Contributor: Laurence Wormald
It’s a truth universally acknowledged that a fund manager seeking to take effective control of investment risk across asset classes must be using or looking to use risk budgeting. However, it may be a lesser known fact that to implement risk budgeting successfully, the most important element is a proper estimation of cross-asset correlations.
There are those who allocate risk to country/region/asset class desks without managing their interaction. This group may be faced with regular surprises, as these interactions can either lead to risk increasing or risk reducing. There can be no expectation of reward without a proper estimate of the risk incurred in these interactions, and thus the portfolio is probably less efficient than it could be.
On the other side, there are those who budget properly to create a risk reduction via cross-asset-class correlations. This group can take more risk in the alpha generation process – a genuine benefit to the investment manager from diversification.
Being able to monitor actual risk expenditure against budget, via what is sometimes called “covariance accounting,” to break down tracking error or total volatility in terms of the cross-asset-class correlations is a vital investment capability that requires a robust underlying risk model.
The robustness of the correlation estimates in the multi-asset class models is ensured by using the same principal components methodology which SunGard’s APT uses for single-asset-class models. Principal components analysis is a powerful technique for separating signals from noise in any dynamic system, and that is what is required to robustly estimate the systematic correlations between assets within different classes (such as sovereign and corporate bonds, equities, and commodities). Historical data is always noisy, but by using principal components techniques, that noise can be effectively filtered out before estimating the systematic correlations.
However, as the last few years have shown, risk management is an art as well as a science. The art comes in choosing what we call the “Estimation Core,” that is the set of assets and macro factors used in estimating the APT components and in choosing the most appropriate number of principal component factors for each multi-asset class model.
Our first criterion for inclusion in the estimation core is that the historical data be fully validated, so that problems associated with stale pricing or missing returns are minimized. For a model based on weekly data, we require 180 weeks of scrubbed returns data for any asset which will be included in the estimation core.
SunGard APT’s research team uses extensive back testing in selecting the best-validated datasets for each asset class, including the most important macro series (selected from a set of approximately 30 macro series such as FX rates, key interest rates and credit spreads, equity, commodity, and volatility Indices) across all assets before extracting the principal components. In selecting, we look for significant explanatory power both in periods of normal market behavior and during market crises, considering crisis scenarios back to the 1990s.
Next, we separately choose estimation cores for each of the major asset classes (consistent with those for our single-asset-class models) before combining them in the multi-asset class estimation. In this way, corporate bonds, for example, inherit exposure to their issuers as well as to rates and credit spreads, while the cross-correlation of equities to commodities is simultaneously estimated with that to FX.
Finally, we test on random matrix theory to check that the principal components are in fact capturing the behavior of the systematic driving terms which drive cross-asset-class correlations. This approach makes the multi-asset class models robust enough for FactSet users to estimate a complete set of cross-asset correlations for use in risk reporting, risk budgeting, and optimized portfolio construction. For more information on optimized portfolio construction, view our recorded webinar on robust optimization.