You are on page 1of 8

On the meaning of a black swan in a risk context

Terje Aven

University of Stavanger, 4036 Stavanger, Norway
a r t i c l e i n f o
Article history:
Received 5 July 2012
Received in revised form 21 January 2013
Accepted 22 January 2013
Available online 24 February 2013
Keywords:
Black swan
Risk
Probability
Uncertainty
Knowledge
a b s t r a c t
In recent years there has been much focus on the so-called black swans in relation to risk management
and decision making under uncertainty. A key issue has been the ability of risk assessment and probabil-
ity theory to capture the black swans. In this paper we carry out an in-depth analysis of what a black
swan means in relation to risk, uncertainty and probability: is a black swan just an extreme event with
a very low probability or is it a more surprising event in some sense, for example an unknown unknown?
We question how the black swans are linked to the risk concep t, to expected values and probabilities, and
to the common distinction between aleatory uncertainties and epistemic uncertainties. The main aim of
this paper is to contribute to a clarification of the issue in order to strengthen the foundations of the
meaning and characterisation of risk, and in this way provide a basis for improved risk management.
The paper concludes that the black swan concept should be associated with a surprising extreme event
relative to the present knowledge.
Ó 2013 Elsevier Ltd. All rights reserved.
1. Introduction
In recent years I have heard numerous speeches and read a large
number of papers which refer to the black swan logic in a risk con-
text. The metaphor is intuitively appealing and has been very pop-
ular to illustrate the idea of surprisin g events and outcomes. The
black swan concept was first introduced by the Latin poet Juvenal,
who wrote ‘‘rara avis in terris nigroque simillima cygno’’ (a rare
bird upon earth, and exceedingly like a black swan), but, accordin g
to Hammond (2009), that was imaginative irony. Juvenal’s phrase
was a common expression in 16th century London, as a statement
of something impossible. Up to that point in time, all observed
swans in the Old World had been white. Taleb (2007), p. xvii,
writes:
Before the discovery of Australia people in the Old World were
convinced that all swans were white, an unassailable belief as it
seemed completely confirmed by empirical evidence.
But then in 1697 a Dutch expedition to Western Australia , led
by Willem de Vlarningh, discovered black swans on the Swan Riv-
er, and the concept of black swans developed to mean not only
something extremely rare (a rarity), but also that a perceived
impossibility might later be disprove n: a logical fallacy, meaning
that if one does not know about something, it is therefore impos-
sible. Taleb (2007) comments that in the 19th century John Stuart
Mill used the black swan logical fallacy as a new term to identify
falsification. John Stuart Mill wrote: ‘‘No amount of observati ons
of white swans can allow the inference that all swans are white,
but the observation of a single black swan is sufficient to refute
that conclusion.’’ It became a classic example in elementary philos-
ophy (Hammond , 2009 ).
In 2007, Nassim Nicholas Taleb further defined and popularised
the concept of black swan events in his book, The Black Swan (Ta-
leb, 2007 ) (a second edition was issued in 2010 with a new section
which among other things discusses various aspects of the proba-
bility concept). Taleb refers to a black swan as an event with the
following three attributes. Firstly, it is an outlier, as it lies outside
the realm of regular expectations , because nothing in the past
can convincingly point to its possibility. Secondly, it carries an ex-
treme impact. Thirdly, in spite of its outlier status, human nature
makes us concoct explanation s for its occurrence after the fact,
making it explainabl e and predictabl e.
Taleb’s (2007, 2010) book has inspired many authors; see for
example the many references in Taleb (2011). However , some
scholars are sceptical of Taleb’s work. Professor Dennis Lindley,
one of the strongest advocato rs of the Bayesian approach to prob-
ability, statistics and decision making, has made his view very clear
in a review of Taleb’s book (Lindley, 2008 ): Taleb talks nonsense.
Lindley lampoons Taleb’s distinction between the lands of Medioc-
ristan and Extremistan, the former capturing the placid random-
ness as in tosses of a coin, and the latter covering the dramatic
randomn ess that provides the unexpected and extreme outcomes.
Lindley provides an example of a sequence of independent trials
with a constant unknown chance of success – clearly an example
of Mediocristan. Each trial is to be understood as a swan and suc-
cess a white swan. Using simple probability calculus, Lindley
shows that a black swan is almost certain to arise if you are to
0925-7535/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ssci.2013.01.016

Tel.: +47 51831000/2267; fax: +47 51831750.
E-mail address: terje.aven@uis.no
Safety Science 57 (2013) 44–51
Contents lists available at SciVerse ScienceDi rect
Safety Scien ce
j our nal homepage: www. el sevi er . com/ l ocat e/ ssci
see a lot of swans, although the probability that the next swan ob-
served is white, is nearly one. Lindley cannot be misunderstood :
‘‘Sorry, Taleb, but the calculus of probability is adequate for all kind
of uncertainty and randomness’’.
The purpose of the present paper is to provide a thorough anal-
ysis of this issue: the concept of the black swan in relation to risk.
What is the meaning of this term in a professional /scientific set-
ting? I question to what extent the ideas of Taleb, and in particular
the distinction between Mediocristan and Extremistan, can be gi-
ven a proper scientific justification in view of existing risk theories
and perspectives . Clearly, if Taleb has made some important points,
Lindley cannot be right. More specifically, the paper studies several
interpretations of a black swan, starting from these four:
1. A surprising extreme event relative to the expected occurrence
rate (extreme event in the sense that the consequences are
large/severe , this understand ing also applies to the interpreta-
tions 2 and 3 below).
2. An extreme event with a very low probability.
3. A surprising, extreme event in situations with large uncertainties.
4. An unknown unknown.
The discussion will be linked to a set of issues, including:
(a) The common distinctio n between aleatory uncertainties and
epistemi c uncertainti es. Does a black swan mean that
chances (frequentist probabili ties) cannot be meaningfully
defined for this event? A frequentist probability exists as a
proportion of infinite or very large populations of similar
units to those considered. It represents the aleatory
uncertainties .
(b) The ability of risk assessme nt to identify the black swans.
(c) The problem of establishing an accurate prediction model
for the black swan.
(d) The situation being characterised by large consequences and
high uncertainties (‘‘post normal sciences’’). Is the occur-
rence of black swans linked to or restricted to such
situations?
