TAIL RISK WORKING PAPERS

On the statistical properties and tail risk of violent

conflicts

arXiv:1505.04722v2 [stat.AP] 19 Oct 2015

Applied

Pasquale Cirillo , Nassim Nicholas Taleb

Probability Group, Delft University of Technology, Tandon School of Engineering, New York
University

as defined by [26]. But should the distribution of war casualties

have an infinite mean, the annihilation of the human species
would be just a matter of time, and the sample properties we
can compute from data have no meaning at all in terms of
statistical inference. In reality, a simple argument allows us
to rule out the infiniteness of the mean: no event or series of
events can kill more than the whole world population. The
support of the distribution of war casualties is necessarily
bounded, and the true mean cannot be infinite.
Let [L, H] be the support of the distribution of war casualties today. L cannot be smaller than 0, and we can safely
fix it at some level L >> 0 to ignore those small events
that are not readily definable as armed conflict [44]. As to H,
its value cannot be larger than the world population, i.e. 7.2
billion people in 20151 [40].
If Y is the random variable representing war casualties, its
range of variation is very large and equal to H L. Studying
the distribution of Y can be difficult, given its bounded but
extremely wide support. Finding the right parametric model
among the family of possible ones is hard, while nonparametric approaches are more difficult to interpret from a risk
analysis point of view.
Our approach is to transform the data in order to apply the
powerful tools of extreme value theory. Since H < 1 we
suggest a log-transformation of the data. This allows to use
tools such as [9]. Theoretical results such as the Pickands,
Balkema and de Haans Theorem allow us to simplify the
choice of the model to fit to our data.
Let L and H be respectively the lower and the upper bound
of a random variable Y, and define the function

H Y
'(Y ) = L H log
.
(1)
H L

AbstractWe examine statistical pictures of violent conflicts

over the last 2000 years, finding techniques for dealing with
incompleteness and unreliability of historical data.
We introduce a novel approach to apply extreme value theory
to fat-tailed variables that have a remote, but nonetheless finite
upper bound, by defining a corresponding unbounded dual
distribution (given that potential war casualties are bounded by
the world population).
We apply methods from extreme value theory on the dual
distribution and derive its tail properties. The dual method
allows us to calculate the real mean of war casualties, which
proves to be considerably larger than the sample mean, meaning
severe underestimation of the tail risks of conflicts from naive
observation. We analyze the robustness of our results to errors
in historical reports, taking into account the unreliability of
accounts by historians and absence of critical data.
We study inter-arrival times between tail events and find that
no particular trend can be asserted.
All the statistical pictures obtained are at variance with the
prevailing claims about "long peace", namely that violence has
been declining over time.

I. I NTRODUCTION
Since the middle of last century, there has been a multidisciplinary interest in wars and armed conflicts (quantified in
terms of casualties), see for example [5], [15], [24], [30], [31],
[32], [44] and [45]. Studies have also covered the statistics
of terrorism, for instance [8], [35], and the special issue of
Risk Analysis on terrorism [33]. From a statistical point of
view, recent contributions have attempted to show that the
distribution of war casualties (or terrorist attacks victims)
tends to have heavy tails, characterized by a power law
decay [8] and [14]. Often, the analysis of armed conflicts
falls within the broader study of violence [6], [30], with
the aim to understand whether we as human are more or
less violent and aggressive than in the past and what role
institutions played in that respect. Accordingly, the public
intellectual arena has witnessed active debates, such as the
one between Steven Pinker on one side, and John Gray on
the other concerning the hypothesis that the long peace was
a statistically established phenomenon or a mere statistical
sampling error that is characteristic of heavy-tailed processes,
[16] and [27] the latter of which is corroborated by this paper.
Using a more recent data set containing 565 armed conflicts
with more than 3000 casualties over the period 1-2015 AD, we
confirm that the distribution of war casualties exhibits a very
heavy right-tail. The tail is so heavy that at first glance
war casualties could represent an infinite-mean phenomenon,

