When large risks occur that have far-reaching consequences, for example when companies have to be rescued or even become insolvent, it is often tempting to say that a black swan has occurred.
In what follows, we first clarify what is meant by a black swan and why three-point distributions are so popular. Then we look at the assumptions hidden in them and at what happens in severe crises. Finally, we show how to model risks that are unbounded above, so that we are better prepared for large risk events.
What is a black swan?
A black swan means a risk whose magnitude could not have been foreseen.
Our thesis is that the impression of such an extraordinary situation often arises from a very elementary modelling effect, namely the use of three-point distributions. A three-point distribution describes a risk with three values, usually the minimum, the most likely value and the maximum. In doing so, it stipulates that the risk does not go beyond the maximum. If the risk does exceed this limit, the model makes it look like an event that nobody could have foreseen.
Why are three-point distributions so popular?
The natural home of the three-point distribution is the risk inventory in enterprise risk management, the company-wide management of risks. In the risk inventory, the company’s risks are regularly recorded and reassessed. This situation has several typical characteristics:
- The number of risks that have to be recorded and reassessed is relatively high.
- The risk assessment is carried out manually. You really have to pick up every single risk, check what it looks like today and assess it again.
- This requires a great many people. The size of the company alone means that nobody can do it alone, because nobody knows all areas well enough to be able to.
- Most risk owners, that is the people who are responsible for a risk and assess it, are not familiar with assessment methods. Risk management is not part of their day-to-day business. They assess their risks only on the side, once a quarter or once every six months.
- The risk managers therefore have to support them, and this support should of course cause the risk managers as little work as possible.
Out of this context comes the call to keep the risk assessment as simple as possible: “Keep it simple”. A decisive precondition is that the assessment method is intuitively understandable for the risk owners and easy for the risk managers to explain. Then the assessment is quick and causes little friction.
The answer that practice gives to this is precisely the three-point distribution. In many companies, a three-point distribution is, as it were, the basic form of a risk. People therefore often assume that all risks have this basic form and consider the assessment done as soon as the parameters of the three-point distribution have been set.
Which three-point distributions are there?
One example is the triangular distribution in Figure 1. It has three distinctive corner points: the minimum of the distribution, the most likely value and the maximum. Once these three values have been given, the impact of the risk has been described in its magnitude and in its shape. The horizontal axis shows the impact, here from 0 to 350,000. The vertical axis shows the density: the higher the orange area above a value, the more likely impacts in that region are. The area begins at the minimum on the left edge, rises to the peak at the most likely value and falls back to zero at the maximum on the right edge.
Figure 1: The triangular distribution as the basic form of a risk. The minimum, the most likely value and the maximum define it completely (German slide).
More on this: Our article “When do I use a triangular distribution in a Monte Carlo simulation?” shows how the triangular distribution is defined, when it fits and why its maximum can be an illusion.
The triangular distribution is not the only three-point distribution used in practice. The PERT distribution is also very popular. It has exactly the same parameters as the triangular distribution, namely the minimum, the most likely value and the maximum. It has a smooth shape, however, and on the long side it falls off somewhat faster than the triangular distribution.
A third example is the multinomial distribution. With it, you can specify any number of values, which are then taken as genuine individual points. To make it a three-point distribution, we specify three values here. In the simulation and in the risk aggregation, that is when the individual risks are combined into an overall risk, one of these three values always occurs. All the values in between do not exist. We can already see here that the essential feature of these distributions is not that they have three points, but that they have end points.
The fourth example is the uniform distribution between two points, a minimum and a maximum. All values in between are equally likely. Each of them is therefore a most likely value.
Figure 2 shows the four distributions side by side: top left the triangular distribution from Figure 1, top right the PERT distribution, bottom left the multinomial distribution with its three individual values as narrow bars, and bottom right the uniform distribution. All four are drawn on the same axis from 0 to 350,000. The PERT distribution falls almost to zero well before the maximum, whereas the triangular distribution runs out in a straight line all the way to the maximum. What all four have in common is that there are no values beyond the maximum.
Figure 2: Four distributions with fixed limits. At the top the triangular and the PERT distribution, at the bottom the multinomial distribution with three values and the uniform distribution (German slide).
More on this: Our article “When do I use a PERT distribution in a Monte Carlo simulation?” shows where the PERT distribution comes from, how it differs from the triangular distribution and why it is not suitable for quantities that are open above.
What assumptions are hidden in a three-point distribution?
The problem with three-point distributions is that they make assumptions that are not readily visible from the outside. The title of the slide in Figure 2 names them: the risk is bounded, and we know the maximum.
