A Gravity Model for Bilateral Home Bias:

Beyond Distance
∗
Crina Pungulescu

Abstract

This paper uses recent technology from the fields of cognitive neuroscience and psycho-linguistics
to introduce a new variable into the established framework of gravity models in finance. This is
a measure of country similarity, computed based on the overlap of the semantic fingerprints (i.e.
numerical representations of meanings) of any pair of country names within an artificial human-like
brain called the Retina. The resulting variable, measuring the conceptual commonality of country
pairs is added to the set of typical gravity variables (distance, common language, border) in a
model for bilateral home bias. Using both the traditional benchmark for computing home bias (i.e.
the world market capitalization shares) and an alternative methodology that takes into account
the evolution of asset returns (i.e. the performance of financial markets) in determining optimal
investment, this paper firmly establishes the significance of the new similarity variable. Countries
appear to be less underinvested in similar (even if not geographically close) partners and affinity
even leads to overinvestment. If similarity is interpreted as a proxy for trust, it suggests that the
need for trust in the investment partner counteracts the financial rationale of risk diversification.

JEL classification: F36, G11, G15

Keywords: Home Bias, Gravity Models, International Portfolio Choice, Model Uncertainty,
Natural Language Programming

∗ John Cabot University, Via della Lungara, 233, 00165, Rome, Italy email: cpungulescu@johncabot.edu, phone: +39
06 681 9121, fax: +39 06 589 2088.

1
1 Introduction
The fact that diversification improves portfolio performance is a tenet of finance that appears to have
been largely ignored in international settings. Since the seminal contribution of French and Poterba
(1991), the difference between optimal and observed (far too low) foreign holdings has been known as the
home bias puzzle. A rich literature delved into explaining the puzzle initially by blaming direct barriers
and transaction costs (Tesar and Werner, 1995). As home bias persists even when such deterrents
are minimal (see Glassman and Riddick, 2001; Warnock, 2001), attention refocused towards access to
information (see Gehrig, 1993; Brennan and Cao, 1997; Coval and Moskowitz, 1999; Ahearne et al.,
2004; Veldkamp and Van Nieuwerburgh, 2006; Bae et al., 2007) or investors’ perceptions, and related
psychological or behavioural factors (see Jenske, 2001; Huberman, 2001; Grinblatt and Keloharju, 2000;
Li, 2004; Lütje and Menkhoff, 2004; Clark, 2015).
For the largest part of the research into the home bias puzzle, data constraints have forced an
aggregate view on home bias (i.e. domestic vs. rest-of-the-world holdings). A relatively new initiative
by the International Monetary Fund (the Coordinated Portfolio Survey) collects data on bilateral equity
holdings for a large number of countries since 2001 and allows for novel approaches to investigating home
bias, in bilateral terms and on a truly worldwide scale.
In this context, gravity models that have been the realm of the international trade literature provide
now a particularly promising tool for studying home bias. Since the important contribution of Portes
and Rey (2005), gravity models have enjoyed substantial attention in international finance and appear
to fit financial data well (Okawa and van Wincoop, 2012). They answer important questions such as
the impact of various phenomena and policy (i.e. globalization, trade agreements etc.) on cross-border
trade in financial assets. Notably, while in the gravity models of international trade, distance relates to
transportation costs, financial assets are ‘weightless’ - distance cannot proxy transportation cost. Portes
and Rey (2005) remark, however, that “a gravity model explains international transactions in financial
assets at least as well as goods trade transactions” and interpret distance as indicative of “barriers
to interaction and, more broadly, to cultural exchange”. Bekaert and Wang (2009) find that gravity
variables (distance in particular) are robust determinants of bilateral home bias, and their interpretation
does not distinguish whether they proxy for information or a so-called “familiarity bias”(Chan et al.,
2005).
This paper makes several contributions relevant to this line of investigation into home bias. Firstly,
it introduces an entirely new and promising variable to the set typically used by gravity models. State-
of-the-art research in cognitive neuroscience and psycho-linguistics provides a novel technology that can
be applied to produce a new measure of country similarity. The recently developed Cortical.io Retina
engine is able to emulate the way the human brain works in synthesizing information and can convert any

2
text into a numerical representation of its meaning: its semantic fingerprint. The semantic fingerprint
takes shape on a 2D map of meanings (a 128×128 matrix where each location stores a pre-defined
commonly occurring semantic contexts) by setting a value of 1 to a location if the meaning pre-stored
there applies to the word or text analyzed and 0 otherwise. An immediate measure of similarity, the
proportion of overlap between the semantic fingerprints of any two countries in the context of this paper
can be applied alongside distance in gravity models. A fine-tuned similarity measure encompassing a
large array of possible contexts and perceptions (as opposed to mere geographical distance) is arguably
richer in meaning. Empirically, the measure appears statistically significant and informative over and
above the contribution of distance and other gravity variables.
Secondly, using an alternative Bayesian approach (see Pástor, 2000; Pástor and Stambaugh, 2000) for
defining optimal investment by taking into account the evolution of asset returns (i.e. financial market
performance of source and destination countries) shows that the phenomenon of underinvestment in
foreign assets is matched by that of overinvesting in selected partners. Similarity (not geographical
proximity) as well as membership in the EMU appear to attract a higher allocation of resources than
optimal. A way of distinguishing between the information channel and the familiarity bias suggests
itself. Well-used information should result in better investment, whereas similarity (interpreted as
familiarity and possibly trust) might entice investors into suboptimal choices.
While these hypotheses are warrant further research, the results of this paper firmly establish simi-
larity as a relevant variable in gravity models in the setting of bilateral home bias.
Section 2 details the procedure of computing optimal portfolio weights, Section 3 discusses the
particularities of computing bilateral home bias and data issues, and most importantly the technology
behind the similarity measure based on the overlap of semantic fingerprints of countries. The results
are presented in Section 4 and Section 5 concludes.

2 Optimal Portfolio Weights

This section presents alternative ways to calculate theoretically optimal portfolio weights with which
observed weights can be compared. Section 2.1 reviews the standard mean-variance model of portfo-
lio choice. Section 2.2 discusses the International CAPM. Sections 2.3 and 2.4 present the Bayesian
modeling approaches of Pástor (2000) and Garlappi et al. (2007), respectively.

3
2.1 Classical Mean-Variance Portfolio Model

The common starting point is the mean-variance framework of Markowitz (1952) and Sharpe (1963)
where the investor makes his portfolio choice in order to maximize his expected utility,

γ 0
max ω 0 µ − ω Σω, (1)
ω 2

where ω is the N -vector of portfolio weights allocated to N assets, i.e. domestic and foreign equity
holdings (N = 2), µ is the N -vector of expected returns, Σ is the N × N variance-covariance matrix and
γ is the coefficient of relative risk aversion. Under the assumption that ω 0 ι = 1 (the budget constraint),
the solution of the portfolio problem becomes

1 −1
ω∗ = Σ (µ − ηι), (2)
γ

where η denotes the expected return on the zero-beta portfolio corresponding to the optimal portfolio
and ι is a N -vector of ones. The budget constraint effectively fixes γ for a known value of the zero-beta
expected return through γ = ι0 Σ−1 (µ − ηι) and determines uniquely the optimal portfolio weights (De
Roon and Nijman, 2001). If a risk-free rate is available and chosen as the zero-beta portfolio, the
coefficient of risk aversion becomes γ = ι0 Σ−1 µe , where µe is the vector of the expected excess returns
(over the risk-free rate). The analytical portfolio choice solution in the mean-variance framework, when
short sales are allowed is:
Σ−1 µe
ω∗ = . (3)
ι0 Σ−1 µe
The solution of the optimization problem involves the true (unobserved) expected returns and
variance-covariance matrix of the returns. Available returns data enables us to use the sample mo-
ments as estimates of the true parameters. However, Merton (1980) shows that the sample variance-
covariance matrix gives an accurate estimate of the true parameter but the estimation of the expected
returns based on historical data is very unreliable due to the high volatility of returns. The impact
of the mean estimated imprecisely, is amplified in the context of international portfolio choice, as the
inverse of the variance-covariance matrix tends to be a large number when the correlations between the
countries are high (Jenske, 2001). Therefore, the “data-based” approach (i.e. substituting the sample
mean and variance in equation 3) directs investors to take extreme and volatile positions.

