The decomposition methodology of Heston and Rouwenhorst (1994) has a number of attractive features. Above all, it is capable to clearly and separately attribute the variation in asset returns to the factors incorporated in the model. In the case of my sample of 6,440 eurobond returns from predominantly European countries for May 1990 to March 2008, it is thus established that country effects account for more of the return variance than industry effects (in Chapter 4). The portion that cannot be attributed to either country or industry effects resides with the common factor of which it is determined that this is largely the conversion effect of local currency to common currency returns. This is again a clear-cut result for eurobond returns and a great quality of the decomposition model that it can so unambiguously extract and capture these distinct effects.
It is tempting to conclude from these results that a geographical portfolio allocation yields better results than an industry sector portfolio allocation for eurobonds in Europe. I am careful not to draw such stark conclusions from the decomposition analysis on superior or even preferable bond portfolio allocation strategies for two reasons. One, because the assumptions innate to the decomposition model can prejudice the measure for country and industry effects. Two, because there are limitations with respect to the interpretation of the results from the model calculations. Each of these deserve further elaboration, if only because it points out directions for further research.
The two assumptions inherent to the Heston and Rouwenhorst decomposition model that have drawn the most criticism are of unit and constant factor exposures. First, consider the assumption of unit factor exposure, which is built in because all assets under consideration within a country and within an industry sector have a unit exposure, or beta, to global market shocks. This is in contrast with mainstream asset pricing theory (such as CAPM), which contends that differences in systemic risks across firms are precisely to be explained by their betas, i.e. their exposure to common market shocks. Brooks and Del Negro (2002, 2004) are often quoted in this respect because they demonstrate, as one of the first, that the assumption of unit betas is less well supported by the data. Their latent factor model relaxes the restriction that all stocks with exposure to a given shock have the same exposure to that shock and nests the fixed effects model of Heston and Rouwenhorst. Their results also uncover that many industry betas are negative while almost all country betas are positive, at least for their sample of USD-denominated stock returns from 1985 to 2002 from a broad range of countries. These differences within groups in beta-heterogeneity are thought to be the reason why country factors have historically outweighed industry factors in
explaining international return variation. Their research has set up a strand of literature where factor coefficients are left unconstrained (though zero restrictions on the exposures to other countries or industry sectors are maintained, which would otherwise lead to the identification problem familiar from standard
exploratory factor analysis). De Moor and Sercu (2006) is a more recent example of this type of studies. Via a two-stage estimation approach, provisionally estimated factor returns determine sensitivities through time series OLS and then the revised factors are extracted from cross-section regressions on these estimated sensitivities. They also verify that this makes a difference.
Secondly, the assumption of constant factor exposures is inherent in Heston and Rouwenhorst’s decomposition model because the factors driving country and industry-affiliation in the standard
decomposition model have very little to no dynamics. This is in contrast with considerable evidence from current literature of time-varying betas. Consequently, there is another strand of literature motivated by this evidence that attempts to overcome these limitations. Examples are Bekaert, Hodrick and Zhang (2009) who use an arbitrage pricing theory (APT) model where the identity of the important systemic factors may change over time. Likewise in Baele and Inghelbrecht (2009, 2010), a GARCH-framework explicitly allows both factor exposures and asset-specific volatilities to vary over time.
It is entirely plausible that models that overcome the restrictive assumptions of unit and constant betas provide a better fit for asset returns data, as all studies discussed so far claim. That in itself has made a commendable contribution to the research field. Yet, it is doubtful that this direction of research will provide any further insights into my key question of optimal portfolio diversification strategies in European bond markets. For one, both strands of literature, with their various additions and estimated in two or sometimes even three stages, continue to rely on linear factor specifications, which is the main trait of the decomposition model. Secondly, results from these studies have by and large left the overall conclusion of domineering country effects over industry effects in stock return variations intact. If a rise in the primacy of industry versus country effects is detected, then this is not recognized as a lasting trend but rather as a temporary phenomenon (e.g. Bekaert, Hodrick and Zhang, 2009). Thirdly, these studies, and especially the time-varying type, are mostly concerned with the explanation of the empirically observed time variation of stock (and recently also bond) return correlations and covariances, as the most recent study from this group of authors confirms (Baele, Bekaert and Inghelbrecht, 2010). Through the better identification of asset return correlation and covariance structures and varying determinants over time, these studies aim to dynamically identify factors as sources that most contribute to risk diversification in a portfolio setting. By extension of the standard Heston and Rouwenhorst model, which these studies essentially are, therein also lies the limitations of the interpretation of the results that are obtained from it.