The first assumption is that the risk is bounded in principle, that is, that a maximum exists at all that one could state. The second assumption is that we actually know this maximum. It does not help us to say that there may well be a maximum but we do not know its value. These four distributions then get us nowhere, because for them we absolutely have to know the value of the maximum.
On the one hand, these distributions have an advantage: their parameters are all easy to explain. More or less everyone knows what a minimum is, and the same goes for a maximum. The most likely value is not quite so simple, because it is not something people know from everyday life. In practice, it is often filled with what one expects in principle.
The problem, however, is that the minimum and the maximum are not as easy to name as we believe, because we often do not readily know the maximum of risks. And if we are honest, we often cannot even say for certain whether a risk is really strictly bounded. The slide in Figure 2 sums this up in two bullet points: easy to explain, but not easy to name.
This assumption therefore runs into crisis precisely when the risk exceeds the assumed limits.
What happens in severe crises?
Let us take an example from the economic situation in spring 2023, when the video was made. Figure 3 shows the interest rates for the euro area and for the United States from January 2021 into spring 2023: the euro rate in orange and the US dollar rate in blue.
Figure 3: The USD rate (blue) and the EUR rate (orange) in percent. After a long phase of low interest rates, both rise sharply from the end of 2021 (German slide).
In both worlds, the period began with a long phase of low interest rates. In the dollar world, rates were close to zero, and in the euro world they were even slightly below, at about −0.7%. At the end of 2021, this situation suddenly changed, and rates rose sharply in both worlds: within a little over a year by 500 basis points in the dollar area and by just under 400 basis points in the euro area. A basis point is one hundredth of a percentage point, so 100 basis points make one percentage point. These were sensationally high rises in interest rates.
The interest rate risk models themselves were not prepared for this rise. They capture the small fluctuations that we see as jagged edges on the curves in Figure 3 and project them into the future. In a certain sense, however, the rise was not random at all, because the shifts in the yield curve were due to central bank policy. The problem was that these shifts were so large that they had often not even been assumed in the maximum realistic scenario.
Banks in crisis
In the course of the rise in interest rates, there were several casualties. The first was Silicon Valley Bank. It had invested a great deal in government bonds, which lost massive value after interest rates had risen so sharply. After that, Credit Suisse, one of the largest European banks, Signature Bank in the US and finally First Republic Bank also got into serious trouble. In all of these cases, a bank run came on top: customers began to withdraw their money as quickly as possible.
All of these institutions were very large, each with a balance sheet total of over a hundred billion, and highly renowned. Nevertheless, Silicon Valley Bank had not assumed that a change in interest rates could reach this magnitude, and it was not hedged against this scenario. In other words, it had assumed that the risk has a maximum and that it knows this maximum. This assumption turned out to be wrong, and the bank stumbled over it.
Banks are not the only ones affected
Banks are of course not the only ones that can be affected. The same applies to all other indebted organisations, for example companies that have invested relatively heavily and are therefore highly indebted. When they gradually refinance their loans, they have to pay considerably more interest. They have to earn this interest from the returns of the projects they financed with the money. Creditreform had predicted long ago that many companies would not manage this and might therefore slide into insolvency. Private individuals who are highly indebted are hit in the same way, and so are states. The effect is the same in all cases.
So we see: when the risk exceeds the assumed limits, a black swan quickly arises, so to speak on paper.
Further risks without a known maximum
At the time, interest rate risks were not the only risk that had swung far and reached magnitudes that had been unimaginable in the past. The same was true of commodity prices, for example for rare earths and metals, and of energy prices for natural gas, mineral oil and electricity. Security of supply was back on the agenda: would Germany be able to supply itself with gas? How would the stability of the power grids develop in the future?
What magnitude can we assume for inflation? Is there a maximum for it that we can rely on and that we might even need to know? In the euro area alone, the inflation rates of the individual countries during this period were between 5 and 22%. That is an extremely wide range. Where was the maximum, and had the end of the line already been reached?
On top of this came renewed geopolitical conflicts. Is there a maximum for what can happen between China and the US, between China and Taiwan, between China and its other neighbours, or between Russia, Ukraine and other countries in the region? How do the activities of activists affect some companies, and how does ever more comprehensive regulation? Cyber risks were in any case among the top places on the risk map. With them, too, it was not clear how they would develop and what magnitudes they would still reach and exceed.
The assumption that a risk always has a maximum impact and that we also know this maximum may therefore stand on feet of clay. We therefore have to think about alternatives, because if we no longer want to assume a maximum, the three-point distribution as a workhorse will no longer get us anywhere.