It is easy to verify that

1) ' is "smooth": ' 2 C 1 ,

2) ' 1 (1) = H,
3) ' 1 (L) = '(L) = L.
Then Z = '(Y ) defines a new random variable with lower
bound L and an infinite upper bound. Notice that the transformation induced by '() does not depend on any of the
parameters of the distribution of Y . In what follows, we will
call the distributions of Y and Z, respectively the real and the
dual distribution.
1 Todays world population can be taken as the upper bound, as never before
humanity reached similar numbers.

TAIL RISK WORKING PAPERS

B. Log-rescaled War Casualties w.r.t. Population

A. Actual War Casualties w.r.t. Population

WW2

80'000'000

1'400'000'000

1'050'000'000

Casualties

Casualties

60'000'000

Three!
Kingdoms

40'000'000

An Lushan

20'000'000

Three!
Kingdoms

410

820

700'000'000
An Lushan
WW2

350'000'000

1230

1640

2050

1230

1640

C. Actual War Casualties w.r.t. Impact

D. Log-rescaled War Casualties w.r.t. Impact

1'400'000'000

60'000'000

2050

Three!
Kingdoms

WW2

1'050'000'000
Three!
Kingdoms

Casualties

Casualties

820

Time

80'000'000

40'000'000

20'000'000

410

Time

410

820

An Lushan
350'000'000

An Lushan

700'000'000

1230

1640

2050

Time

WW2

410

820

1230

1640

2050

Time

Fig. 1: War casualties over time, using raw (A,C) and dual (B,D) data. The size of each bubble represents the size of each
event with respect to todays world population (A,B) and with respect to the total casualties (raw: C, rescaled: D) in the data
set.
By studying the tail properties of the dual distribution (the
one with an infinite upper bound), using extreme value theory,
we will be able to obtain, by reverting to the real distribution,
what we call the shadow mean of war casualties. We will show
that this mean is at least 1.5 times larger than the sample mean,
but nevertheless finite.
We assume that many observations are missing from the
dataset (from under-reported conflicts), and we base on analysis on the fact that war casualties are just imprecise estimates
[37], on which historians often have disputes, without anyones
ability to verify the assessments using period sources. For
instance, an event that took place in the eighth century, the An
Lushan rebellion, is thought to have killed 13 million people,
but there no precise or reliable methodology to allow us to
trust that number which could be conceivably one order of
magnitude smaller.2 . Using resampling techniques, we show
that our tail estimates are robust to changes in the quality and
2 For a long time, an assessment of the drop in population in China was
made on the basis of tax census, which might be attributable to a drop in the
aftermath of the rebellion in surveyors and tax collectors.[4]

reliability of data. Our results and conclusions will replicate

even we missed a third of the data. When focusing on the more
reliable set covering the last 500 years of data, one cannot
observe any specific trend in the number of conflicts, as large
wars appear to follow a homogeneous Poisson process. This
memorylessness in the data conflicts with the idea that war
violence has declined over time, as proposed by [15] or [30].
The paper is organized as follows: Section I describes our
data set and analyses the most significant problems with the
quality of observations; Section III is a descriptive analysis of
data; it shows some basic result, which are already sufficient
to refute the thesis [15] that we are living in a more peaceful
world on the basis of statistical observations; Section IV
contains our investigations about the upper tail of the dual
distribution of war casualties as well as discussions of tail
risk; Section V deals with the estimation of the shadow mean;
in Section VI we discuss the robustness of our results to
imprecision and errors in the data; finally, Section VII looks
at the number of conflicts over time, showing no visible trend
in the last 500 years.