2.2 International CAPM

An asset pricing model, such as the International Capital Asset Pricing Model (I-CAPM), provides an
alternative to the “data-based” approach. The I-CAPM is valid in a perfectly integrated world, where
the law of one price holds universally and markets clear (total wealth is equal to total value of securities).
The world market portfolio can then be defined as the sum of all individual portfolios weighted by the

4
positions held by mean-variance investors. The portfolio implication of the CAPM is that the average
mean-variance investor holds the market portfolio (Lintner, 1965). In an international setting, the
optimal investment weights of a country according to this so-called “model-based” approach, are given
by the relative shares of domestic and foreign equities in the world market capitalization. For a US
investor, this implies that domestic equity holdings should have been about 40% in 2004. The actual
domestic allocations figures for the US were as high as 80%. The resulting “model-based” home bias
has been traditionally used in the literature and gives the first measure of home bias in the present
study.
The I-CAPM results in the well-known linear beta relationship between risk premium on the domestic
portfolio and the expected excess return on the world market benchmark1 :

E(rd ) − rf = β [E(rw ) − rf ] , (4)

cov(rw ,rd )
where rd is the return on the domestic market portfolio, rf is the risk free rate, β ≡ var(rw ) is the
world beta of the domestic market and rw is the return on the world market portfolio. The empirical
counterpart of equation 4 is given by

rd − rf = α + β (rw − rf ) + ε, (5)

where α and ε are respectively the intercept and the disturbance term. The I-CAPM is considered valid
if estimates of the intercept, α̂ are zero. However, an intercept different than zero, even if insignificant,
can be used by a Bayesian investor to question the optimality of the portfolio prediction of the I-CAPM,
and therefore the reliability of the traditional “model-based” measure of home bias.

2.3 Bayesian Mean-Variance Portfolio Weights

Considering the stringency of the assumptions of the I-CAPM, it is reasonable to expect that some
investors do not accept the model unconditionally.
When the I-CAPM holds, the world benchmark fully describes the asset returns and captures all
sources of priced risk. In terms of the beta pricing relationship (5), a valid model results in a zero value
for the intercept α̂. In the Bayesian framework developed by Pástor (2000), when there is mistrust in
the I-CAPM, the data becomes informative and is involved in the portfolio allocation decision. The
degree of trust (i.e. the belief that the intercept α̂ is zero) is expressed in values of the standard errors
of the intercept σα . A small value indicates a strong belief that the theoretical model is valid and results
in optimal portfolio weights that closely correspond to the “model-based” approach. A higher value
1 This model makes the additional assumption that currency risk is not priced. See De Santis and Gérard (2006) and
Fidora et al. (2007) for an analysis of exchange rate risk on home bias measures.

5
involves data to a larger extent in the computation of optimal weights leading thus to a different set of
optimal weights and brings us closer to the results of the “data-based” approach. Full mistrust in the
model (i.e. σα → ∞) coincides with the “data-based” optimal weights. This Bayesian interpretation is
an insightful reconciliation of the “model” and “data-based” approaches. For instance, a nonzero value
for α̂, even if insignificant according to a standard t-test (and therefore failing to reject the I-CAPM),
could become instrumental in explaining why observed allocations deviate from the model prescriptions.
The starting point of the Bayesian analysis is a prior (non-data) belief in the the model, in this case,
the belief in a zero intercept and no mispricing. The prior is updated using returns data to a certain
extent depending on the chosen degree of mistrust in the model. The sample mispricing, α is “shrunk”
accordingly towards the prior mean of α to obtain the posterior mean of α.
Using the data in combination with the model prediction ultimately results in different estimates
for the mean and variance covariance matrix of returns, since now the moments of the predictive
distribution are used to compute the portfolio weights. These Bayesian mean-variance optimal weights
are computed as:

Σ∗−1 µ∗e
ω∗ = (6)
ι0 Σ∗−1 µ∗e
where µ∗e and Σ∗ are the predictive mean and variance that replace in this approach the sample moments
of the distribution of returns.
The predictive density of returns (entering the utility function of the investor that maximizes next
period wealth) is defined as:
Z
p (rt+1 |Φ) = p (rt+1 |θ, Φ) p (θ|Φ) dθ (7)
θ

where p (rt+1 |Φ) is the probability density function of excess returns conditional on Φ (the sample data)
and θ is the set of parameters of the statistical model that describes the stochastic behavior of asset
returns. To treat the estimates of the parameters θ̂ as the true values, is to ignore estimation risk. An
alternative is to use Bayesian analysis to account for estimation risk. The predictive density (equation
7) involves p (θ|Φ), the conditional probability of the parameters of the model given the data available.
According to Bayes’ Rule, the posterior density, p (θ|Φ), is proportional to the product of the likelihood
function, or probability distribution function for the data given the parameters of the model, p (Φ| θ),
and the prior density, p (θ), that reflects the non-data information available about θ (see Koop, 2003):

p (θ|Φ) ∝ p (Φ|θ) p (θ) . (8)

In the present setting, the prior of a zero intercept follows from assuming a valid I-CAPM and is
subsequently updated through incorporation of the information revealed by the data. The methodology

6
and the analytical solutions for the mean and variance of the predictive density are presented in further
detail in the Appendix.
A degree of mistrust in the I-CAPM depending on the empirical performance of the model on specific
country data, may result in optimal weights that are closer to the observed allocations and thereby imply
for certain countries, a lower home bias than the deviation from the market capitalization share.

2.4 Bayesian Multi-Prior Framework

The alternative, “data-based” approach, using historical return data for estimating the optimal alloca-
tions, results in volatile and extreme investment predictions (see Baele et al., 2007, for an example).
Garlappi et al. (2007) tackle the problem of volatile data by extending the mean-variance framework
to incorporate the investors’ aversion to uncertainty around the estimate of the mean returns. This
changes the standard mean-variance problem in two ways: (1) it binds the expected returns to a con-
fidence interval around their estimate, thus taking into account the eventual estimation error and (2)
it allows the investor to minimize over the choice of expected returns, thus manifesting its aversion to
uncertainty. The Multi-Prior framework of Garlappi et al. (2007) is defined by the following problem:

γ 0
max min ω 0 µ − ω Σω, (9)
ω µ 2

subject to
f (µ, µ̂, Σ) ≤  (10)

ω0 ι = 1 (11)

where µ̂ is the sample mean of asset returns. If the confidence intervals are defined jointly for all assets,
T (T −N ) 0
f can be taken as (T −1)N (µ̂ − µ) Σ−1 (µ̂ − µ) and  as a quantile for the F -distribution2 , where N is
the number of assets and T , the number of observations. The constraint translates into P (f ≤ ) = 1−p
for a corresponding probability level. This framework can also be extended to include uncertainty over a
chosen return-generating model, such as the I-CAPM. The solution to the Multi-Prior max-min problem
is a set of optimal weights with considerably smoother behavior compared to the ones obtained through
the direct influence of the data.
This methodology is applied to obtain the second measure of home bias, the volatility corrected
Bayeasian home bias.
2 If asset returns are normally distributed and Σ is known, f has a χ2 distribution with N d.f. If Σ is not known, it
follows a F -distribution with N ,T − N d.f. (Garlappi et al., 2007)

7
3 Home Bias Measures and Data Issues
The previous section presented alternative ways of defining optimal portfolio allocations. This section
defines the measure of home bias in terms of actual and optimal portfolio weights, and introduces the
main characteristics of the data, particularity the novel measure of similarity made possible by recent
advances in the the fields of cognitive neuroscience and psycho-linguistics.