How do you model risks that are unbounded above?
The alternative assumption we have to make is this: the risk can be unbounded. So how can we represent a risk for which we give corner points but no maximum?
Three quantiles instead of three points
A first example is the Expert distribution, whose input dialog is shown in Figure 4. It, too, aims to be very easy to explain and to parameterise. First, we define what it means for something to be fairly certain to happen. The field “Ziemlich sicher” (fairly certain) contains 0.1. This is the probability with which we are allowed to be wrong. Fairly certain here therefore means 90% certain.
Figure 4: The Expert distribution with the minimum 0, the median 100,000 and two values that are fairly certain to be exceeded and not exceeded respectively. The field “Maximum” stays empty, so the distribution is open above (German interface).
Then we specify three quantiles of the distribution. A quantile is a value below which the impact stays with a certain probability. The first point is the value where we are fairly certain that the impact is larger, in this case 50,000. The second point is the median, the middle of the distribution, here 100,000. It takes the place of the most likely value, so to speak. That is a good thing, because the median is clearly defined: 50% of the values lie below it and 50% above. The third point is the value where we are fairly certain that the impact is smaller, here 200,000. With the definition we made first, this means that 50,000 is undershot with a probability of only 10%, and 200,000 is exceeded with a probability of only 10%.
We do not have to specify a minimum or a maximum for this distribution, as they are optional. This gives us a great deal of flexibility. If there is a minimum, for example zero, which exists for many risk impacts, we can certainly specify it. In the example, the field “Minimum” contains 0. We can leave the maximum open. If we do not know which inflation rate, which gas price or which electricity price is the maximum, we cannot set a value here anyway. Conversely, if we are in a situation where we do know the maximum, we can use this parameter too. The standard case of the three-point distribution, in which we know the minimum and the maximum, is therefore included here. But we are not tied to always having to specify both.
On the right of the dialog, we see the shape of the distribution. The orange area peaks between 0 and 150,000 and runs out to the right in a long, flat tail. So impacts can still occur above 200,000, and some of them are substantially higher than 200,000.
More on this: Our article “Quantifying expert estimates in a Monte Carlo simulation: the Expert distribution” shows how the Expert distribution is defined and how it represents long tails and open limits.
As many points as you like
Figure 5 shows a second variant, the Poly distribution bounded on the left. Bounded on the left means that it has a lower limit. Here we specify various values, each of which we expect with a certain degree of certainty. The column “Verlust” (loss) contains the value, and the column “Grad der Sicherheit” (degree of certainty) contains the probability with which this value is not exceeded.
Figure 5: The Poly distribution bounded on the left. Seven points define the degree of certainty with which a loss is not exceeded. There is no maximum (German interface).
The minimum is 0, that is the smallest value the distribution takes, just as with the Expert distribution. We are 10% certain that 50,000 will not be exceeded, 50% certain that 100,000 will not be exceeded and 90% certain that 200,000 will not be exceeded. These are the same values as in Figure 4. But now we can specify further quantiles and thereby define how the upper part of the distribution is structured: 250,000 is not exceeded with 95% certainty, 400,000 with 99% and 550,000 with 99.9%.
So here we are not limited to specifying only three corner points. We can define as many points of the distribution as we like and thus draw the distribution precisely, so to speak. As with the Expert distribution, we could add a maximum here too if we wanted to. It would correspond to a degree of certainty of 100%. But we can also leave the distribution open above, as in Figure 5.
What strategy should risk assessment follow?
Figure 6 summarises the strategy that risk assessment should follow. When you have a risk in hand and are assessing it, it is a good idea to check first what characteristics the risk has. In particular, you should ask yourself two questions: is there a worst case, or is the risk open above? And is this worst case known?
Figure 6: Three rules for risk assessment. Check the characteristics of the risk, use no assumptions that you already know to be false, and represent the essential characteristics of the risk correctly (German slide).
If we answer both questions with yes, that is, if there is a worst case and we also know its value, then a classic three-point distribution may not be so wrong for this risk after all. In that case we may simply carry on as before.
If, on the other hand, we conclude that this worst case may not exist, or that we do not know it even if it did exist, then we should be careful. Then the rule applies that we should not make assumptions that we already know to be false. So if we know that there is no maximum, we should not assume a maximum either. Instead, we should try to represent the essential characteristics of the risk correctly, for example with a distribution that is open above, as in Figures 4 and 5.
This way we avoid a large risk event looking like a black swan merely because our model assumed a limit that does not exist in that form in reality.