(e) The situation being characterise d by scientific uncertainties
as in applying the precautionar y principle. Again, is the
occurrence of black swans linked to or restricted to such
situations?
The remaining part of the paper is organised as follows. In Sec-
tion 2 we discuss the meaning of a black swan – addressing the
four interpretations 1–4 mentioned above. The discussion is based
on some fundamenta ls concerning the concepts of risk, probability
and uncertainty, which are presente d in Appendix A. To be able to
clarify the meaning of the black swans, it is essential to be precise
on these concepts. As we know there are different perspecti ves and
definitions of these terms, and in our analysis we need to distin-
guish between these to be able to carry out a thorough argumenta-
tion. The closing Section 3 provides some final remarks and
conclusions .
2. Discussion of what a black swan means
In the following we will discuss the four interpretations 1–4
introduced in Section 1.
2.1. Is a black swan a surprising extreme event relative to the expected
occurrence rate?
Let N(t) denote the number of times the event under consider-
ation (called A) occurs in the period [0, t], and assume that the
expected number of events per unit of time, E[N(t)]/t, converge s
to k as t goes to infinity (the expectation here is with respect to a
frequent ist probability distribution, see Appendix A). We refer to
k as the expected occurrence rate of the event A. Suppose k = 1/
100 (year
÷1
), i.e. the event A is expected to occur once every
100 years.
Now, is the occurrence of A to be considered a black swan? It is
clearly a rare event, but the probability that it occurs in a period of
10 years could be relatively high. Assuming that the occurrence
process N(t) is a Poisson process, we know that the probability of
at least one event during 10 years is 1 ÷ exp{÷k10} ~ 0.10, and
the occurrence of the event can hardly be said to be surprising.
If, however , the occurrence rate k is equal to say 1/10,000, an
occurrence of the event during the next 10 years is rather surpris-
ing as the probabili ty is about 0.1%. However, say that we are fac-
ing not only one such event occurrence process, but many similar
processes , for example 200. Then the occurrence of at least one
event is about 20%. Hence the occurrence of one event in this per-
iod is not particularly surprising. Focusing on a specific type of
event A, the occurrence may be surprisin g but not when consider-
ing a large number of such processes.
In a society we are facing a number of different types of extreme
events. Let us assume for the moment that we know what these
types are and what their occurrence rates are. Then we are in a sit-
uation as described above. Considered in isolation, one type of ex-
treme events may be considered surprising but not if we open up
for all types of events.
Of course in practice, the situation is not as idealised as this; we
may not know all types of events and how often they occur. These
situation s will be discussed in the coming subsectio ns. It is too
early to conclude on the question of whether a surprisin g extreme
event relative to the expected occurrence rate should be consid-
ered a black swan.
2.2. Is a black swan an extreme event with very low probability?
Let us return to the case presented by Lindley (2008) mentioned
in the introduction section. In this example we consider a sequence
of independen t trials with a constant unknown chance of success.
Lindley shows that a black swan (failure of trial) is almost certain
to arise if you are to see a lot of swans, although the probabili ty
that the next swan observed is white (success of trial) is nearly
one. This example is similar to the one studied above for the occur-
rence rate. When we focus on the occurrence rate of the first black
swan we get a very low number, whereas if we consider the occur-
rence rate when studying a large set of swans the probability of
occurrence of a black swan becomes large. But there are some dif-
ferences between these two examples , and these are important. To
reveal these, we need to dive deep into the assumptions of Lind-
ley’s example.
Lindley assumes that we are facing a sequence of independent
trials with a constant unknown chance of success, and to obtain
his probabilitie s he assumes a prior probability distribution over
this chance, namely a uniform distribut ion over the interval
[0, 1]. This means that Lindley tacitly has assumed that there is a
zero probability that all swans are white – there is a fraction of
swans out there that are black (non-white). From this point on,
his analysis cannot change this assumption. Of course then, the
probabili ty calculus will show that when considering a sufficient
number of swans, some black ones will be revealed; see the Appen-
dix B. By the assumpti ons made, the analyst has removed the main
uncertainty aspect of the analysis. In real life we cannot of course
exclude the possibility that all swans are white. The uncertainty
about all swans being white is a key issue here, and Lindley has
conceale d it in the assumptions . This is the problem raised by
many authors, including Taleb (2007, 2010) and Aven et al.
T. Aven / Safety Science 57 (2013) 44–51 45
(2011): the probabili ty-based approach to treating the risk and
uncertainties is based on assumptions that could hide critical
assumptions and therefore provide a misleading description of
the possible occurrence of future events. Let us reconsider Lindley’s
example to allow for a positive probabili ty that all swans are white.
Let us assume that there are only two possibilities, either the
fraction p of white swans are 100% or 99%. Hence p is either 1 or
0.99. Suppose the analyst assigns prior probabili ties to these values
as 0.2 and 0.8, respectivel y. Now suppose that the analyst has ob-
served n swans and they are all white, what then are his posterior
probability for the next m swans to be all white? (n and m large
numbers). Using Bayes’ formula in the standard way, we find that
this probability is close to one, i.e. the probabili ty of a black swan
occurring is very small, in contrast to what Lindley computed in his
analysis. See the Appendix B for the details.
This example shows the importance of the assumptions made
for the probabilistic analysis. Depending on the assumpti ons made,
we get completely different conclusions about the probabili ty of a
black swan occurring.
Lindley’s example also fails to reflect the essence of the black
swan issue in another way. In real life the definition of a proba-
bility model and chances cannot always be justified, see Appendix
A. Lindley’s set-up is the common framework used in both tradi-
tional statistics and Bayesian analysis. Statisticians and others of-
ten simply presume the existence of this framework and the
elements of surprise that Taleb and others are concerned about
fall outside the scope of the analyses. This is the key problem
of the probability-ba sed approach to risk analysis, and a possible
interpretation of Taleb’s work is the critique of the lack of will
and interest for seeing beyond this framewor k among statisticians
and others when analysing risk. I share this concern; see e.g. Aven
(2011a,b). Let us look into an example to clarify what the issue is
all about.