3.1 Home Bias Measures

Bilateral home bias of source country i with respect to destination country j is quantified as the relative
difference between actual (ACTij ) and optimal (OP Tij ) foreign portfolio weights:

ACTij
BHBij = 1 − . (12)
OP Tij

Optimal portfolio weights are calculated using the alternative methodologies described in Section
2. The actual portfolio holdings of source country i in the equity of destination country j (ACTij )
are determined using data Coordinated Portfolio Survey conducted by the IMF. The domestic equity
holdings are calculated as the difference between the market capitalization of the country (M Ci ) and
the total domestic equity stocks held by foreign investors (F Li ):

F Aij
ACTij = . (13)
F Ai + M Ci − F Li

In the typical case, when actual foreign involvement is lower than the optimal share of international
assets that a country should hold, and the country is subject to home bias, the measure takes values
between 1 (when the investors do not hold any foreign assets) and 0 (when actual and optimal portfolio
weights are equal). However, bilateral data offers cases when the actual weights exceed optimal weights,
for instance when negative or very low weights are assigned to the world market index in the optimization
framework. This can be the case when the world market index has a high variance and covariance with
the domestic index and a lower mean. In such instances the source country appears not home biased,
but on the contrary, overinvesting abroad and the former measure of home bias would be misleading.
Therefore, the formula is extended to take into account the case of overinvestment abroad (negative
‘home bias’) and obtain comparable results (as in Bekaert and Wang, 2009)3 , as follows:

(ACTij − OP Tj )
BHBij = − . (14)
(1 − OP Tj )

This formula is used to compute a negative measure of ‘home bias’ (a so-called ‘foreign bias’) when
optimal allocations are lower than the observed foreign investment4 .
3 In their paper linking home bias to exchange rate volatility, Fidora et al. (2007) make a different choice when computing
bilateral home bias, along the same lines.
4 If equation 12 were applied in the cases of overinvestment, the (much lower) optimal weights in the denominator

8
3.2 Data

3.2.1 Financial Data

The current exercise on bilateral home bias is made possible by the initiative of the International
Monetary Fund (IMF) to conduct the Coordinated Portfolio Investment Survey (CPIS) which covers the
geographical distribution of 80 source economies (reporting) and 241 destination economies (reported).
The first trial was successfully conducted in 1997 and starting with 2001, the CPIS takes place every
year. This paper uses the 12 years of data on cross-border stocks and bilateral foreign trade (imports)
available since 2001.
While the value-added in information brought about by such detailed statistics is substantial, it has
to be noticed that several possible bias are in-built in this database. If for instance, country A uses a
broker company in country B in order to ultimately invest in a third country C, the database will not
allow us to uncover country A’s holdings of country C’s equity. Also, participating economies may (and
do) on occasion choose to withhold information on the amounts of their holdings.
Data on weekly market index prices and their respective market capitalization is obtained from
Datastream. Weekly US$-denominated total returns have been computed for 58 of the 241 economies
covered by the CPIS. Where available (34 cases), Datastream’s total market indices have been used.
The coverage starts in January 1973 for the more established markets and as late as 2006 for Uruguay
and Luxembourg. The world market index has been computed as a weighted average of the available
market indices at each point in time. In order to compute Bayesian optimal weights every period, only
the countries with 520 weekly observations of returns data available prior to the year-end observation,
have been taken into consideration.
Annual market capitalization figures for the full sample are obtained from the World Development
Indicators database.
The risk-free rate is the one-month Treasury Bill rate from Ibbotson and Associates Inc., available
on Kenneth French’s website5 .

3.2.2 Gravity Variables and Beyond: Similarity

The typical gravity variables (distance between capital cities, common border and common language)
are obtained from the GeoDist database explained in Mayer and Zignago (2011). Gravity models have
been shown to be as successful in explaining international finance as international trade and (see Okawa
and van Wincoop, 2012, and references therein, for instance).
might result in extreme values for home bias that are practically irrelevant for subsequent analysis. Equation 14 redefines
the formula for computing home bias in such cases, maintaining therefore the scale of the resulting home bias.
5 http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html

9
 This paper introduces a completely new variable to the widely used framework of gravity models,
a variable that is both conceptually intuitive and apt to refine and increase their explanatory power.
State-of-the-art research in cognitive neuroscience and psycho-linguistics delivers a new and accurate
tool for performing semantic comparisons (see De Sousa Weber, 2015, for a detailed description). This
tool is the Retina engine6 which mimicks the functioning of the human brain (i.e. assuming language is
stored in the human brain in the form of a distributed memory as opposed to a single symbolic unit) and
can be used to analyze text in a similar way and with a high degree of accuracy. Like the human brain
accumulating and synthethising information, the Retina engine’s ongoing training on a large number
pre-selected Wikipedia articles (4.4 million for Version 2.2) results in a 128×128 matrix of “contexts”
(a 2D map of commonly occurring meanings where topology is important, in the sense related concepts
appear clustered together).
On this map of semantic contexts, each text or word finds its own semantic fingerprint, which is a
numerical (binary) representation of meaning: an element of 1 if the context in a given position in the
128×128 matrix is applicable, 0 otherwise. As countries are the object of study here, some examples
from the semantic fingerprint of “United States” 7 include clusters of meanings around contexts such as:
“federalism, legislate, constitution”, “lawyers” or “nuclear, fission, weapons”, “bilateral, agreements,
NATO”, “diplomacy, united nations, cold war” and also more isolated elements (lone 1s on the map),
such as: “colony, land, indian territory” or “softball, baseball, New York Yankees”. From the semantic
fingerprint of “United Kingdom”, clusters are found around “European Union, European Economic
Area”, “OECD, IMF, GDP”, “parliamentary, constituency, commons, parliament” or “empire, pro-
tectorate, commonwealth” along with more maverick contexts such as “peerage, lord mayor, national
trust”.
The opportunity of comparing two semantic fingerprints is immediate, in its simplest form given by
the proportion of bits that overlap in two fingerprints8 , which is 30% in the case of the United States
- United Kingdom pair. Examples of shared context include: “British empire”, “malls”, “WTO, high
income, OECD”.
This paper proposes the similarity measure defined as the semantic overlap of the two fingerprints
for the pairs of countries in the dataset obtained using this technology as a relevant regressor in gravity
models.
The similarity measure based on semantic overlap does more than exemplify the generally termed
“barriers to interaction” proxied by distance and the common language, border and colonial link dummy
variables that complement distance in gravity models. It is able to capture and synthesize in one
6 invented and developed by Cortical.io the http://www.cortical.io/product.html
7 http://www.cortical.io/expression-builder.html
8 http://www.cortical.io/demos/similarity-explorer/