2.1. Alternative methodologies
In essence, the Heston and Rouwenhorst decomposition of stock returns model determines the extent to which separate factor-related effects explain their return variation. In so far as the objective of the
portfolio manager is to reduce risk from these variation or risk-contributing factors, is a valid conclusion that “diversification across countries within an industry is a much more effective tool for risk reduction than industry diversification within a country” (Heston and Rouwenhorst, 1994, p3). The proper identification of the importance of country and industry (and potentially other) effects in asset return variation is an important achievement in itself, but little more in terms of portfolio design can justifiably be concluded from it. I concur with De Moor and Sercu (2009, p 6) when they state that “the Heston and Rouwenhorst (1994) methodology does not tell us anything about the correlations among sectors or countries, and no conclusion can therefore be made to international risk diversification”. While more recent studies
highlighted earlier go much further in explaining asset return comovements, theirs remain bound too to the identification of the main sources of reduction of variation (or volatilities) in portfolio returns. Portfolio allocation is about more than just risk reduction.
It seems altogether that a complementing tool is needed to enable any qualifications of country versus industry portfolio construction. Certainly if any statements wish to be made on the portfolio allocation strategies in terms of the twin objectives of return optimization in relation to the amount of risk a portfolio manager is willing to take, as is often the case in fund management practice. This is offered by a strand of literature that adopts a model-based approach in the outright comparison of the performance of country and industry-based portfolios. Building on the analytical work of Huberman and Kandel (1987) and De Roon and Nijman (2001), Moerman (2004, 2008) and Eiling et al. (2006) evaluate the risk-return characteristics of equity portfolios constructed from a country and industry allocation. Taking an investor’s perspective, the starting point is a Markowitz-style mean-variance optimization problem from which efficient frontiers can be created for country and industry portfolios (and both combined, which by nature is always best). So-called spanning and efficiency tests are devised which are able to establish whether a country or industry portfolio has superior mean-variance properties. Spanning tests show whether the inclusion of extra investment opportunities enlarges the efficient set of portfolios. They are conducted by Moerman (2004, 2008) using the MSCI equity indexes for Euro zone countries (excluding Luxembourg) over the period 1995-2004. He finds that a stock investor is better off diversifying over industries rather than over countries, both for the full sample period and for two subperiods around EMU, but especially post- EMU. Moerman’s result is robust when sectors affected by the IT-bubble are neglected, though the difference in performance between the country and industry diversification becomes less pronounced. Eiling et al. (2006) conduct similar spanning tests for the Euro zone’s equity markets for the period 1990- 2003. They also introduce efficiency tests, which test for the difference in maximum Sharpe ratios between cross-country and cross-industry diversified portfolios. Eiling et al. find, however, that country and industry- based portfolios cannot be distinguished in terms of mean-variance performance and Sharpe ratios. Their style analysis, which examines the ability to replicate the variation of industry portfolio returns with
country indexes and vice versa, suggests an increasing relative importance of industry effects over country effects. This style analysis is further elaborated upon and comparable results confirmed in a subsequent study of Euro zone equity returns over the 1990 to 2008 period (Eiling et al., 2010). These latter results of the increased importance of industry effects are obtained through a related but different methodology (i.e. style analysis) from spanning and efficiency tests though.
The performance of mean-variance tests of spanning and efficiency provides yet another perspective in addition to the results obtained from the decomposition model in Chapter 4 on optimal portfolio diversification on a country or industry sector basis. For this reason and because this analysis has so far, to the best of my knowledge, not been performed on bonds, do I prefer to go down this route in this second chapter of the empirical analysis.