Let us consider the event A that an individual person carries out
a serious terrorist attack in a country during the next year. An
example of such an attack is the killing in Norway on 22 July
2011, when a man placed a car-bomb outside the government of-
fice and massacred a number of people on the island of Utøya out-
side Oslo. For this type of event, it has no meaning to talk about a
frequentist probability or chance. We cannot repeat situations like
this under similar conditions. We may, however , use a knowledge-
based (subjective, judgemental) probabili ty to express our uncer-
tainty and degree of belief about the occurrence of this type of
event. Say that a group of experts assigns a probability of A equal
to 0.001, given the background knowled ge K at a specific point in
time, i.e. P(A|K) = 0.1%. The group considers it quite unlikely that
such an event will occur given the informat ion it has available.
Then an event A occurs. Was it a black swan? The assigned
probability was very small, and the outcome then somewhat
surprising.
As discussed in Section 2.1, we must be careful in judging an
event as surprising based on an isolated analysis; if we extend
the types of events considered, the assigned probability for at least
one event could be relatively large. Furthermore, the probability is
a knowledge- based probability, and this probabili ty could of
course lead to poor predictions of the actual number of events
occurring. The probabilitie s could be based on assumptions that
turn out to be wrong, and/or on little relevant information and
knowledge. In the terrorist attack example, the relevant police
security services could have based their judgements on many
assumptions concerning the motivation and capability of persons
with extreme views to perform such an attack. We see that the is-
sue about the event being a black swan or not based on probability
assignment s leads us into a discussion about the background
knowledge that the probabilitie s are conditional on, the topic of
the next section.
2.3. Is a black swan a surprising, extreme event in situations with large
uncertainti es?
There is a discussion in the literature concerning the suitabilit y
of subjectiv e probabili ties to reflect epistemic uncertainties (see
e.g. Aven, 2010b; Dubois, 2010; North, 2010 ). The problem with
these probabilitie s is that the knowledge that they are based on
is not reflected in the assigned numbers. In two situations you
may assign the same probability, say 0.2, but the knowledge basis
could be completely different. Let us focus on the situation charac-
terised by poor knowled ge and, to be concrete, think about the ter-
rorist example again. Here the police security services may have
little information indicating such an event taking place, and a pos-
sible occurrence could be seen as surprising relative to this
informat ion.
Situations of poor knowledge and extreme consequences have
been analysed by many authors. One of the earliest works goes
back to Funtowicz and Ravetz (1985), who presented a model for
classifying problem- solving strategies into applied sciences, pro-
fessional consultancy and post-normal sciences. The model is illus-
trated in a diagram based on the two axes: (i) decision stakes – the
value dimensio n (costs, benefits), and (ii) the system uncertainti es
– the knowledge dimension. Large uncertainties and high decision
stakes characterise the post-norm al sciences. An example of this
category is climate change. Surprising extreme events may occur
as we do not fully understand what is going on.
The Funtowicz and Ravetz model’s axes resemble the same two
dimensio ns that characterise the risk descriptions based on the risk
perspecti ves R = (C&U), consequences C and uncertainties U; see
Appendix A. If the risk is judged as large accordin g to these per-
spectives, extreme events may occur.
What large uncertainties in this setting mean is, however, not
obvious (Aven, 2013 ). It is clearly not the same as a high probabil-
ity as discussed above, as the background knowledge also has to be
reflected. Following the risk perspective (C&U), the uncertainty
dimensio n does not only cover the measure Q used to describe
the uncertainties but also the background knowledge K that the
measure is based on. Hence a high score on the uncertainties is
not only associated with high probabilitie s of occurrences of some
severe consequences but also dependent on judgements made on
the degree of knowledge supporting the probability assignment s.
If we judge the background knowledge as poor this would affect
the total score of the uncertainties being high or low. To make a
judgeme nt about K being poor, many types of aspects need to be
taken into account, for example (Flage and Aven, 2009 ):
v The assumpti ons made represent strong simplifications.
v Data are not available, or are unreliable.
v There is a lack of agreement/ consensus among experts.
v The phenomena involved are not well understood; models are
non-existen t or known/believe d to give poor predictions.
To reflect the importance of background knowled ge and key
assumpti ons, it has been suggested to use an uncertainty measure -
ment Q = (P, U
F
), where U
F
covers qualitative importance assess-
ments of uncertainty factors which are based on key
assumpti ons that the probabilities are founded on (Aven, 2010a;
Flage and Aven, 2009 ).
Many other structure s can be used to define what high uncer-
tainties mean. One interesting category is ‘‘scientific uncertainti es’’
as defined in relation to use of the precautio nary principle. There
exist a number of definitions of this principle (see e.g. Sandin
et al., 2002; Aven, 2011c ), but almost all definitions identify ‘‘scien-
tific uncertainties ’’ as the trigger or criterion for its invocation. As
noted by Aven (2011c), the essential feature of the precautionar y
principle when looking at the many definitions of the principle is
46 T. Aven / Safety Science 57 (2013) 44–51
that it applies when the consequences of the activity considered
could be serious but we do not fully understand what could hap-
pen. What this means in practice is a disputed topic (see e.g. Aven,
2011c; Cox, 2011; North, 2010; Vlek, 2011 ), but a common idea is
that scientific uncertainti es can be linked to the difficulty of estab-
lishing an accurate predictio n model for the consequences. Note
that consequences here are to be understood as also covering
events occurring; in other words, the idea states that we have sci-
entific uncertainti es if there is a lack of consensus in the scientific
community about a model for predicting the occurrence of the
event. Clearly, if a probabili ty model can be justified we cannot re-
fer to scientific uncertainties, as the phenomenon is to a large ex-
tent understood, following this line of argument. The parameters of
the model may be subject to uncertainties, but as long as a proba-
bility model has been justified, the situation is not characterise d by
scientific uncertainti es.
Black swans (surprising, extreme events) can occur in situations
of high risk in the (C&U) sense, in the post-normal science area and
in the face of scientific uncertainti es as discussed above, presumin g
that there is a reference for what can happen (some knowledge/be-
liefs). However, a black swan could also occur in a situation judged
as having rather small or moderate risk in the (C&U) sense. The ex-
perts are convinced that all swans are white, based on thorough re-
search. Then a discovery happens that completely changes the
understand ing of the phenomena, a white swan – a new type of
virus for example – is discovered. The dramatic black swans would
in fact happen when risk is judged as rather small or moderate , as,
according to the prevailing thinking, then there is no reason to be-
lieve that a black swan should occur.
