Put allocation at the core
As shown by Solnik (1993), global asset allocation is the largest source of differences in performance among global portfolios.1 However, institutional investors have to be reluctant and refrain from giving too much attention to active asset allocation versus within-asset class selection. The risk of active asset allocation strategies is high. The reason for this is that asset allocation typically involves 3 asset classes (equities, bonds and cash) and 20 or more countries, so that the total number of investable ‘securities’ is relatively small when comparing it with asset selection portfolios that comprise investment spectrums with 100 to several thousand ‘securities’.2 This implies that the average ‘bet size’ is relatively large. According to Dick Grinold’s so-called Fundamental Law of Active Management this will automatically translate into a situation in which asset allocation managers will have to overcome a high skill hurdle to compete with asset selection.3 However, it is also true that these strategies offer interesting opportunities for increased portfolio diversification.
Observing the behaviour of a lot of well-respected international pension plans, we do see that they indeed tend to focus more on asset selection than asset allocation in their investment process. The tactical asset allocation decision is more or less statically derived from the target or strategic asset allocation. The target allocation is the result of an asset-liability management (ALM) study that focuses on the long-term return and risk profiles of the various asset classes and countries within these asset classes, their correlations and liability-related factors. This creates a situation in which the asset allocation process resembles ‘benchmarking’, nothing more and nothing less. Using the fact that asset allocation has a larger overall influence on performance than asset selection one could say that the ‘winners’ in WM’s surveys of pension plan returns are in many cases the ones that:
q either had less stringent restrictions on the liability side, so that they could invest more in equities; or
q were able to generate better return, risk and correlation forecasts as inputs for their ALM optimisation process; or
q were just lucky!
Most pension plans use some set of historical inputs with respect to return, risk and correlation coefficients, which implies that the secondexplanation is probably less important than the other two. Does this mean that pension plans knowingly and willingly translate the potentially high risks surrounding (tactical) asset allocation decisions into some form of passive management of this part of the investment process? In my opinion, the answer to this question is both ‘yes’ and ‘no’. ‘Yes’ in that they refrain from excess switching between asset classes and/or taking too large an active bet compared to the strategic benchmark. ‘No’ in that decisions regarding strategic benchmark choice and especially deviations from that active benchmark are traditionally based on relatively simple decision models such as:
q mean-variance optimisations with historical returns, risks and correlations; or
q ‘ad hoc’ decisions based on implicit ‘models’ or investment policy committee meetings without any robust ex-post performance attribution analysis.
In many instances pension plans spend relatively little time and manpower on the asset allocation decisions although they know that it is the most important explanatory variable when it comes to disentangling their overall performance. In another contribution to IPE, (June 1999, p44) I concluded that a change is on its way with more and more plans structuring the asset allocation decisiontaking process and giving it a more prominent place within the overall investment process. Some of the larger and more sophisticated plans might opt for internal development of asset allocation models or frameworks, others might turn to the asset management community for help. Markowitz & van Dijk (2000) shows that an important first step has to be the integration of strategic and tactical asset allocation as a two-way information processing system in which:
q long-term benchmark weights are derived and applied when deviating tactically; and
q shorter-term information (on the market or macro level, or from the performance attribution analysis of previous decisions) translates into important signals for future strategic benchmark deduction.4
Knowing that the average ‘bet size’ is large, pension plans should ‘create’ as many different asset class/country combinations as possible. The more explicit and detailed the asset allocation process, the easier it is to control the risks and add value. This goes way beyond the old idea that performance is all about ‘picking bonds or stocks’. And with the world getting smaller, the correlation coefficients increase.Hence the need for assets with ‘non-standard’ return-risk-correlation characteristics is growing even further.
Grinold & Kahn (2000) correctly stated that one has to take into account that there is a structural difference between asset selection and asset allocation models. In the asset selection case the potential for positive information ratios is mainly enlarged by:
q proper benchmarking and guidelines for the portfolio manager in combination with relatively-high intra-asset class correlations; and
q the availability of relatively large spectrums of investable securities at any specific moment in time.