10
number deeper and more fundamental similarities and possibly, perceptions of similarity (models of
government and institutions, society values, common interests, historical patterns etc.) with a level
of fine tuning that has been missing so far from gravity models. Moreover, when gravity models are
applied to ‘weightless’ financial assets and distance loses its original use as a proxy for transportation
costs, semantic overlap is conceivably more informative as than distance between the capitals of two
countries (i.e. the typical distance measure in gravity models). Suffice the example of the pairs Austria
- Slovakia (which, using semantic fingerprinting are 31% similar, while the distance between Vienna
and Bratislava is merely 59km) and Austria - Germany (41% similar, with 523 km between Vienna and
Berlin).
Figures 1 and 2 give examples relevant to the range of the similarity measure in the dataset: Spain
and Portugal are 52% similar, the semantic fingerprints of Poland and Bermuda overlap only 5%9 . The
distribution is slightly skewed to the right (see Figure 3), similarity between many pairs of countries
being just below the average of 21% (and median of 20%).
Similarity correlates significantly with the typical gravity variables (see Table 2), most particularly
with the logarithm of distance (the correlation coefficient is 0.5810 ), followed by the common border
dummy (0.38), imports (0.32) and common language (0.14). Also, among the international organiza-
tions, EU, OECD and EMU are the ones that bring together most similar countries.
Similarity based on semantic overlap has been introduced to the world of finance by Kogan et al.
(2015) who show that it is a viable predictor of stock return correlations (outperforming traditional
industry classifications) for the 30 companies covered by the Dow Jones Industrial Average (DJIA).
This interpretation of similarity as an accurate predictor of stock return correlations is central to the
scope of this paper. In their comprehensive examination of possible determinants of bilateral home bias,
Bekaert and Wang (2009) use two measures related to the diversification potential of a partner country:
the correlations between the market returns of two countries (albeit its well-documented time variability
is a disadvantage) and the difference between the industrial structures of the two countries. They note a
persistent return correlation puzzle (countries with higher correlation induce significantly lower bilateral
home bias), which is not sufficiently mitigated by the use of the industrial structure (whose explanatory
power is largely negligible). The gravity variables, of which distance is most powerful, are interpreted
in the home bias context as proxies of information/familiarity. The contribution of the (orthogonalized)
semantic overlap measure of similarity to explaining home bias over and above the documented impact of
11
distance suggests it carries a deeper interpretation. Taking a hint from anthropology , a link between
9 For comparison, “Romeo” and “Juliet” share 70% of their semantic fingerprints.
10 For this reason, subsequent estimations use the similarity variable orthogonalized on distance.
11 In a plea for interdisciplinary study and the value of anthropology in understanding fundamentals of the global

financial crisis, Tett (2013) points out that “there was a reason why the credit markets come from the Latin word credit,
or rather from credere, meaning to believe”.

11
similarity and trust is conceivable. Bhuiyan (2010) for instance documents a positive relationship
between trust and interest similarity for users of online networks. Moreover, Guiso et al. (2009) show
for a sample of European countries that the economic exchanges (international trade, portfolio and
foreign direct investment) depend significantly on trust. In the context of the return correlation puzzle
noted by Bekaert and Wang (2009), this interpretation of similarity suggests that a need for trust might
be at work counteracting the diversification motif that is the ground finance is built on.

4 Results

4.1 Descriptive Statistics

Table 1 presents descriptive statistics for the full sample for the two measures of bilateral home bias
obtained under the the I-CAPM framework (HB I) and under the Bayesian framework (HB B). As data
availability constrains the Bayesian home bias to a smaller sample, the measure of I-CAPM home bias
is also restrained for comparative purposes to the same data points, giving an alternative measure of
I-CAPM home bias (HB IC).
Several patterns capture attention. Firstly, home bias across the full sample is high for the I-CAPM
measures (0.89) and significantly lower for the measure that takes into account country performance,
the Bayesian home bias, which averages at 0.50. Secondly, the phenomenon of overinvestment is notable
and consistently identified in bilateral equity holdings. 7% of the bilateral home bias observations for
the full sample of I-CAPM home bias (significantly higher - over 20% - when the sample is constrained
to members of EU, EMU and ASEAN) and almost half (47%) for the full sample Bayesian home bias
are negative (showing a persistent ‘partner bias’). The larger number of negative instances of home
bias apparent when the performance of the asset returns in the target markets is taken into account
(using the Bayesian approach to estimate optimal portfolio weights), a large number of countries seem
to “miss the mark” and overinvest in countries that are doing poorly.
Table 2 shows that the correlation between the I-CAPM and the Bayesian home bias is positive and
significant (at 0.15) and the correlation coefficients of both dependent variables with the similarity and
distance regressors are significant and comparable.

4.2 Econometric Specification

This section presents the results of panel data estimation allowing for fixed effects for both source and
destination countries, using the efficient estimation procedure proposed by Guimarães and Portugal
(2009):

12
 BHBijt = αi + αj + βInitialBHBij2001 + γln(Distanceijt )+

δImportsijt + ζCommonLanguageijt + ηCommonBorderijt +

θEUijt + τ EM Uijt + λOECDijt + µN AF T Aijt + πASEANijt , (15)

where the Initial bilateral home bias is the variable in 2001, Distance is the distance between capitals,
Imports is defined as a percentage of total imports of country i (similar to Bekaert and Wang, 2009),
Common Language/ Common Border are dummy variables that take the value 1 when countries share
a border/ a language spoken by more than 20% of the population and EU, EMU, OECD, NAFTA,
ASEAN are dummy variables that take the value 1 when both the source and the destination countries
are part of the respective international organizations: European Union (EU), European Monetary
Union (EMU), Organization for Economic Co-operation and Development (OECD), North American
Free Trade Association (NAFTA) and Association of Southeast Asian Nations (ASEAN).
Table 3 reports the results of using a gravity model to explain the dynamics of the bilateral home
bias, where the dependent variables are in turn: the I-CAPM bilateral home bias for the the full sample,
the I-CAPM bilateral home bias for the comparative sample and the Bayesian home bias. The empirical
analysis shows that bilateral home bias exhibits persistence (a high initial value leads to higher values
of home bias in the future) and confirms distance as a significant regressor, as higher distance results in
higher values of home bias. Trade links (proxied by bilateral imports as proportion of total imports of a
given country) matter robustly and home bias appears lower for countries sharing a common language
and more importantly a common border. Consistent to the descriptive statistics, sharing EU and
especially EMU membership lowers bilateral home bias significantly.
The similarity variable (orthogonalized on distance) is relevant and will be interpreted in terms of the
information versus trust paradigm proposed in Section 3.2.2. Similarity enters with correct (negative)
sign in all regressions, suggesting that home bias is lower towards similar (but not geographically
close) partners. However, the coefficient is smaller and no longer statistically significant when the
dependent variable (HB B) takes into account the financial market performance of the destination
country (the coefficient of the distance variable also diminishes in this model but maintains statistical
significance). The situation is repeated when the dependent variable becomes in turn the home bias in
absolute values (i.e. deviations from optimal investment) and the sample is restricted to the positive
values of the dependent variable (see Tables 4 and 5 respectively). The proposed interpretation and
future line of investigation is strengthened by the fact that similarity becomes once more statistically
significant (unlike distance) when the dependent variable is constrained to only the negative instances
of Bayesian home bias, i.e. overinvestment (see Table 6). Distance and other conveyors of information
(language, border) do not significantly attract a suboptimally high amount of resources, but similarity

13
(and imports) as well as participation in the international organization with the strongest ties, EMU do.
Notably, participation in looser international organizations EU, OECD, NAFTA appears significantly
linked with lower overinvestment (i.e. higher values of the dependent variable). While well used
information might lead the holder away from bad investments, trust (in this interpretation of similarity)
might be blind to negative consequences. Extreme situations such as the financial crisis could be the
eye opener. The coefficients of the similarity variable tend to increase during the financial crisis (see
Figure 4) suggesting that countries might be pursuing active strategies of diversification in reshuffling
their portfolios moving away from some of their similar counterparts.
While further work is needed to gauge the interpretation and impact of the variable in international
portfolio choices, the estimation results presented in this section firmly establish the significance and
information potential of similarity in gravity models. Above the informational content of distance (which
is linked with lower home bias), similarity significantly explains overinvestment in affine countries despite
their poor financial performance.