Some types of events occur completely as a surprise based on
the knowled ge available : no one has thought about the type of
event before it occurs – we talk about unknown unknowns, which
is the topic of the next subsection. Clearly, for these types of events
there are scientific uncertainties in this sense; we are not able to
establish an accurate predictio n model for the event.
2.4. Is a black swan an unknown unknown?
The United States Secretary of Defense, Donald Rumsfeld, made
the term ‘‘unknown unknowns’’ familiar to us all on 12 February
2002 at a press briefing, where, addressing the absence of evidence
linking the government of Iraq with the supply of weapons of mass
destruction to terrorist groups, he said:
There are known knowns; there are things we know we know.
We also know there are known unknowns; that is to say we
know there are some things we do not know. But there are also
unknown unknowns – the ones we don’t know we don’t know.
The term has, however, been used long before this. It is men-
tioned for example by Furlong (1984), and in relation to climate
change it is has been commonl y used for many years (e.g. Myers,
1993).
In a risk setting, the idea of unknown unknowns intuitively cap-
tures the fact that the actual events occurring are not covered by
the events identified in the risk description/ris k assessment. Our
focus here is of course on events with extreme consequences. Con-
sider the risk perspective (C&U), or reformulate d by specifically
showing some events A included in C: (A, C&U) (A may for example
represent a terrorist attack or a gas leakage in a process plant).
When speaking about the risk (A, C&U), there are no unknown un-
knowns, as A and C simply express the actual events and conse-
quences of the activity. However , in relation to a risk description
(A
/
, C
/
, Q, K), we may have unknown unknowns (here A
/
and C
/
are
the events and consequences respectively, specified in the risk
assessment, Q is the measure of uncertainty used and K is the back-
ground knowledge, see Appendix A). The point is simply that the A
/
events do not capture the A; we may experience some surprises
relative to A
/
reflecting that A is different from A
/
. For example,
when studying the life of a young person, he or she may die of a
disease not known today; the A
/
events do not cover the true A.
Hence the unknown unknowns are included in the risk concept
but not captured of course by the risk description.
Also if risk is (C&P), i.e. (A, C&P), unknown unknowns reflect
that the actual types of events occurring are not captured by the
risk description. If P is a knowledge- based probability, the relevant
perspecti ves of risk do not, however , cover the unknown un-
knowns, as the risk and the risk description coincide. Seeing P as
a frequentist probability, we are back to the (A, C&U) type of argu-
ments; however , these perspecti ves have limited applicability, as
frequent ist probabilities cannot be justified for non-repetit ive
events and situation s, as commented in Appendix A. Similar types
of analyses can be made for the other risk perspecti ves defined in
Appendix A.
The interesting question is now whether a black swan is to be
considered an unknown unknown. Yes, it is a possible interpreta-
tion: a black swan is an event occurring which was not captured
by the knowledge reflected in the risk description, i.e. A
/
using
the risk perspecti ve (A, C&U). If we study discussions on the inter-
net concerning unknown unknowns and black swans, for example
related to climate change, we see that many people use these
terms more or less interchange ably. Remember Taleb’s definition
of a black swan, where he refers to it as an event which lies outside
the realm of regular expectati ons, because nothing in the past can
convincingly point to its possibility. A reasonable interpretation of
this statement is that if the risk description of the risk assessment
is not able to capture the event, it is an unknown unknown and a
black swan – nothing in the past can convincingly point to its pos-
sibility, interpreting the past in a wide knowled ge sense.
3. Final remarks and conclusions
Based on the above analysis, two main ways of looking at the
black swan concept seems most appropriate : (i) as a rare event
with extreme conseque nces, or as a term for expressing (ii) an ex-
treme, surprisin g event relative to the present knowledge.
Before making a conclusion about the preferred terminology, let
us consider some practical examples. Firstly, think of the Fukushi-
ma Daiichi nuclear disaster in Japan in March 2011. Was this event
a black swan? According to (i), yes, but not according to (ii). The
scientific communi ty had the knowledge to understand that in
the case of a tsunami, which we know occurs from time to time,
extreme consequences would be likely. The situation is better
characteri sed as one where the risk associated with such tsunamis
was accepted, not that the event was surprisin g in relation to the
available knowledge. A reviewer of an earlier version of the present
paper commented that one can oppose this conclusion:
until this event, no one had conceived it a possibility that a tsu-
nami would destroy all back-up systems simultaneou sly, as
well as it (the earthqua ke) would prevent support from outside
to reach the site.
It is tempting to say that if this was the case, something must
have been wrong with the risk assessme nts, but I will not go fur-
ther into this discussion here. The example clearly demonstrat es
the importance of being precise on whose knowled ge we are talk-
ing about.
As another example, think of the terrorist attack in Norway 22
July 2011. Again we conclude that it is a black swan in the sense
of (i) and as for the tsunami case it can be discussed whether it
was also a black swan with respect to (ii). Clearly, the event came
T. Aven / Safety Science 57 (2013) 44–51 47
as a big surprise for the police security services, relative to their
knowledge. It was thus a black swan for them. We can discuss
whether this was a failure of the police security services, but that
is not the issue here. It is not natural to classify the attack as an un-
known unknown as similar types of events have happened before.
The third and last example is the recent financial crisis. Was this
a black swan? Again the answer is a yes for (i), but as above the
conclusion is not so clear for understand ing (ii). There were many
signals and warnings ahead of the problems we have experienced
in the economy, but still it came as a surprise to many, probably
most, people. Some experts would say that the event could have
been foreseen given all the informat ion at the time, others
acknowledged that there were signals of a catastrop he but the ac-
tual situation turned out to be much more severe than one could
have predicted.
It is of course possible to use the term ‘black swan’ in both
senses, (i and ii). However I believe (ii) should be employed , as
the former understand ing would result in too big a class of events,
including those that are simply rare but well understo od, as the
analysis of interpretations 1 and 2 shows in Sections 2.1 and 2.2 .
