Technology

Customised

Using a customised risk model enables an investment manager to explicitly identify and quantify the various exposures being made in his or her portfolio, in terms that correspond directly to their investment process. The model will clearly show whether the manager is making any significant factor bets besides those that were intended.

Our risk models are Customised in almost every respect, from the extent of the stock universe included and the industry classification used, to the definition and ordering of the specified factors.

Transparent

All our models are fully transparent in the sense that all intermediate computations are visible to the user. The obvious advantage of this is that it enhances understanding of the sources of risk, which we believe to be the most important function of risk profiling.

Emphasis on accuracy in covariance matrices

The covariance matrix is the engine of the risk model: any errors in its estimation will flow directly through to the portfolio risk estimation and decomposition. We therefore give priority to its accuracy. By contrast, errors in individual stock betas can be expected to cancel out at the portfolio level and are therefore unlikely to bias the resulting risk estimates.

Realistic betas

A popular approach to assigning betas is to use “dummy variables” – in effect an arbitrary value of 1 or 0 (on or off) to determine the relationship of a stock to a factor. For example, UBS, being a Swiss bank, would have a 1 for Switzerland and Banks, and 0 for all other countries and industries. This flies in the face of common observation and, unsurprisingly, compromises the results given by models that use them.

The R-Squared approach is to estimate factor betas from the model’s sample data, so UBS can be expected to have non-zero betas to countries other than Switzerland and possibly industries other than banks (consider General Electric’s relationship to industries other than Consumer Durables).

The result is both more intuitive and more realistic – and accurate.

Time series factors

Many risk models use cross-sectional factors to estimate the covariance matrix.

Hybrid

R-Squared’s risk models are Hybrid in the sense that they combine the use of both specified factors and statistical factors.

Most risk models are based on a set of specified factors chosen by the vendor. Such models have the obvious advantage that the portfolio risk analysis is then expressed in terms of factors that can be clearly understood – even though they may not necessarily relate directly to the bets the manager had in mind when building the portfolio.

However, such models also have the potential drawback that the particular set of factors chosen may not capture all the common factor effects at work in the portfolio, thus leading to a mis-estimation of its risk.

Statistical factor models, on the other hand, are more likely to capture all the significant common factor effects at work in the portfolio, but suffer from the obvious disadvantage that their portfolio risk analysis is not easily interpretable in real-world terms that have economic or intuitive meaning.

The Hybrid approach captures the best of both worlds, in that it includes recognisable specified factors, which will usually capture most of the common factor effects at work in the portfolio, but also uses a small number of statistical factors to capture any other (possibly transient) common factor effects present in the stock returns data.

The Reference Day effect

Nearly all monthly models currently available rely on end-of-month data to construct their covariance matrices. For some time it has been recognised that this can introduce bias into the resulting risk estimates because end-of-month asset prices are not always representative of intra month prices. This error will be manifest risk forecasts that are inaccurate, sometimes dangerously so.

To overcome this bias, R-Squared developed a technique for sampling returns throughout the month to give a more valid estimation of covariances and hence risk and risk decomposition.