TAIL RISK WORKING PAPERS

II. A BOUT THE DATA

The data set contains 565 events over the period 1-2015
AD; an excerpt is shown in Table I. Events are generally
armed conflicts3 , such as interstate wars and civil wars, with
a few exceptions represented by violence against citizens
perpetrated by the bloodiest dictatorships, such as Stalins
and Mao Zedongs regimes. These were included in order
to be consistent with previous works about war victims and
violence, e.g. [30] and [32].
We had to deal with the problem of inconsistency and lack
of uniformity in the attribution of casualties by historians.
Some events such as the siege of Jerusalem include death
from famine, while for other wars only direct military victims
are counted. It might be difficult to disentangle death from
direct violence from those arising from such side effects as
contagious diseases, hunger, rise in crime, etc. Nevertheless,
as we show in our robustness analysis these inconsistencies
do not affect out results in contrast with analyses done on
thin-tailed data (or data perceived to be so) where conclusions
can be reversed on the basis of a few observations.
The different sources for the data are: [5], [18], [25], [29],
[30], [32], [44], [45] and [46]. For websites like Necrometrics
[25], data have been double-checked against the cited references. The first observation in our data set is the Boudiccas
Revolt of 60-61 AD, while the last one is the international
armed conflict against the Islamic State of Iraq and the Levant,
still open.
We include all events with more than 3000 casualties
(soldiers and civilians) in absolute terms, without any rescaling
with respect to the coeval world population. More details about
rescaling in Subsection II-D. The choice of the 3000 victims
threshold was motivated by three main observations:
Conflicts with a high number of victims are more likely to
be registered and studied by historians. While it is easier
today to have reliable numbers about minor conflicts
although this point has been challenged by [36] and [37]
it is highly improbable to ferret out all smaller events
that took place in the fifth century AD. In particular,
a historiographical bias prevents us from accounting for
much of the conflicts that took place in the Americas and
Australia, before their discovery by European conquerors.
A higher threshold gives us a higher confidence about
the estimated number of casualties, thanks to the larger
number of sources, even if, for the bloodiest events
of the far past, we must be careful about the possible
exaggeration of numbers.
The object of our concern is tail risk. The extreme value
techniques we use to study the right tail of the distribution
of war casualties imposition of thresholds, actually even
larger than 3000 casualties.
To rescale conflicts and expressing casualties in terms of
todays world population (more details in Subsection II-D),
3 We refer to the definition of [44], according to which An armed conflict
is a contested incompatibility which concerns government and/or territory
where the use of armed force between two parties, of which at least one is
the government of a state, results in at least 25 battle-related deaths". This
definition is also compatible with the Geneva Conventions of 1949.

we relied on the population estimates of [22] for the period

1-1599 AD, and of [39] and [40] for the period 1650-2015
AD. We used:

Century estimates from 1 AD to 1599.

Half-century estimates from 1600 to 1899.
Decennial estimates from 1900 to 1949.
Yearly estimates from 1950 until today.

A. Data problems
Accounts of war casualties are often anecdotal, spreading
via citations, and based on vague estimates, without anyones
ability to verify the assessments using period sources. For
instance, the independence war of Algeria has various estimates, some from the French Army, others from the rebels,
and nothing scientifically obtained [19].
This can lead to several estimates for the same conflict,
some more conservative and some less. Table I, shows different estimates for most conflicts. In case of several estimates,
we present the minimum, the average and the maximum one.
Interestingly, as we show later, choosing one of the three as
the true" estimate does not affect our results thanks to the
scaling properties of power laws.
Conflicts, such as the Mongolian Invasions, which we refer
to as named" conflicts, need to be treated with care from a statistical point of view. Named conflicts are in fact artificial tags
created by historians to aggregate events that share important
historical, geographical and political characteristics, but that
may have never really existed as a single event. Under the
portmanteau Mongolian Invasions (or Conquests), historians
collect all conflicts related to the expansion of the Mongol
empire during the thirteenth and fourteenth centuries. Another
example is the so-called Hundred-Years War in the period
1337-1453. Aggregating all these events necessarily brings to
the creation of very large fictitious conflicts accounting for
hundreds of thousands or million casualties. The fact that, for
historical and historiographical reasons, these events tends to
be more present in antiquity and the Middle Ages could bring
to a naive overestimation of the severity of wars in the past.
Notice that named conflicts like the Mongolian Invasions are
different from those like WW1 or WW2, which naturally also
involved several tens of battles in very different locations, but
which took place in a much shorter time period, with no major
time separation among conflicts.
A straightforward technique could be to set a cutoff-point
of 25 years for any single event an 80 years event would
be divided into three minor ones. However it remains that the
length of the window remains arbitrary. Why 25 years? Why
not 17?
As we show in Section IV, a solution to all these problems
with data is to consider each single observation as an imprecise
estimate, a fuzzy number in the definition of [43]. Using Monte
Carlo methods we have shown that, if we assume that the
real number of casualties in a conflict is uniformly distributed
between the minimum and the maximum in the available
data, the tail exponent is not really affected (apart from the

