors read as a finding about the goods-trade literature (§5, Table 6). On a matched panel of manufactures trade among the same countries, the standard specification reproduces the stylised fact: a distance elasticity of -0.547 (0.048), which the authors place against the literature’s benchmark – “Leamer and Levinsohn (1995) cite a ‘consensus elasticity’ of -0.6; our point estimate of -0.55 in column (1) is one standard deviation away from this.” Adding normalised telephone traffic (0.123, standard error 0.010) and bank branches (0.141, 0.019) changes the picture: “[t]he information variables do indeed enter with sizeable, very well-determined coefficients; and they improve the regression considerably. The EU dummy becomes significant, the proportion of the variance explained rises substantially, and most importantly, the coefficient on distance falls sharply. The elasticity is now only -0.28!” The inference is stated as a suggestion rather than a demonstration: “[i]t seems that information flows may be at least as important” as transport costs, and in a footnote, that the empirical goods-trade literature “overestimates the importance of transportation costs (proxied by distance) and considerably underestimates the importance of information asymmetries (also proxied by distance).”

No – bilateral trade enters the equity equation significantly, but distance and the information variables survive it (§4.2, §5, Table 6). Obstfeld and Rogoff’s model in which asset trade mirrors goods trade “can therefore potentially explain why the distribution of asset flows obeys a ‘gravity’ model like the distribution of trade flows, even without any transaction costs or information costs on asset markets.” Testing it, “trade flows do enter significantly in the equation but that distance and the other information variables remain strongly significant”: in the unnormalised equity regression, adding trade at 0.364 (0.048) moves distance only from -0.666 (0.040) to -0.455 (0.046), and in the normalised specification distance stands at -0.529 (0.042) with trade at 0.224 (0.031). Within Europe, with intra-European trade controlled for, distance remains at -0.451 (0.163), which the authors read as meaning “the type of information needed to trade equities within Europe cannot be summarized by trade linkages.” The same test disposes of a specific omitted-variable worry: since “the benefits of diversification may be correlated with the intensity of trade between countries (and thus with geographical distance),” that “might have generated an omitted variables bias in our regressions. Our results here, however, also dismiss this possibility.” The conclusion drawn about Obstfeld-Rogoff is correspondingly partial: their model “may capture part but not all the determinants of asset flows.”

A descriptive illustration of national information sets, explicitly not offered as a general result (§4.2, Table 4). The authors searched FT Profile for country keywords in headlines of four general-interest European newspapers (Le Monde, The Guardian, the Frankfurter Allgemeine Zeitung, La Stampa) and three financial ones (the Financial Times, Les Echos, Il Sole 24 Ore), tabulating the share of headlines devoted to each country. “The results are suggestive: there is a much broader coverage of Spain and Italy by French newspapers compared to that of the British and to a lesser extent the German press. On the other hand, Switzerland is followed much more closely by Germany than by the UK (or France).” The relevant summary statistic is that “the correlation between the number of articles written in country i about country j and the distance between the countries is indeed negative: -0.23 for the general interest newspapers and -0.33 for the financial newspapers.” The authors attach an explicit disclaimer in a footnote: “[t]hese results are illustrative rather than claiming to be general.”

A near-unit elasticity of transactions with respect to holdings and a similar distance elasticity for holdings – on 80 observations, which the authors treat as illustrative (§6). Using the two US Treasury benchmark surveys of US holdings of foreign long-term securities, covering some 40 countries for two years, a between-estimator gives log(US transactions) = 1.05 (0.053) x log(US holdings) + 6.66 (0.127), with R-squared 0.87, so “[t]he elasticity of US residents’ transactions in foreign corporate equities with respect to US holdings in those equities is close to one.” A holdings regression then reproduces the transactions result: log(US holdings) = 0.47 (0.082) x log(market capitalisation) + 0.24 (0.098) x sophistication - 0.71 (0.262) x log(distance), R-squared 0.63, “produc[ing] for distance an elasticity which is very similar to the ones we found in Section 4 for the transaction data.” A further detail supports proportionality: “regressing our turnover ratio variable on distance does not give anything: this tells us that our information variables impact holdings and transactions in a proportionate way.” The hedge is unambiguous: “[w]e are unable, however, to check the robustness of these results as thoroughly as we could for our previous results on transactions data, because of the small number of observations in the holdings data and the special status of the US. We therefore consider these results as illustrative.”

That the geographical pattern of asset trade is real and is shaped mainly by information, and that the theory of asset trade should be rebuilt accordingly – without a claim to have identified the mechanism or measured its welfare cost (§1, §7). The headline conclusion: “[w]e view our empirical work as strong evidence that there is a very important geographical component in international asset flows. International capital markets are not frictionless: they are segmented by informational asymmetries or familiarity effects.” The authors are careful that the two candidates are not separated: “[s]eparating out ‘familiarity’ effects from pure informational symmetries remains a challenge for the empirical literature,” though they note Coval and Moskowitz’s finding that locally investing funds earn substantially higher returns “suggesting that for this class of investors, local investments reflect a true informational advantage.” The programmatic claim is that modelling should “shift away from models based on factor endowments, comparative advantage and autarky prices … towards models including differentiated assets, transaction costs, information asymmetries and possibly models based on some type of ‘familiarity effect’.” And the open question is left open: “[w]hether theoretical dynamic models based on asymmetric information and heterogeneous beliefs are more appropriate or whether the theory should also emphasize issues like ‘familiarity’ and behavioral explanations … remains an open issue.”