5 Concluding Remarks
This paper applies tools provided by state-of-the-art research in artificial intelligence to propose a
new variable of country similarity to gravity models. The Retina engine developed by Cortical.io
replicates the way the human brain works in synthesizing information and converts any text into a
numerical representation of its contextual meaning: its semantic fingerprint. The country similarity
measure used in this paper is the proportion of overlap between the semantic fingerprints of any two
countries. Similarity correlates well with distance and other typical gravity variables (common border
and language, for instance) but adds significant value to a gravity model applied to bilateral home bias
for a large set of countries. Using an alternative Bayesian measurement of home bias (that takes into
account financial market performance expressed in the evolution of asset returns) shows that aside from
the phenomenon of underinvestment in foreign partners, there are stable patterns of overinvestment,
when countries direct resources towards foreign partners whose financial performance does not seem
to justify the decision. Similarity as well as participation in a tight international agreement, such as
sharing the euro currency in the EMU, are significant drivers of overinvestment.

14
Appendix

A The Bayesian Framework

This appendix outlines the steps of deriving the moments of the predictive distribution of excess returns,
rt+1 , conditional on the set of sample data, Φ in terms of the prior and the likelihood function.
The Prior
The way in which the prior distribution incorporates the information given by the estimated intercept
reflects the degree of belief in the model. Complete belief in the model assumes that the eventual nonzero
intercepts are merely a result of sampling or estimation error and ignores them when computing the
expectations of excess returns (the fitted value of the dependent variable) while complete disbelief in
the model uses the sample mean as the estimate of expected returns. As our main interest lies in the
intercept it sufficient to construct a prior which is informative only with respect to α and diffuse (highly
volatile, non-informative) for the other parameters. Pástor (2000) chose a normal inverted Wishart
prior for the intercept:   
1
α|Σ ∼ N 0, σα2 , (A-1)
s2
with Σ following a inverted Wishart distribution: Σ−1 ∼ W H −1 , υ , H −1 the parameter matrix of the

Wishart distribution and υ, the degrees of freedom. The expectation of the inverted Wishart distribution
H
is given by E (Σ) = (υ−N −1) , where N is the number of asset returns in our time series. We can rewrite
the expectation for the prior residual covariance matrix, as E (Σ) = s2 IN , for H = s2 (υ − N − 1)
The prior involves a diagonal and homoskedastic covariance matrix for the residuals, which is set to
be non-informative by choosing υ=15, the equivalent of the sample of 15 observations. The prior of
homoskedasticity can easily be reversed under the pressure of data that enters the computation of the
posterior density. At this point, taking expectation of the conditional prior distribution of α, leads to
an unconditional distribution in the from:

α ∼ N 0, σα2 IN ,


(A-2)

where σα2 incorporates the degree of disbelief in the model. Based on the interpretation that the
intercepts that are different than zero reflect omitted sources of risk from the model, the size of this
mispricing is directly linked to the size of the residual covariance matrix. If the variance of the intercepts
has been large, the model is consequently less trusted. The asset pricing model is linear in the benchmark
risk factor, the world returns under the I-CAPM 12 : Rt = α+βFt +εt , assuming E (εt ) = 0, E (εt ε0t ) = Σ,
12 Pástor (2000) derives the results for the general case of N assets and K benchmarks. In the case of International
CAPM, the only benchmark is given by the world returns. Notation follows closely Asgharian and Hansson (2006).

15
  0
E (Ft ) = µt , E (Ft − µt ) (Ft − µt ) = ΩF , cov (Ft , εi,t ) = 0, ∀i = 1, N . The prior joint distribution is:

p (θ) = p (α|Σ) p (Σ) p (β) p (µF ) p (ΩF ) , (A-3)

where only the priors on the last three distributions are diffuse as derived by Pástor and Stambaugh
(2000): ( −1 )
1 0 σα2

− 12
p (α|Σ) ∝ |Σ| exp − α Σ α , (A-4)
2 s2
 
−
(υ+N +1) 1 −1
p (Σ) ∝ |Σ| 2 exp − trHΣ , (A-5)
2
p (β) ∝ 1, (A-6)

p (µF ) ∝ 1, (A-7)

p (ΩF ) = Ω−1
F . (A-8)

The Likelihood
In the linear model for asset returns, the disturbances are assumed uncorrelated and homoskedastic.
The benchmark returns are assumed i.i.d., normal, independent over time and independent of the error
terms. Under these independence assumptions, the likelihood function can be written as a product of
two normal likelihood functions, for the returns on the assets and respectively for the returns on the
benchmark factor:

p (Φ|θ) = p (R|θ, F ) p (F |θ) . (A-9)

The product terms are further expanded using computational results of Pástor and Stambaugh
(2000) into:
 
− T2 T −1 1 
−1 0


p (R|θ, F ) ∝ |Σ| exp − trΣ̂Σ − b − b̂ Σ ⊗ F F b − b̂ , (A-10)
2 2
 
T T 1 0 −1
p (F |θ) ∝ |ΩF |− 2 exp − trΩ̂F Ω−1
F − (µ F − µ̂F ) (µF − µ̂F ) Ω F , (A-11)
2 2
13 0
where b = vec (B) and B = (α β) .
The Posterior Density
We return to the key relation of Bayesian analysis, that defines the posterior distribution via pro-
portionality with the product of prior density and likelihood functions. Pástor and Stambaugh (2000)
combine the results for the priors with the ones for the likelihood functions separately for the regression
parameters and for the benchmark returns.
13 The transformation vec applied to a matrix, stacks its columns resulting into a vector.

16
 The posterior means of the model parameters result from:

b ≡ E (b|Φ) = IN ⊗ P −1 X 0 X b̂,


(A-12)

where b̂ the vector of OLS estimates of the model on the dataset, X = (ιT F ), P = S + X 0 X, D[2×2] is
s2
a matrix with the first element d(1,1) = 2
σα and the rest of the elements d(m,n) = 0, with m,n 6= 1.
The posterior variance of the model parameters is given by:

var (b|Φ) = Σ̃ ⊗ P −1 , (A-13)

(H+T Σ̂+B̂ 0 QB̂ )

, Q = X 0 IT − XP −1 X 0 X and Σ̂ and B̂ result from estimating


where Σ̃ = E (Σ|Φ) = T −υ−N −K−1

the model on the available sample.