TAIL RISK WORKING PAPERS

TABLE I: Excerpt of the data set of war casualties. The original data set contains 565 events in the period 1-2015 AD. For
some events more than one estimate is available for casualties and we provide: the minimum (Min), the maximum (Max), and
an intermediate one (Mid) according to historical sources. Casualties and world population estimates (Pop) in 10000.
Event
Boudiccas Revolt
Three Kingdoms
An Lushans Reb.
Sicilian Vespers
WW1
WW2

Start
60
220
755
1282
1914
1939

End
61
280
763
1302
1918
1945

obvious differences in the smaller decimals).4 Similarly, our

results remain invariant if we remove/add a proportion of the
observations by bootstrapping our sample and generating new
data sets.
We believe that this approach to data as imprecise observations is one of the novelties of our work, which makes our
conclusions more robust to scrutiny about the quality of data.
We refer to Section IV for more details.
B. Missing events
We are confident that there are many conflicts that are
not part of our sample. For example we miss all conflicts
among native populations in the American continent before
its European discovery, as no source of information is actually
available. Similarly we may miss some conflicts of antiquity
in Europe, or in China in the sixteenth century. However, we
can assume that the great part of these conflicts is not in the
very tail of the distribution of casualties, say in the top 10 or
20%. It is in fact not really plausible to assume that historians
have not reported a conflict with 1 or 2 million casualties, or
that such an event is not present in [25] and [29], at least for
what concerns Europe, Asia and Africa.
Nevertheless, in Subsection VI, we deal with the problem
of missing data in more details.
C. The conflict generator process
We are not assuming the existence of a unique conflict
generator process. It is clear that all conflicts of humanity
do not share the same set of causes. Conflicts belonging to
different centuries and continents are likely to be not only
independent, as already underlined in [32] and [18], but also
to have different origins. We thus avoid performing time
series analysis beyond the straightforward investigation of the
existence of trends. While it could make sense to subject the
data to specific tests, such as whether an increase/decrease in
war casualties is caused by the lethality of weapons, it would
be unrigorous and unjustified to look for an autoregressive
component. How could the An Lushan rebellion in China (755
AD) depend on the Siege of Constantinople by Arabs (717
AD), or affect the Viking Raids in Ireland (from 795 AD on)?
But this does not mean that all conflicts are independent: during WW2, the attack on Pearl Harbor and the Battle of France
4 Our

results do not depend on choice of the uniform distribution, selected

for simplification. All other bounded distributions produce the same results
in the limit.