Definition (ii) (an extreme, surprisin g event relative to the present
knowledge) is in line with Taleb’s definition but is not only focused
on surprises relative to the past; remember Taleb’s ‘‘nothing in the
past can convincingly point to its possibility’’. The key is the knowl-
edge available. It would be better to say ‘‘nothing in our knowledge
can convincingly point to its possibility’’.
What knowled ge means can however be discussed. If we con-
sult the knowledge (knowledge managemen t) literature we see
that there exist a huge number of ideas and definitions out there
for this concept, see Zins (2007), who documents 130 definitions
of data, informati on , and knowledge formulated by 45 scholars.
Rowley (2006, 2007) also provide many definitions. A conceptu al
framework often referred to in this context is the DIKW hierarchy,
which covers the data (D), informat ion (I), knowledge (K) and wis-
dom (W) dimensions (see e.g. Adler, 1986; Ackoff, 1989; Zeleny,
1987; Frické, 2009 ). It is beyond the scope of the present paper
to provide a detailed analysis of this issue, but some comments
are in place. Data and information can be seen as a part of knowl-
edge (given that they are cognitivel y assimilated, Hansson, 2002 ),
but knowledge is also about beliefs. For example, we may think
of a situation where some analysts believe that some potential
attackers do not have the intentions and capacity to perform an at-
tack. Their belief can be based on data and informat ion, modellin g
and analysis. Hence a black swan can be an extreme, surprising
event relative to the historical data present, but it can also be an
extreme, surprising event relative to some relevant beliefs. This
view on knowledge obviously means that it cannot be objective
as a belief is someone’s belief. In general knowledge then needs
to be considered subjective or at best inter-subj ective among peo-
ple, for example experts.
In my view, the above analysis has shown that Taleb has a point
when stressing the need for seeing beyond the standard probabilis-
tic analysis when addressing risk. Lindley’s criticism of Taleb
shows, in my opinion, a lack of understand ing of this need: the
standard probabilistic methods and models used for analysing
uncertainties are not able to predict black swans (in sense ii).
The assumptions that the analyses and models are based on are of-
ten not given the attention that they deserve; refer to the example
of Lindley (2008) discussed in Section 2.2.
The limitations of the probabili ty model-based approach to risk
have been addressed by many risk researchers and analysts (see
e.g. Renn, 1998; Aven, 2011a ) and frameworks have been sug-
gested to be able to meet the challenges raised above. A main cat-
egory of these alternative perspectives on risk solves the
‘‘narrowness ’’ of the probability-ba sed approach by replacing prob-
ability with uncertainties in the definition of risk, and makes a
clear distinctio n between the concept of risk and how it is de-
scribed or measured; see Appendix A.
One may question why it is important to discuss the meaning of
a black swan. Whether a black swan is to be interpreted in line
with (i or ii), could this affect the risk assessme nt and risk manage-
ment? Yes, it could. This paper is motivated by the need to
strengthen the scientific platform of the risk discipline by provid-
ing new insights into the relationship between surprisin g events,
risk, probability and uncertainty. For this discipline , as for all other
scientific disciplines, it is essential that the conceptual basis is so-
lid. However, the present paper is not only of theoretical and foun-
dational interest. The main contributi on of the work is not the
definition of a black swan as such, but the structure develope d to
understa nd and analyse the related features of risk and uncertain-
ties in a risk assessment and risk managemen t context. We saw
this clearly demonstrat ed when studying the example of Lindley
(2008). According to Lindley, there is no need for uncertainty
assessme nt that extends beyond the probabili stic one. Taleb and
many others, including the present author, reject this idea and seek
to build a scientific platform for a more complete approach to risk
and uncertainti es. The present paper can be seen as a contribution
to this end.
In my view, Taleb’s book represents in this respect an important
contributi on, although his work lacks a proper scientific framing. It
has been an aim of the present work to contribute to giving his
ideas and related work a stronger basis with reference to the risk
field, in particular in relation to key concepts such as risk and
probabili ty.
It is beyond the scope of the present paper to discuss in detail to
what degree and how risk assessment and risk managemen t can
help protectin g against black swans. For some recent work
addressing the issue, see Paté-Cornell (2012), Cox (2012) and Aven
(submitted for publication). I would like however to make two
comments . Firstly, although risk assessment can never fully cap-
ture the black swans, improvem ents can and should be made com-
pared to the probabili stic approach that dominates the present
quantitat ive risk assessment practice. To this end, it is considered
essential to establish risk-unce rtainty frameworks that are so
broad that they also capture such events ÷ as what was unknown
at time t could be known by time s, and what is unknown to per-
sons x could be known to persons y. The knowledge dimension
needs to be highlight ed, and much more than what is typically
seen today in risk assessment applicati ons. More research is
needed on how to do this in practice. Secondly, I believe that it is
essential to acknowledge that we cannot produce ‘‘optimal strate-
gies’’ meeting black swans. Different types of optimisa tion meth-
ods seeking robust solutions can be effective in many cases, but
we always need what I refer to as managerial review and judge-
ment, where the decision maker sees beyond the formal decision
support, and gives weight to uncertainties and other concerns
not captured by the formal assessments, including the cautionary
and precautionary principles. See Aven (submitted for publication).
One of the reviewer s of an earlier version of the present paper
stated in a comment to the paper that
there is no need to try to agree on the definition of a ‘‘black
swan’’ as a scientifically based term in this field, therefore I feel
that this issue of a black swan is a non-issue, there is no need to
use this term when dealing with risk and safety, one can
express the problems of unexpected events in relation to risk
management in more precise terms, adjusted to the context at
hand.