Finally, the predictive means and variance of asset returns are defined using the posterior moments.
The predictive means can be computed as:

µ∗ ≡ E [RT +1 |Φ] = µ̃ = α̃ + β̃ µ̃F , (A-14)

where µ̃, α̃, β̃, µ̃F are posterior means and parameters.
The predictive variance-covariance matrix of asset returns is given by:

0
cov (Ri,T +1 Rj,T +1 |Φ) ≡ β̃i0 Ω∗F β̃j + tr [Ωcov (βi , βj |Φ)] + σ̃i,j + [1 µ̃0F ] cov bi , b0j |Φ [1 µ̃0F ] , (A-15)

where σ̃i,j is the respective (i, j) element of the posterior variance covariance matrix, Σ̃ and Ω∗F is
the predictive covariance matrix factor employed by the model explaining the returns: Ω∗F = Ω̃F +
T Ω̂F Ω̂F
var (µF |Φ), where Ω̃F = T −3 , var (µF |Φ) = T −3 .

The analytical result for the predictive variance-covariance matrix for the asset returns is:

cov (R, F |Φ) = β̃ Ω̃F + β̃var (µF |Φ) . (A-16)

B The Multi-Prior Framework

Garlappi et al. (2007) prove that the Multi-Prior optimization problem in the case when uncertainty
about the estimation of expected returns is expressed jointly for all assets, is equivalent to the maxi-
mization problem:

γ 0 √
max ω 0 µ − ω Σω − εω 0 Σω, (B-1)
ω 2
subject to
ω 0 ι = 1, (B-2)

17
where
(T − 1) N
ε= . (B-3)
T (T − N )
Without imposing short sales constrains, the problem can be solved analytically and the optimal
weights are given by:
√
σP∗ ε + γσP∗
   
∗ −1 1
ω =√ Σ µ̂ − B− ι , (B-4)
ε + γσP∗ A σP∗

where σP∗ is the variance of the optimal portfolio and the (unique) positive real solution to the polynomial
equation:
√
Aγ 2 σP4 + 2AγσP3 + Aε − AC + B 2 − γ 2 σP2 − 2γ εσP − ε = 0,


(B-5)

and A = ι0 Σ−1 ι, B = µ̂0 Σ−1 ι and C = µ̂0 Σ−1 µ̂.

18
References
Ahearne, A. G., W. L. Griever, and F. E. Warnock (2004). Information costs and home bias: An
analysis of U.S. holdings of foreign equities. Journal of International Economics 62, 313–336.

Asgharian, H. and B. Hansson (2006). Home bias among European investors from a Bayesian perspec-
tive. Journal of International Financial Markets, Institutions and Money 16, 397–410.

Bae, K., R. Stulz, and H. Tan (2007). Do local analysts know more? a cross-country study of the
performance of local analysts and foreign analysts. Working Paper .

Baele, L., C. Pungulescu, and J. R. Ter Horst (2007). Model uncertainty, financial market integration
and the home bias puzzle. Journal of International Money and Finance 26, 606–630.

Bekaert, G. and X. Wang (2009). Home bias revisited. Working Paper .

Bhuiyan, T. (2010). A survey on the relationship between trust and interest similarity in online social
networks. Journal of Emerging Technlogies in Web Intelligence 2, 291–299.

Brennan, M. J. and H. H. Cao (1997). International portfolio investment flows. Journal of Finance 52,
1851–1880.

Chan, K., V. Covrig, and L. Ng (2005). What determines the domestic bias and the foreign bias?
evidence from mutual fund equity allocations worldwide. Journal of Finance 60, 1495–1534.

Clark, G. L. (2015). Behavior, cognition and context. Working Paper .

Coval, J. D. and T. J. Moskowitz (1999). Home bias at home: Local equity preference in domestic
portfolios. Journal of Finance 54, 2045–2073.

De Roon, F. A. and T. E. Nijman (2001). Testing for mean-variance spanning: A survey. Journal of
Empirical Finance 8, 111–156.

De Santis, R. A. and B. Gérard (2006). Financial integration, international portfolio choice and the
European Monetary Union. ECB Working Paper No. 626 .

De Sousa Weber, F. (2015). Semantic folding theory and its applications in semantic fingerprinting.
Cortical.io White Paper .

Fidora, M., M. Fratzscher, and C. Thimann (2007). Home bias in global bond and equity markets: The
role of real exchange rate volatility. Journal of International Money and Finance 26, 631–655.

French, K. R. and J. M. Poterba (1991). Investor diversification and international equity markets.
American Economic Review 81, 222–226.

19
Garlappi, L., R. Uppal, and T. Wang (2007). Portfolio selection with parameter and model uncertainty:
A multi-prior approach. Review of Financial Studies 20, 41–81.

Gehrig, T. (1993). An information based explanation of the domestic bias in international equity
investment. Scandinavian Journal of Economics 95, 97–109.

Glassman, D. A. and L. A. Riddick (2001). What causes home asset bias and how should it be measured?
Journal of Empirical Finance 8, 35–54.

Grinblatt, M. and M. Keloharju (2000). The investment behavior, and performance of various investor
types: A study of Finland’s unique data set. Journal of Financial Economics 55, 43–67.

Guimarães, P. and P. Portugal (2009). A simple feasible alternative procedure to estimate models with
high-dimensional fixed effects. IZA Discussion Paper No. 3935 .

Guiso, L., P. Sapienza, and L. Zingales (2009). Cultural biases in economic exchange? Quarterly
Journal of Economics 124, 1095–1131.

Huberman, G. (2001). Familiarity breeds investment. Review of Financial Studies 14, 659–680.

Jenske, K. (2001). Equity home bias-can information cost explain the puzzle? Economic Review 86,
31–42.

Kogan, S., G. Salganik, F. Ibriyamova, and D. Stolin (2015). Using semantic fingerprinting in finance.
Working Paper .

Koop, G. (2003). Bayesian Econometrics. John Wiley and Sons Ltd.

Li, K. (2004). Confidence in the familiar: An international perspective. Journal of Financial and
Quantitative Analysis 39, 47–68.

Lintner, J. (1965). The valuation of risky assets and the selection of risky investments in stock portfolios
and capital budgets. Review of Economics and Statistics 47, 13–37.

Lütje, T. and L. Menkhoff (2004). What drives home bias? evidence from fund managers’ views.
Working Paper .

Markowitz, H. M. (1952). Portfolio selection. Journal of Finance 7, 77–91.

Mayer, T. and S. Zignago (2011). Notes on CEPII’s distance measures: The GeoDist database. CEPII
Working Paper No. 2011-25 .

Merton, R. C. (1980). On estimating the expected return on the market. Journal of Financial Eco-
nomics 8, 323–361.

20
Okawa, Y. and E. van Wincoop (2012). Gravity in international finance. Journal of International
Economics 87, 205–215.

Pástor, L. (2000). Portfolio selection and asset pricing models. Journal of Finance 55, 179–223.

Pástor, L. and R. F. Stambaugh (2000). Comparing asset pricing models: An investment perspective.
Journal of Financial Economics 56, 335–381.

Portes, R. and H. Rey (2005). The determinants of cross-border equity flows. Journal of International
Economics 65, 269–296.

Sharpe, W. (1963). A simplified model for portfolio analysis. Management Science 9, 277–293.

Tesar, L. L. and I. Werner (1995). Home bias and high turnover. Journal of International Money and
Finance 14, 467–492.

Tett, G. (2013). Anthropology and the global financial crisis. Kroeber Anthropological Society Pa-
pers 102, 25–32.

Veldkamp, L. and S. Van Nieuwerburgh (2006). Information immobility and the home bias puzzle.
Working Paper .

Warnock, F. E. (2001). Home bias and high turnover reconsidered. International Finance Discussion
Paper No. 702 .

