A once-in-a-lifetime change has taken place in the behaviour of share prices since late 1998. Individual share price volatility has risen to levels markedly higher than at any previous time bar a brief period in the mid-1970s. This change has had a direct impact on portfolio risk levels – nearly doubling typical active risk (‘tracking error’) over the past three years. Portfolio risk has risen because of a change in the market environment, not because of actions taken by managers.
This dramatic secular change has profound implications for portfolio managers, risk analysts and traders. Most risk models have performed poorly as a result of these changes with portfolios generating returns far outside predicted ranges. The models are under scrutiny from sceptical users and must evolve if they are to survive as practical management tools.
A useful way of measuring share price behaviour is ‘cross-sectional variability’ (CSV). This is a measure of the spread of share price changes for a specified period of time. To estimate CSV, we calculate the standard deviation of share price changes for a period of time (one week, for example) for a group of shares (the constituents of an index). It is convenient to annualise the statistic to give it a similar interpretation to standard volatility and risk statistics. This CSV measure is shown in figure 1 for the period from 1990 to date. This clearly shows that there has been a break with past patterns of share price behaviour since late 1998.
This is not a UK phenomenon. This shift in share price patterns can be observed across global equity markets, and US or Japanese equity markets show an almost identical pattern. Furthermore, the message from the CSV analysis is mirrored on the options markets over this period with a general risk in stock option implied volatilities (both in absolute terms and relative to index option implied volatilities). The changes highlighted in the chart cannot be explained by the movements of technology stocks or other specific sectors.
Why should this change matter? The answer is that if individual share prices exhibit more variability so, too, will portfolios. More important, the dispersion of ‘active’ returns on portfolios will increase. Tracking errors will rise. If tracking error (active risk) forecasts are based on the past behaviour of shares – as most models were over this period – then forecast risk levels were likely to have been too low.
It should be no surprise to see that portfolio-delivered tracking errors did rise over this period, wrong-footing many risk analysts and portfolio managers. The market environment changed and this affected portfolio be-haviour, even when no mat-erial changes had been made to the structure of portfolios.
One way of illustrating this is to look at the dispersion of equity portfolio returns.
Figure 2 shows the spread of returns reported for UK equity portfolios in WM’s All Funds sample between 1991 and 2000. The increase in the spread of portfolio returns matches the elevation in CSV observed in the same period (plotted with the white boxes and solid line). Apart from WM’s figures, there is plenty of anecdotal evidence to suggest that active risk forecasts dur-ing 1998 and 1999 were far too low in many asset management organisations.
Analysts have offered two explanations for the apparent increase in the individual volatility of shares – one fundamental and one technical. The fundamental explanation is broadly as follows. An individual share’s idiosyncratic risk (stock-specific risk) is normally related to company-specific factors. The argument here is that the rapid emergence of new technologies over the past few years (including the internet) has created risks to the value of all businesses.
Most of us agree that tech stocks are risky and are exposed to internet-related risk. This line of argument is different. It suggests that uncertainty has increased about the value of all business assets. Investors simply do not know what successful business models will emerge in the coming years. Some retailers will exploit the new technologies and their share prices will soar. Others will make the wrong choices and their share prices will suffer. On average, the new technologies will produce growth and profit, but it will be very unevenly spread. The winner takes all. If this truly is the new rapidly-changing, highly-competitive en-vironment in which businesses must operate, then it seems fair to expect company-specific risk to have risen. This line of argument does not appeal to everyone.
There is a second, more everyday, technical explanation on offer for the increase in CSV. Some investment banks report a change in recent years to the way large investors trade. In the US markets they have observed that the volume of limit orders relative to market orders has declined substantially. The argument goes that these market orders are more likely to cause large price moves which will induce high short-term volatility into share prices. We should observe that the increase in CSV is quite marked, even when measured using monthly share price changes. While trading activity might be expected to introduce volatility into price changes measured over short periods, it seems an unlikely culprit for a change measured using monthly data.
There is no escaping the fact that risk models have performed poorly in the recent past. Events that were assigned small probabilities by analysts have occurred with alarming frequency. Statistical models which were designed to mimic the past behaviour of shares failed to forecast the extreme price movements and portfolio returns experienced.
Does this mean that the statistical model is not useful? Our answer would be: “No, the model is still useful, but you must understand how vulnerable its assumptions are”. Model builders usually recognise that their models are only an approximate representation of the real world. However, the modeller usually likes to believe that the most important real-world factors are captured by the model. When modellers talk about ‘model risk’ they acknowledge the chance that the model may have missed something important or that the world can change in important ways.
The recent experience also suggests that statistical modellers may have got their priorities mixed up. Much effort has been expended in building expensive models of factor influences on share prices. If it turns out that individual factors and market-wide changes in share price behaviour are what really counts, it could suggest a simpler and cheaper approach with a different emphasis. Naturally, such a change would not be in the commercial interests of the providers of complex statistical risk models.
This does not mean that the quantitative analyst should pack up his box of tricks and go home. The changes in the market environment present some challenges, but they do not mean that the analyst cannot say anything useful about portfolio risk.
First, risk analysts and portfolio managers can monitor the sort of information presented above. Second, there is useful information on the prospective behaviour of shares from the options markets. An option price depends – in part – on the future uncertainty about a share price. As uncertainty increases so too does the option price and the so-called ‘implied volatility’. We can use this option implied volatility (measured across a basket of shares and adjusted for general market volatility) to say something about prospective risk conditions.
Although the message from this sort of analysis is pretty consistent with the analysis of CSV (because option investors and traders look at CSV) it does provide a real-time, forward-looking assessment of risk conditions. It is worth pointing out that sell-side quants and traders use this sort of information routinely. You might say that it is odd that the buy-side has not made more use of this information. You would be right and asset managers have paid a price (in poor risk forecasts) in recent years.
Quantitative analysts and model providers now have a long history of over-selling their wares. Really good financial models are simple and transparent. This means, first, that users can appreciate ‘it really is only a model’ and cannot capture all the complexity out there in the real world and, second, it is straight-forward to answer simple ‘what if?’ questions. Analysts who had asked: ‘what if stock-specific risk doubles?’ in spring 1999 would have been better equipped to explain the extreme returns generated by many portfolios that year. There is a need to think beyond the day-to-day risk management tool.
So, model providers deliver complex tools, but they have missed a once-in-a-lifetime change in risk dynamics. Buy-side users of risk models need to meet these challenges by tapping other sources of information to monitor market risk conditions and by developing a more eclectic banking-style approach to risk forecasting.
John Hibbert is founder and David Carruthers is senior analyst at consultants Barrie & Hibbert in Edinburgh