I disagree with this reviewer that black swan is a non-issue as
here stated. Firstly, the concept of black swan exists out there
and is commonl y used in relation to risk and safety. The idea has
48 T. Aven / Safety Science 57 (2013) 44–51
gained a lot of attention and is a hot topic in many forums that dis-
cuss safety and risk. We as a scientific and professional environ-
ment cannot just ignore this. We need to provide perspectives
and guidance on what this concept is saying. We need to place this
concept in the framewor ks that the risk field has developed over
the years. And this is exactly what this paper is doing. Secondly,
the risk field needs suitable concepts for reflecting this type of phe-
nomena. The popularity of the black swan concept clearly demon-
strates this, but I would like to add that also from a strictly
professional point of view, there is a need for concepts that de-
scribe what a ‘‘surprise’’ really means in this context. These con-
cepts cannot and should not be limited by the probability based
thinking and ideas as the phenomena that we are trying to charac-
terise extend beyond this paradigm. In the extended non-probabi l-
ity context we need to develop proper terminology – the present
situation is rather chaotic – and the ‘‘black swan’’ concept repre-
sents in my view a useful contribution to this end. We can use
statements as 1–4 in Section 1, but when communicating and dis-
cussing issues linked to ‘‘surprising events’’ my experience is that it
is very helpful to have at hand a term like black swan, that people
can easily relate to. Using the black concept I have noticed in-
creased interest and enthusias m for discussing risk issues. Thirdly
and last, I am convinced that studying the black swan concept pro-
vides new insights into the risk field, about the links between risk,
probability and uncertainties , as was highlighted also in the above
discussion.
To summarise, I conclude that a black swan is to be seen as a
surprising extreme event relative to the present knowledge/beli efs.
Hence the concept always has to be viewed in relation to whose
knowledge/bel iefs we are talking about, and at what time. In a risk
assessment context and following the (A, C&U) risk perspective, a
black swan can be seen as an extreme event A occurring not spec-
ified by the A
/
events of the risk assessme nt - it is a surprise relative
to the knowledge defined by the A
/
events.
4. Uncited references
IRGC (2005), Verma and Verter (2007) and Willis (2007).
Acknowled gment
The author is grateful to several anonymous reviewers for their
useful comments and suggestions to earlier versions of this paper.
Appendix A
A.1. Fundamentals about the concepts of risk, probabilit y and
uncertainty
The following overview provides a lists of the main categories of
definitions/perspectives of risk as used in professional /scientific
contexts (Aven, 2012 ):
(1) Risk = Expected conseque nces (R = EC) or expected utility
(R = EU).
(2) Risk = Probability of an (undesirable) event (R = P).
(3) Risk = Objective Uncertainty (R = OU).
(4) Risk = Uncertain ty about a loss (R = U).
(5) Risk = Potential/pos sibility of a loss (R = PO).
(6) Risk = Probability and scenarios/cons equences/sever ity of
consequences (R = P&C).
(7) Risk = Event or consequence (R = C).
(8) Risk = Conseque nces/damage/sev erity of these + Uncertainty
(R = C&U).
(9) Risk = The effect of uncertainty on objectives (R = ISO).
An example of risk definition (6) is the well-known triplet (s
i
, p
i
,
c
i
), where s
i
is the ith scenario, p
i
is the probability of that scenario,
and c
i
is the conseque nce of the ith scenario, i = 1, 2, . . . , N; i.e. risk
captures : What can happen? How likely is that to happen? If it
does happen, what are the consequences ? (Kaplan and Garrick,
1981).
Rosa (1998, 2003) provides an example of (7): risk is a situation
or event where something of human value (including humans
themselv es) is at stake and where the outcome is uncertain. Exam-
ples of (8) are the definitions used by Aven (2007), expressing that
risk is equal to the two-dimensi onal combinati on of events/conse-
quences (of an activity) and associate d uncertainties , and that of
Aven and Renn (2009), which states that risk is uncertainty about
and severity of the conseque nces (or outcomes) of an activity with
respect to somethin g that humans value. The consequences may be
seen in relation to a reference level (ideal states, planned values,
expected values, objectives ). The category (9) definition is the
one used by ISO (2009a,b).
When risk is defined by consequences and uncertainti es
(R = C&U) (and also when R = C), risk is described by specifying
the events/cons equences C and using a description (measure) of
uncertainty Q. Specifying the events/cons equences means to iden-
tify a set of events/q uantities of interest C
/
that characteri se the
events/co nsequences C. Examples of C
/
are the profit from an
investme nt and the number of injuries in a safety context. Depend-
ing on the principles adopted for specifying C and the choice of Q,
we obtain different perspectives on how to describe/measur e risk.
As a general description of risk we are led to the triplet (C
/
,Q,K),
where K is the knowled ge that C
/
and Q are based on. The most
common tool for representing or expressing the uncertainties U
is probabili ty P, but other tools also exist, including imprecise
(interval) probability and representation s based on the theories
of evidence (belief functions) and possibilit y (Dubois, 2010; Aven
and Zio, 2011 ).
For the definitions that are based on probabili ties and expected
values, different interpretations may apply. Basically, there are two
ways to understand the probability of an event A in a practical set-
ting (Aven, 2012 ):
(i) as a frequent ist probability, which we denote by P
f
(A). This
probability is defined by the fraction of times the event A
occurs when consideri ng an infinite population of similar
situations or scenarios to the one in focus. This concept is
a model concept, a paramete r of a probability model. As
the frequentist probability P
f
(A) is unknown , it has to be
estimated. In this way we obtain a clear distinction
between the underlying concept P
f
(A) and its estimate
P
f
(A)

(say), or
(ii) as a subjective (judgemental, knowledge- based) probability
P(A) = P(A|K), expressing the assessor’s uncertainty (degree
of belief) of the occurrence of event A given the background
knowledge K. To interpret this probabili ty, an uncertainty
standard approach is commonly used: the probabili ty
P(A) = 0.2 (say) means that the assessor compares his/her
uncertainty (degree of belief) about the occurrence of the
event A with the standard of drawing at random a specific
ball from an urn that contains five balls.
The subjectiv e (knowledge-based, judgeme ntal) probabilities
express epistemic uncertainties, whereas the variation generating
the frequentist probabili ties and the probabili ty models, express
aleatory (stochastic) uncertainty . Consider a special designed die
with six outcomes 1, 2, . . . , 6 as usual but without the same sym-
metry of the standard die. In this case we would establish a prob-
ability model expressing that the distribution of outcome s is
T. Aven / Safety Science 57 (2013) 44–51 49
given by (p
1
, p
2
, . . . , p
6
), where p
i
is the frequent ist probabili ty of
outcome i, interpreted as the fraction of outcomes showing i,
and having sum equal to 1. However, in a risk assessment con-
text, repeating the situations may be more difficult, making the
establishment of such models problematic, and conseque ntly also
the distinctio n between aleatory uncertainty and epistemic
uncertainty. The frequentist probability and the probability model
cannot that easily be defined as in the die example. In many cases
they cannot be meaningfull y defined at all, as for example the
frequentist probabili ty of a terrorist attack (Aven and Renn,
2010, p. 80 ).