21
 Table 1: Descriptive Statistics
This table presents some descriptive statistics (total number of observations (No. Obs.), percentage of
negative observations (% Negative ), mean of the sample (Mean), standard deviation (Std. Deviation),
minimum (Min) and maximum (Max) for three measures of home bias: Home Bias I-CAPM Full Sample
(HB I), Home Bias I-CAPM Comparative Sample (HB IC) and Bayesian Home Bias (HB B) for the full
sample as well as for subsamples with both source and destination countries members of an international
organization: OECD, EU, EMU, NAFTA and ASEAN.

Sample Measure No. Obs. % Negative Mean Std. Deviation Min Max
Full HB I 48628 0.06 0.89 0.27 -0.81 1
HB IC 30866 0.05 0.88 0.26 -0.81 1
HB B 30866 0.47 0.50 0.51 -0.82 1
OECD HB I 9288 0.11 0.72 0.35 -0.51 1
HB IC 8396 0.09 0.74 0.32 -0.16 1
HB B 8396 0.43 0.50 0.49 -0.27 1
EU HB I 5323 0.23 0.60 0.42 -0.51 1
HB IC 4414 0.20 0.61 0.40 -0.19 1
HB B 4414 0.59 0.34 0.48 -0.27 1
EMU HB I 1949 0.27 0.52 0.43 -0.51 1
HB IC 1685 0.22 0.56 0.40 -0.19 1
HB B 1685 0.56 0.35 0.48 -0.27 1
NAFTA HB I 66 0.00 0.73 0.21 0.40 1
HB IC 66 0.00 0.73 0.21 0.40 1
HB B 66 0.06 0.78 0.28 -0.02 1
ASEAN HB I 132 0.23 0.71 0.43 -0.02 1
HB IC 132 0.23 0.71 0.43 -0.02 1
HB B 132 0.70 0.28 0.45 -0.06 1

22
 Table 2: Correlation Coefficients: Main Variables
This table presents the correlation coefficients between the main variables: Home Bias I-CAPM Full Sample (HB I), Bayesian Home Bias (HB B),
the semantic overlap measure (Similarity), logarithm of the distance between capital cities (Distance), the proportion of bilateral imports in
total imports of source country (Imports), common language (Language), common border (Border), as well as dummy variables for both source
and destination countries belonging to EU, EMU, OECD, NAFTA or ASEAN.

HB I HB B Similarity Distance Imports Language Border OECD EU EMU NAFTA ASEAN

HB I 1 0.15 -0.39 0.47 -0.27 -0.11 -0.29 -0.33 -0.42 -0.29 -0.03 -0.04
HB B 1 -0.09 0.09 0.02 0.04 -0.05 0.01 -0.12 -0.06 0.03 -0.03
Similarity 1 -0.58 0.32 0.14 0.38 0.28 0.43 0.22 0.02 0.10

23
Distance 1 -0.34 -0.03 -0.41 -0.32 -0.63 -0.27 -0.05 -0.07
Imports 1 0.19 0.42 0.16 0.18 0.14 0.27 0.05
Language 1 0.16 -0.05 -0.10 -0.04 0.08 0.04
Border 1 0.12 0.17 0.14 0.15 -0.01
OECD 1 0.35 0.23 0.08 -0.04
EU 1 0.48 -0.02 -0.03
EMU 1 -0.01 -0.02
NAFTA 1 0.00
ASEAN 1
 Table 3: Gravity Model Results (Full Sample)
This table reports the results of (fixed effects) panel regressions of bilateral home bias computed in
the I-CAPM framework for the full sample (HB I) and comparative sample (HB IC) and the Bayesian
framework (HB B) on: the initial level of bilateral home bias (Initial), logarithm of distance between
capital cities (Distance), the semantic overlap measure, orthogonalized on distance (Similarity), the
proportion bilateral imports in total imports per source country (Imports), common language (Lan-
guage), common border (Border), as well as dummy variables for both source and destination countries
belonging to EU, EMU, OECD, NAFTA or ASEAN. Fixed effects are included for source and destina-
tion countries. Values of the coefficients, robust T-statistics and R2 , are reported. Significance of the
coefficients is denoted by *** (at 1%), ** (at 5%) and * (at 10%).

Dependent Var. : HB I HB IC HB B HB I HB IC HB B
No. Obs. 46463 22246 22246 46463 22246 22246
Initial 0,5740*** 0,5955*** 0,2798*** 0,5673*** 0,5887*** 0,2795***
(17,74) (15,62) (13,23) (17,54) (15,42) (13,25)
Distance 0,0306*** 0,0360*** 0,0130*** 0,0262*** 0,0329*** 0,0126***
(3,90) (4,17) (3,88) (3,42) (3,81) (3,75)
Similarity -0,3101*** -0,2211*** -0,0297
(-4,38) (-3,06) (-0,72)
Imports -0,4221*** -0,3110** -0,2162*** -0,3949*** -0,3028*** -0,2150***
(-3,06) (-2,59) (-2,82) (-2,93) (-2,59) (-2,82)
Language -0,0322** -0,0207* -0,0061 -0,0207* -0,0118 -0,0049
(-2,44) (-1,83) (-1,55) (-1,68) (-1,05) (-1,20)
Border -0,0360** -0,0597** -0,0163 -0,0330* -0,0599** -0,0162
(-2,00) (-2,34) (-1,01) (-1,78) (-2,34) (-1,01)
EU -0,0887*** -0,0571*** 0,0121* -0,0808*** -0,0513** 0,0129*
(-4,59) (-3,00) (1,74) (-4,16) (-2,44) (-1,80)
EMU -0,0478** -0,0621** -0,0224*** -0,0473** -0,0612** -0,0222***
(-2,11) (-2,48) (-2,94) (-2,06) (-2,44) (-2,94)
OECD -0,0353*** -0,0194 -0,0009 -0,0283** -0,0149 -0,0003
(-2,64) (-1,20) (-0.11) (-2,13) (-0,92) (-0,03)
NAFTA 0,1178** 0,1323*** 0,0174 0,0860* 0,1135*** 0,0147
(2,28) (3,73) (0,38) (1,68) (3,22) (0,32)
ASEAN -0,0040 0,0056 -0,0092 -0,0024 0,0072 -0,0089
(-0,06) (0,09) (-0,41) (-0,03) (0,11) (-0,40)
2
R 0.62 0.71 0.58 0.62 0.72 0.58
24
Table 4: Dependent Variable: Bilateral Home Bias in Absolute Terms (i.e. Deviations from
Optimality)
This table reports the results of (fixed effects) panel regressions of the absolute values of bilateral home
bias computed in the I-CAPM framework for the full sample (HB I) and comparative sample (HB
IC) and the Bayesian framework (HB B) on: the initial level of bilateral home bias (Initial), distance
between capital cities (Distance), the semantic overlap measure, orthogonalized on distance (Similarity),
the proportion of bilateral imports in total imports of source country (Imports), common language
(Language), common border (Border), as well as dummy variables for both source and destination
countries belonging to EU, EMU, OECD, NAFTA or ASEAN. Fixed effects are included both for
source and destination countries. Values of the coefficients, robust T-statistics and adjusted R2 , are
reported. Significance of the coefficients is denoted by *** (at 1%), ** (at 5%) and * (at 10%).

