Plotting the frequency of earthquakes higher than a given magnitude on a logarithmic scale gives a straightish line that suggests we might expect a 9.2 earthquake every 100 years or so somewhere in the world and a 9.3 or 9.4 every 200 years or so (the Tohoku earthquake which led to the Fukushima disaster was 9.0). Such a distribution is known as a power-law distribution, which gives more room for action at the extreme ends than the more familiar bell-shaped normal distribution, which gives much lower probabilities for extreme events.
Similarly, plotting the annual frequency of one day falls in the FTSE All Share index higher than a given percentage on a logarithmic scale also (as you can see below) gives a straightish line, indicating that equity movements may also follow a power-law distribution, rather than the normal distribution (or log normal, where the logarithms are assumed to have a normal distribution) they are often modelled with.
However the similarity ends there, because of course earthquakes normally do most of their damage in one place and on the one day, rather than in the subsequent aftershocks (although there have been exceptions to this: in The Signal and the Noise, Nate Silver cites a series of earthquakes on the Missouri-Tennessee border between December 1811 and February 1812 of magnitude 8.2, 8.2, 8.1 and 8.3 respectively). On the other hand, large equity market falls often form part of a sustained trend (eg the FTSE All Share lost 49% of its value between 11 June 2007 and 2 March 2009) with regional if not global impacts, which is why insurers and other financial institutions which regularly carry out stress testing on their financial positions tend to concern themselves with longer term falls in markets, often focusing on annual movements.
How you measure it obviously depends on the data you have. My dataset on earthquakes spans nearly 50 years, whereas my dataset for one day equity falls only starts on 31 December 1984, which was the earliest date from which I could easily get daily closing prices. However, as the Institute of Actuaries’ Benchmarking Stochastic Models Working Party report on Modelling Extreme Market Events pointed out in 2008, the worst one-year stock market loss in UK recorded history was from the end of November 1973 to the end of November 1974, when the UK market (measured on a total return basis) fell by 54%. So, if you were using 50 years of one year falls rather than 28.5 years of one day falls, a fall of 54% then became a 1 in 50 year event, but it would become a 1 in 1,000 year event if you had the whole millennium of data.
On the other hand, if your dataset is 38 years or less (like mine) it doesn’t include a 54% annual fall at all. Does this mean that you should try and get the largest dataset you can when deciding on where your risks are? After all, Big Data is what you need. The more data you base your assumptions on the better, right?
Well not necessarily. As we can already see from the November 1973 example, a lot of data where nothing very much happens may swamp the data from the important moments in a dataset. For instance, if I exclude the 12 biggest one day movements (positive and negative) from my 28.5 year dataset, I get a FTSE All Share closing price on the 18 July 2013 of 4,494 rather than 3,513, ie 28% higher.
Also, using more data only makes sense if that data is all describing the same thing. But what if the market has fundamentally changed in the last 5 years? What if the market is changing all the time and no two time periods are really comparable? If you believe this you should probably only use the most recent data, because the annual frequency of one day falls of all percentages appears to be on the rise. For one day falls of at least 2%, the annual frequency from the last 5 years is over twice that for the whole 28.5 year dataset (see graph above). For one day falls of at least 5%, the last 5 years have three times the annual frequency of the whole dataset. The number of instances of one day falls over 5.3% drop off sharply so it becomes more difficult to draw comparisons at the extreme end, but the slope of the 5 year data does appear to be significantly less steep than for the other datasets, ie expected frequencies of one day falls at the higher levels would also be considerably higher based on the most recent data.
Do the last 5 years represent a permanent change to markets or are they an anomaly? There are continual changes to the ways markets operate which might suggest that the markets we have now may be different in some fundamental way. One such change is the growth of the use of models that take an average return figure and an assumption about volatility and from there construct a whole probability distribution (disturbingly frequently the normal or log normal distribution) of returns to guide decisions. Use of these models has led to much more confidence in predictions than in past times (after all, the print outs from these models don’t look like the fingers in the air they actually are) and much riskier behaviour as a result (particularly, as Pablo Triana shows in his book Lecturing Birds on Flying, when traders are not using the models institutional investors assume they are in determining asset prices). Riskier behaviour with respect to how much capital to set aside and how much can be safely borrowed for instance, all due to too much confidence in our models and the Big Data they work off.
Because that is what has really changed. Ultimately markets are just places where we human beings buy and sell things, and we probably haven’t evolved all that much since the first piece of flint or obsidian was traded in the stone age. But our misplaced confidence in our ability to model and predict the behaviour of markets is very much a modern phenomenon.
Just turning the handle on your Big Data will not tell you how big the risks you know about are. And of course it will tell you nothing at all about the risks you don’t yet know about. So venture carefully in the financial landscape. A lot of that map you have in front of you is make-believe.