If probability is understo od as a frequent ist probability, the def-
inition categories (1, 2 and 6) are based on a model concept (prob-
ability model with unknown parameters), and risk has to be
estimated. In the case that probability is subjective, these defini-
tions would, however, represent judgeme nts (degrees of belief)
of the assessors.
If the definitions are based on subjectiv e probabilitie s, they do
not allow for a distinctio n between the concept of risk and how
to measure/de scribe risk; the concept of risk is then the same as
the measureme nt of the risk. If, on the other hand, risk is based
on frequentist probabilities, such a distinction is obtained ÷ be-
tween the concept of risk based on the underlying true frequent ist
probability and the measure ment/descripti on of risk based on esti-
mation of these probabilities. However, in the latter case, the risk
concept is based on modellin g, which means this dichotomy does
not always exist, as this modelling will only be justified in situa-
tions of repeatability .
Statistical data analysis is based on one or the other of two
alternative conceptual foundations : the traditional frequentist ap-
proach and the Bayesian approach. The former is based on well-
known principles of statistical inference, the use of probability
models, the interpretation of probabilities as relative frequencies,
point estimates, confidence interval estimation and hypothesis
testing. By contrast, the Bayesian approach is based on the concept
of subjective probabilitie s and is typically applied in situations in
which there exists only a limited amount of data. In a Bayesian
context, the frequentist probabilitie s are often referred to as
chances (Singpurwalla, 2006; Lindley, 2000 ). Probability models
also constitute a basic pillar of a Bayesian analysis. The idea in such
an analysis is to first establish probability models that adequate ly
represent the aleatory uncertainties . The epistemic uncertainties ,
reflecting incomplete knowledge or lack of knowledge about the
values of the paramete rs of the models, are then represented by
prior subjectiv e probability distribut ions. When new data on the
phenomena studied become available, Bayes’ formula is used to
update the representat ion of the epistemi c uncertainties in terms
of the posterior distribut ions. Finally, the predictiv e distribution s
of the quantities of interest and the observables (for example, the
number of system failures) are derived by applying the law of total
probability. The predictive distribution s are epistemic, but they
also reflect the inherent variability of the phenomena being stud-
ied, i.e. the aleatory uncertainties .
Appendix B
(a) This first part of this appendix covers the black swan exam-
ple presented by Lindley (2008) and shows how he has
arrived at his results.
Let p be the fraction of successe s (white swans) in the infinite
series of trials (total population of swans), and let X
n
be the number
of successes (white swans) in n trials (swans). Furthermore, let Y
m
be the number of successes (white swans) in m new trials (swans).
We will compute the probability of m successes in the new trials
given only successes in the first n trials, i.e. P(Y
m
= m| X
n
= n). By
conditionin g on the true value of p, we find that
P(Y
m
= m[X
n
= n) =
Z
[0;1[
P(Y
m
= m[X
n
= n; p) dH(p[X
n
= n)
=
Z
[0;1[
P(Y
m
= m[p) dH(p[X
n
= n) =
Z
[0;1[
p
m
dH(p[X
n
= n);
(B:1)
where H(p| X
n
= n) is the posteri or distribution of p. Lindley (2008)
assumes a uniform distribu tion for H, and hence the posterior den-
sity f of p given X
n
= n equals:
f (p[X
n
= n) = c P(X
n
= n[p) f (p) = cp
n
1 = (n ÷1) p
n
;
where c is a constant such that the integral over this density equals
one.
Hence (B.1) equals
Z
[0;1[
p
m
(n ÷1)p
n
dp = (n ÷1)=(m÷n ÷1);
as presented by Lindley (2008). We see that if m is equal to one and
n is large, this probability is close to one; i.e., the probability that
the next swan is black is negligib le, but if m is large the probab ility
(B.1) is close to zero; i.e., the probability of a least one black swan in
the large sample of size m is close to one.
(b) The second part of this appendix shows the computati onal
result of the black swan example presented in Section 2.2.
Here the prior probabilitie s give mass 0.2 and 0.8 to the p
values 1.0 and 0.99, respectively .
The task is again to compute P(Y
m
= m| X
n
= n). Following argu-
ments as above, we find
P(Y
m
= m[X
n
= n) = P(Y
m
= m[X
n
= n; p = 1)P(p = 1[X
n
= n)
÷P(Y
m
= m[X
n
= n; p = 0:99)P(p = 0:99[X
n
= n)
= 1 P(p = 1[X
n
= n) ÷0:99
m
P(p = 0:99[X
n
= n):
Now, using Bayes’ formula, it is not difficult to see that
P(p = 1| X
n
= n) = c P(X
n
= n|p = 1) P(p = 1) = c 10.2 and
P(p = 0.99| X
n
= n) = c P(X
n
= n|p = 0.99) P(p = 0.99) = c 0.99
n
0.8,
leading to
P(Y
m
= m| X
n
= n) = [0.2/(0.2 + 0.99
n
0.8)] + [0.8 0.99
m+n
/(0.2 +
0.99
n
0.8)].
We see that in this case if n is large this probabili ty is close to
one; i.e., the probability of at least one black swan occurring is
close to zero; this is also the case for large m values, which is in
contrast to Lindley’s result in (a) above.
References
Ackoff, R.L., 1989. From data to wisdom. Journal of Applied Systems Analysis 16, 3–
9.
Adler, M.J., 1986. A Guidebook to Learning for the Lifelong Pursuit of Wisdom.
Collier Macmillan, New York.
Aven, T., 2007. A unified framework for risk and vulnerability analysis and
management covering both safety and security. Reliability Engineering and
System Safety 92, 745–754.
Aven, T., 2010a. On how to define, understand and describe risk. Reliability
Engineering and System Safety 95, 623–631.
Aven, T., 2010b. On the need for restricting the probabilistic analysis in risk
assessments to variability. Risk Analysis 30, 354–360 (With discussion pp. 381–
384).
Aven, T., 2011a. Selective critique of risk assessments with recommendations for
improving methodology and practice. Reliability Engineering and System Safety
96, 509–514.