A state-of-the-art cross-sectional analysis (without worrying too much about the time-series of returns, risks and explanatory variables) will help increase portfolio information ratios.
Time-series analysis is much more important in asset allocation models and that might be a totally new experience for a lot of decision takers within the pension plan community. And it might even be new to many European asset managers who are now confronted with the need to know more about sophisticated asset allocation and country selection with the definition of ‘domestic’ now being Europe or Euroland instead of an individual country. In asset allocation models or frameworks it is not so much about ‘security X or security Y’ but much more about ‘Asset class X in country A: Now or later?’ In the latter case the whole sequence of ‘now’ or ‘later’ type of answers for all asset classes and countries will then translate into an asset allocation portfolio. Sometimes this might even imply that a number of ‘later’ decisions translate into ‘now’ actions anyway! The currency factor and the smaller spectrum make a direct cross-sectional comparison of individual ‘securities’ much harder in the asset allocation case.
And it is not just the structural difference in model type (time-series instead of cross-sectional regression framework) which makes asset allocation modelling difficult. Another problem is, we are probably forced to work with monthly observations because there are not enough high-quality quarterly observations available on many relevant variables. For example, a model using quarterly data on gross domestic product to predict quarterly returns on stocks or bonds would, even if reliable GDP data were available for the whole post-second world war period, still have only a mere 220 (55 years x 4) observations. Working with monthly data quadruples the number of observations, but independent and reliable monthly data are often not available for important (macro-) economic time-series. Another problem is that macroeconomic data are often published relatively late. We also observe in many instances that initially reported values are later corrected. And it is exactly the factor block with macro variables that seems to play a more important role in asset allocation modeling than in asset selection modeling. We do therefore need to accept that even our ‘ideal’ model will not show tremendous R-squares.5 Many variables that we would love to incorporate in the model structure fall immediately out of it due to lack of observations or statistical significance with the relatively small data set we have. When we take into account that the data set will also have to be split somehow to enable out-of-sample testing the problem gets even bigger. It is therefore probably inevitable to construct a framework that integrates quantitative and qualitative techniques since we do not a-priori believe that ‘man’ cannot add value.
Notwithstanding the problems and difficulties that we described in the previous paragraphs, we do feel that sophisticated players in the institutional money management community should take asset allocation seriously and make it the core of their investment process. The combination of potentially high excess returns and high risk warrant above-average attention instead of a seemingly passive treatment of the issue. Especially because passivity is in many instances based on a not too sophisticated strategic benchmarking process that suffers from subjectivity, infrequent updating and naïve basic assumptions. And even if the result of our structured research efforts is that we are not able to generate statistically significant excess returns for all asset class/country combinations, we should still try to make the most of our information. This means that we have to be active when possible and passive when information coefficients are close to zero. Relatively low R-squares and models that are not as ‘ideal’ as we would want them to be will then – in combination with the high risk profile of asset class deviations compared to benchmark weight deviations in asset selection strategies – translate into relatively small absolute active weights vis-à-vis our strategic benchmark. But that is in itself not a big problem, because the relative added value of a one percentage-point benchmark deviation in asset allocation strategies is higher than the same deviation in an asset selection strategy.
Erik van Dijk is CEO at Palladyne Asset Management in Amsterdam
1 Solnik, B., Predictable Time-Varying Components of International Asset Returns, Research Foundation of the Institute of Chartered Financial Analysts, Charlottesville (Virginia), 1993.
2 We can of course add more asset categories (like for example real estate, private equity, commodities etc.) and/or countries, but the relative ‘bet size’ will remain large compared to asset selection.
3 See for example Grinold & Kahn’s excellent handbook on active investment management: Grinold, R.C. & R. Kahn, Active Portfolio Management, McGraw-Hill, New York, 2000.
4 Markowitz, H.M., and Erik L. van Dijk, ‘Single-Period Risk-Return Analysis in a Changing World’, Palladyne Research Paper, November 2000.
5 R-squares in the range of 0 to 30% are to be expected in a well-functioning asset allocation framework with monthly observations with the majority of time-series models for individual asset class/country combinations ending somewhere in the 10-20% range. But researcher will find that a lot of work has to be done to get to that result.