ABSOLUTE VALUES HB I HB IC HB B
No. Obs. 46463 22246 22246
Initial 0,5414*** 0,5720*** 0,2693***
(17,00) (15,77) (13,28)
Distance 0,0260*** 0,0335*** 0,0108***
(3,43) (4,07) (3,62)
Similarity -0,3086*** -0,2224*** -0,0104
(-4,38) (-3,08) (-0,22)
Imports -0,3456*** -0,2351*** -0,1215*
(-2,75) (-2,18) (-1,79)
Language -0,0229* -0,0135 -0,0073*
(-1,90) (-1,23) (-1,89)
Border -0,0304* -0,0537** -0,0107
(-1,72) (-2,16) (-0,70)
EU -0,0787*** -0,0502** 0,0085
(-4,22) (-2,63) (1,19)
EMU -0,0347 -0,0579** -0,0154**
(-1,57) (-2,35) (-2,10)
OECD -0,0233** -0,0164 -0,0053
(-2,11) (-1,03) (-0,72)
NAFTA 0,0721 0,0975*** -0,0094
(1,44) (2,95) (-0,21)
ASEAN -0,0096 0,0037 -0,0121
(-0,13) (0,06) (-0,59)
R2 25
0.61 0.72 0.55
 Table 5: Dependent Variable: Positive Instances of Bilateral Home Bias
This table reports the results of (fixed effects) panel regressions of the positive instances of bilateral
home bias computed in the I-CAPM framework for the full sample (HB I) and comparative sample (HB
IC) and the Bayesian framework (HB B) on: the initial level of bilateral home bias (Initial), logarithm
of the distance between capital cities (Distance), the semantic overlap measure, orthogonalized on
distance (Similarity), the proportion of bilateral imports in total imports of source country (Imports),
common language (Language), common border (Border), as well as dummy variables for both source
and destination countries belonging to EU, EMU, OECD, NAFTA or ASEAN. Fixed effects are included
both for source and destination countries. Values of the coefficients, robust T-statistics and adjusted
R2 , are reported. Significance of the coefficients is denoted by *** (at 1%), ** (at 5%) and * (at 10%).

POSITIVE VALUES HB I HB IC HB B
No. Obs. 43603 21185 11906
Initial 0,3799*** 0,4612*** 0,2701***
(11,34) (13,31) (10,06)
Distance 0,0110*** 0,0207*** 0,0089**
(3,49) (4,42) (3,62)
Similarity -0,1892*** -0,1200** -0,0273
(-5,01) (-2,20) (-0,49)
Imports -0,2380*** -0,1318*** -0,1619**
(-2,38) (-2,68) (-2,40)
Language -0,0141** -0,0206*** -0,0049
(-1,18) (-2,68) (-0,79)
Border -0,0671 -0,0445** -0,0275
(-1,18) (-2,27) (-1,51)
EU -0,0296*** -0,02392** -0,0142
(-3,08) (-2,21) (-1,40)
EMU -0,0573*** -0,0580** -0,0210
(-3,32) (-2,98) (-1,48)
OECD -0,0525*** -0,0261** -0,0128
(-4,94) (-2,35) (-1,57)
NAFTA 0,0728 0,0461 0,0101
(0,13) (1,46) (0,24)
ASEAN 0,0271 0,0268* -0,0057
(1,34) (1,99) (-0,24)
R2 0.54 0.68 0.41
26
 Table 6: Dependent Variable: Negative Instances of Bilateral Home Bias
This table reports the results of (fixed effects) panel regressions of the negative instances of bilateral
home bias computed in the I-CAPM framework for the full sample (HB I) and comparative sample (HB
IC) and the Bayesian framework (HB B) on: the initial level of bilateral home bias (Initial), distance
between capital cities (Distance), the semantic overlap measure, orthogonalized on distance (Similarity),
the proportion of bilateral imports in total imports of source country (Imports), common language
(Language), common border (Border), as well as dummy variables for both source and destination
countries belonging to EU, EMU, OECD, NAFTA or ASEAN. Fixed effects are included both for
source and destination countries. Values of the coefficients, robust T-statistics and adjusted R2 , are
reported. Significance of the coefficients is denoted by *** (at 1%), ** (at 5%) and * (at 10%).

NEGATIVE VALUES HB I HB IC HB B
No. Obs. 2860 1061 10340
Initial 0,0076 0,0021 0,0067**
(1,07) (0,27) (2,25)
Distance 0,0057 0,0045 0,0013
(1,27) (1,20) (1,55)
Similarity -0,0691 0,0149 -0,0158***
(-1,53) (0,40) (-2,96)
Imports -0,0200 -0,0786 -0,1116***
(-0,24) (-0,98) (-2,81)
Language -0,0053 -0,0004 0.0011
(-0,73) (-0,05) (1,06)
Border -0,0018 -0,0062 0,0004
(-0,17) (-1,36) (0,13)
EU -0,0111** -0,0083** 0,0051***
(-2,57) (-2,47) (3,05)
EMU -0,0188*** -0,0212** -0,0049***
(-2,77) (-2,33) (-3,35)
OECD -0,0167** -0,0103 0,0020*
(-2,57) (-0,97) (1,70)
NAFTA 0,0071**
(2,17)
ASEAN -0,0388** 0,0249** 0,0057***
(-2,46) (-2,47) (3,05)
R2 0.51 0.47 0.50
27
 Figure 1: High Semantic Overlap: Spain - Portugal
This figure gives an example of high semantic overlap between the fingerprints of Spain and Portugal,
obtained from the interactive tool of Cortical.io at: http://www.cortical.io/demos/similarity-explorer/.

28
 Figure 2: Low Semantic Overlap: Poland - Bermuda
This figure gives an example of low semantic overlap between the fingerprints of Poland and Bermuda,
obtained from the interactive tool of Cortical.io at: http://www.cortical.io/demos/similarity-explorer/.

29
 Figure 3: Similarity Measure: Histogram
This figure uses a histogram to illustrates the distribution of the similarity measure based on the
semantic overlap between the fingerprints of all country pairs in the dataset.

SIMILARITY

600

500

400
Frequ ency

300

200

100

0
.00 .05 .10 .15 .20 .25 .30 .35 .40 .45 .50 .55

30
 Figure 4: Time Varying Over/Under Investments and Crisis Effects
This figure shows time varying estimated coefficients for the distance and similarity explanatory variables as well as the yearly averages of the
dependent variables: Positive/Negative Instances of Home Bias I-CAPM (Comparative Sample) and Bayesian Home Bias (HB B) based on
yearly repetitions of the regressions presented in full in Table 5 (for positive HB) and respectively Table 6 (for negative HB). The shaded area
covers the period of the financial crisis (2007 - 2009).

Posi tive Hom e Bias I -CAP M (Comparati ve S ample) P osit ive Hom e Bias Bay esian

1.0 1.0
.05 .2

.00 0.8 0.8

.1
-.05 0.6 0.6

-.10 .0

31
0.4 0.4

Coef fic ient s

Coef fic ient s

Av erage HB +
Av erage HB +

-.15
0.2 -.1 0.2
-.20

0.0 0.0
-.25 -.2
2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012

Av erage HB+ Distance S imilarity Av erage HB+ Distance Simil arit y

Negative Home Bi as I -CAP M (Comparat ive Sample) Negat ive Hom e Bias B ay esian

-.010 .00

-.012 -.02

-.014
-.04
-.016
.1 -.06
-.018 .005

.0 -.020 .000 -.08

Coef ficient s
Coef ficient s

Average HB -
Average HB -

-.005
-.1
-.010
-.2
-.015

-.3 -.020
2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012

Av erage HB - Distance Simil arit y Av erage HB - Distance S imilarity