Aven, T., 2011b. Quantitative Risk Assessment. The Scientific Platform. Cambridge
University Press, Cambridge.
Aven, T., 2011c. On different types of uncertainties in the context of the
precautionary principle. Risk Analysis 31 (10), 1515–1525 (With discussion
pp. 1538–1542).
Aven, T., 2012. The risk concept – historical and recent development trends.
Reliability Engineering and System Safety 99, 33–44.
50 T. Aven / Safety Science 57 (2013) 44–51
Aven, T., 2013. On Funtowicz & Ravetz’s ‘‘decision stake – system uncertainties’’
structure and recently developed risk perspectives frameworks. Risk Analysis
33 (2), 270–280.
Aven, T., submitted for publication. On how to deal with deep uncertainties in a risk
assessment and management context. Risk Analysis.
Aven, T., Renn, O., 2009. On risk defined as an event where the outcome is uncertain.
Journal of Risk Research 12, 1–11.
Aven, T., Renn, O., 2010. Risk Management and Risk Governance. Springer Verlag,
Berlin.
Aven, T., Zio, E., 2011. Some considerations on the treatment of uncertainties in risk
assessment for practical decision-making. Reliability Engineering and System
Safety 96, 64–74.
Aven, T., Renn, O., Rosa, E., 2011. The ontological status of the concept of risk. Safety
Science 49, 1074–1079.
Cox, T., 2011. Clarifying types of uncertainty: when are models accurate, and
uncertainties small? Risk Analysis 31, 1530–1533.
Cox, T., 2012. Confronting deep uncertainties in risk analysis. Risk Analysis 32 (10),
1607–1629.
Dubois, D., 2010. Representation, propagation and decision issues in risk analysis
under incomplete probabilistic information. Risk Analysis 30, 361–368.
Flage, R., Aven, T., 2009. Expressing and communicating uncertainty in relation to
quantitative risk analysis (QRA). Reliability & Risk Analysis: Theory &
Applications 2 (13), 9–18.
Frické, M., 2009. The knowledge pyramid: a critique of the DIKW hierarchy. Journal
of Information Science 35 (2), 131–142.
Funtowicz, S.O., Ravetz, J.R., 1985. Three types of risk assessment. In: Whipple, C.,
Covello, V.T. (Eds.), Risk Analysis in the Private Sector. Plenum Press, New York.
Furlong, R.B., 1984. Clausewitz and Modern War Gaming: Losing can be better than
winning. Air University Review 35, 4–7.
Hammond, P., 2009. Adapting to the Entirely Unpredictable: Black Swans, Fat Tails,
Aberrant Events, and Hubristic Models. The University of Warwick Bulletin of
the Economics Research Institute, 2009/10, 1, November.
Hansson, S.O., 2002. Uncertainties in the knowledge society. International Social
Science Journal 54 (171), 39–46.
IRGC International Risk Governance Council, 2005. White Paper on Risk
Governance. Towards an Integrative Approach. Author: O. Renn with Annexes
by P. Graham. International Risk Governance Council, Geneva.
ISO, 2009a. Risk management—vocabulary. Guide 73, 2009.
ISO, 2009b. Risk Management – Principles and Guidelines, ISO 31000:2009.
Kaplan, S., Garrick, B.J., 1981. On the quantitative definition of risk. Risk Analysis 1,
11–27.
Lindley, D.V., 2000. The philosophy of statistics. The Statistician 49 (3), 293–337.
Lindley, D.V., 2008. The Black Swan: the impact of the highly improbable. Reviews.
Significance (March), 42.
Myers, N., 1993. Biodiversity and the precautionary principle. Ambio 22 (2/3), 74–
79 (Biodiversity: Ecology, Economics, Policy).
North, W., 2010. Probability theory and consistent reasoning. Risk Analysis 30 (3),
377–380.
Paté-Cornell, M.E., 2012. On black swans and perfect storms: risk analysis and
management when statistics are not enough. Risk Analysis 32 (11), 1823–1833.
Renn, O., 1998. Three decades of risk research: accomplishments and new
challenges. Journal of Risk Research 1 (1), 49–71.
Rosa, E.A., 1998. Metatheoretical foundations for post-normal risk. Journal of Risk
Research 1, 15–44.
Rosa, E.A., 2003. The logical structure of the social amplification of risk framework
(SARF): metatheoretical foundation and policy implications. In: Pidgeon, N.,
Kaspersen, R.E., Slovic, P. (Eds.), The Social Amplification of Risk. Cambridge
University Press, Cambridge.
Rowley, J., 2006. Where is the wisdom that we have lost in knowledge? Journal of
Documentation 62 (2), 251–270.
Rowley, J., 2007. The wisdom hierarchy: representations of the DIKW hierarchy.
Journal of Information Science 33 (2), 163–180.
Sandin, P., Peterson, M., Hansson, S.O., Rudén, C., Juthe, A., 2002. Five charges
against the precautionary principle. Journal of Risk Research 5, 287–299.
Singpurwalla, N.D., 2006. Reliability and Risk: A Bayesian Perspective. Wiley,
Chichester.
Taleb, N.N., 2007. The Black Swan: The Impact of the Highly Improbable. Penguin,
London.
Taleb, N.N., 2010. The Black Swan: The Impact of the Highly Improbable, second ed.
Penguin, London.
Taleb, N.N., 2011. <http://www.fooledbyrandomness.com/DerivTBS.htm> (accessed
12.12.11).
Verma, M., Verter, V., 2007. Railroad transportation of dangerous goods: population
exposure to airborne toxins. Computers and Operations Research 34, 1287–
1303.
Vlek, C., 2011. Straightening out the grounds for precaution: a commentary and
some suggestions about Terje Aven’s ‘‘On Different Types of Uncertainties...’’.
Risk Analysis 31, 1534–1537.
Willis, H.H., 2007. Guiding resource allocations based on terrorism risk. Risk
Analysis 27 (3), 597–606.
Zeleny, M., 1987. Management support systems: towards integrated knowledge
management. Human Systems Management 7 (1), 59–70.
Zins, C., 2007. Conceptual approaches for defining data, information, and
knowledge. Journal of the American Society for Information Science and
Technology 58 (4), 479–493.
T. Aven / Safety Science 57 (2013) 44–51 51