Where risk models are weak
The concept and practice of risk management have become increasingly important for the investment management industry in recent years. In order to meet the soaring demand for risk figures to be produced as quickly as possible, fund managers have increasingly turned to risk evaluation models which are centered on one common measure – ex ante tracking error.
The simplicity of tracking error is probably the principal reason for its popularity. It encapsulates risk in a single number and is relatively easy to calculate using one of a number of commercial risk models. However, recent market conditions have highlighted various weaknesses that render it unsuitable as the sole measure of portfolio risk.
Tracking error essentially relies on interpreting the historical behaviour of various factors in order to forecast the future. More-and-more complex financial modeling techniques are employed to forecast factors such as market volatility. Typically, a long observation period will be used - usually five years or beyond - although increased weighting is often given to more recent data. This attempt to capture more recent trends is only partially successful since in essence, risk models implicitly assume that volatility is relatively constant.
During 1997 and 1998 in particular, this fundamental assumption led to a significant underestimation of risk, reflecting the sudden increase in the levels of volatility experienced in all global equity markets during the Asian crisis, and lead-up to the Long Term Capital Management crisis in the autumn of 1998. This followed a five-year period of low and falling volatility in equity markets, and risk models subsequently underestimated tracking errors. Interestingly, volatility levels (at the overall market level) have been reasonably stable since then, and this has not been an issue in the estimation of tracking errors in developed markets since 1999.
When one considers that even in ‘normal’ market conditions, predicted tracking errors can underestimate actual tracking errors by between 30% and 50%, the level of underestimation may become unacceptably high in times of extreme market conditions. Tracking error should therefore perhaps be viewed more as a snapshot of the risk profile of a portfolio at a particular point in time, based on historic market conditions rather than as a measure that predicts the performance of a portfolio against its benchmark.
While the observed performance of many portfolios during periods of intense volatility has been either significantly better or worse than that forecast by the standard risk models, it is not simply due to the impact of volatility. Market momentum or auto-correlation is another significant factor, and its impact on the accuracy of the standard risk model has been even greater than that of volatility in the last two years. Tracking error estimates are provided by the study of a series of monthly performance data, which is then converted into an annualised risk statistic.
The underlying assumption is that performance from one month to the next is uncorrelated, ie that one month’s performance is independent and distinct from the previous month’s returns. Not only does this contradict conventional wisdom, which states that certain styles – growth, or value, for example – can dominate markets for a period of time, but also that various market sectors or individual stocks can continue to move in the same direction for a prolonged period. This was illustrated very clearly in 1999’s final quarter when telecoms, media and technology sectors (TMT) rallied sharply, gathering pace as investors rushed to jump on the bandwagon. This was a classic example of auto-correlation, and led to significant inaccuracy in forecast tracking error.
Along with volatility and momentum, stock specific risk is another potential area of weakness for standard risk models. Stock specific risk is independent of industry and style, but in absolute terms, is the largest component of the risk of any individual stock. Diversification helps to reduce stock specific risk, but cannot eliminate it fully, and an active portfolio will generally contain a large amount of this type of risk. Markets have behaved unusually in the last two years in that volatility at the overall market level has been relatively stable, while it has risen sharply at the stock level. The recent rise in individual stock volatility, and in addition, the dispersion of individual stock returns since the rise of the TMT factor in late 1999, have now propelled this factor to centre stage.
A risk model that fails to capture the size and spread of stock returns in a dynamic way will undoubtedly provide inaccurate tracking error estimates. To a certain extent, commercial risk models are dynamic in that they respond to increases in observed levels of specific risk over a relatively short time horizon by focusing on both the average level of specific risk within any given market and the risk peculiar to individual stocks.
However, the method of calculating stock specific risk estimates effectively caps the level of specific risk borne by any individual stock by using regression techniques and then standardising estimates to a three standard deviation limit. Observed extreme stock events will therefore not be fully replicated in the model as they will simply be allocated the highest permitted level of risk. While this may be acceptable in ‘normal’ market conditions, in periods of high volatility such as we have recently experienced, the potential for underestimating tracking errors increases considerably.
One other factor which can also have an adverse impact on the accuracy of tracking error estimates is that standard risk models are based on the analysis of static portfolios and do not take into account the trading and turnover activity that form a normal part of active fund management. As a result, for instance, two funds that have similar tracking errors, can produce quite different returns relative to their benchmarks, simply because of the difference in their levels of turnover. Similarly, even in the absence of significant stock turnover, outperformance of one sector over another will tilt the balance of both the fund and the benchmark and may alter the tracking error.
Looking ahead, investigations could certainly be carried out with the aim of improving the accuracy of tracking error – for example by using adjustment factors to load ex ante tracking errors to reflect any increase in the volatility of markets or auto-correlation effects which are not reflected in the risk models used. But in practice, there is no one measure that entirely encapsulates a portfolio’s risk profile, and a range of measures need to be used in order for clients, consultants and fund managers to have a better understanding of risk.
Relatively simple summary statistics can help provide insight. We also examine measures such as the sum of absolute sector bets, average price/earnings ratio, average price to book ratio, average forecast earnings per share growth and beta. However, we believe no one measure is capable of accurately encapsulating a portfolio’s risk exposure, since all have specific weaknesses. The key to risk management is therefore the careful analysis and control of each source of risk.
In practice, this entails slicing the portfolio in different ways so as to understand the relative positions that might influence the portfolio’s future performance. The fund manager should ‘cut’ the portfolio at country level, currency level, style factor level, industry level, market capitalisation split, and stock level, and analyse the portfolio biases at each level relative to the portfolio’s benchmark. The more thorough the portfolio analysis, the better the fund manager’s understanding of overall sources of risk, and the better it can be conveyed to clients and consultants.
A range of summary risk statistics and a detailed analysis of sources of risk will add considerable insight to the portfolio construction process and house style, enabling a better understanding of portfolio composition, the sources of deviation away from the benchmark and how this compares with recent history. In this way, more valuable information is available to clients, fund managers and consultants alike, and the weaknesses of individual risk models are no longer of particular significance.
Gary Smith is head of the global portfolio team at Gartmore Investment Management in London