Min
70
3600
800
1466
4823

Mid
7.52
3800
1300
0.41
1544
7300

Max
8.04
4032
3600
0.80
1841
8500

Pop
19506
20231
24049
39240
177718
230735

were not independent, despite the time and spatial separation,

and thats why historians merge them into one single event.
And while we can accept that most of the causes of WW2 are
related to WW1, when studying numerical casualties, we avoid
translating the dependence into the magnitude of the events: it
would be naive to believe that the number of victims in 1944
depended on the death toll in 1917. How could the magnitude
of WW2 depend on WW1?
Critically, when related conflicts are aggregated, as in the
case of WW2 or for the Hundred Years war (1337-1453),
our data show that the number of conflicts over time if
we focused on very destructive events in the last 500/600
years and we believed in the existence of a conflict generator
process is likely to follow a homogeneous Poisson process,
as already observed by [18] and [32], thus supporting the idea
that wars are randomly distributed accidents over time, not
following any particular trend. We refer to Section IV for more
details.
D. Data transformations
We used three different types of data in our analyses: raw
data, rescaled data and dual data.
1) Raw data: Presented as collected from the different
sources, and shown in Table I. Let Xt be the number of
casualties in a given conflict at time t, define the triplet
{Xt , Xtl , Xtu }, where Xtl and Xtu represent the lower and
upper bound, if available.
2) Rescaled data: The data is rescaled with respect to
the current world population (7.2 billion5 ). Define the triplet
l
l P2015
u
u P2015
}, where
{Yt = Xt P2015
P t , Yt = X t P t , Yt = X t P t
P2015 is the world population in 2015 and Pt is the population
at time t = 1, ..., 2015.6 Rescaled data was used by Steven
Pinker [30] to account for the relative impact of a conflict.
This rescaling tends to inflate past conflicts, and can lead
to the naive statement that violence, if defined in terms of
casualties, has declined over time. But we agree with those
scholars like Robert Epstein [12] stating [...] why should we
be content with only a relative decrease? By this logic, when
we reach a world population of nine billion in 2050, Pinker
will conceivably be satisfied if a mere two million people are
killed in war that year". In this paper we used rescaled data
5 According to the United Nations Department of Economic and Social
Affairs [40]
6 If for a given year, say 788 AD, we do not have a population estimate,
we use the closest one.

TAIL RISK WORKING PAPERS

Right Tail: 1-F(y)

for comparability reasons, and to show that even with rescaled

data we cannot really state statistically that war casualties have
declined over time (and thus violence).
3) Dual data: those we obtain using the log-rescaling of
equation (1), the triplet {Zt = '(Yt ), Ztl = '(Ytl ), Ztu =
'(Ytu )}. The input for the function '() are the rescaled values
Yt .
Removing the upper bound allows us to combine practical
convenience with theoretical rigor. Since we want to apply the
tools of extreme value theory, the finiteness of the upper bound
is fundamental to decide whether a heavy-tailed distribution
falls in the Weibull (finite upper bound) or in the Frchet
class (infinite upper bound). Given their finite upper bound,
war casualties are necessarily in the Weibull class. From a
theoretical point of view the difference is large [9], especially
for the existence of moments. But from a practical point of
view, heavy-tailed phenomena are better modeled within the
Frchet class. Power laws belong to this class. And as observed
by [11], when a fat-tailed random variable Y has a very large
finite upper bound H, unlikely to be ever touched (and thus
observed in the data), its tail, which from a theoretical point
of view falls in the so-called Weibull class, can be modeled
as if belonging to the Frchet one. In other words, from a
practical point of view, when H is very large, it is impossible
to distinguish the two classes; and in most of the cases this is
indeed not a problem.
Problems arise when the tail looks so thick that even the
first moment appears to be infinite, when we know that this
is not possible. And this is exactly the case with out data, as
we show in the next Section.

Real Tail
Apparent Tail

Fig. 2: Graphical representation (Log-Log Plot) of what may

happen if one ignores the existence of the finite upper bound
H, since only M is observed.
Figure 2 shows illustrates the need to separate real and dual
data. For a random variable Y with remote upper bound H,
the real tail is represented by the continuous line. However, if
we only observe values up to M , and we ignore the existence
of H, which is unlikely to be reached and therefore hardly
observable, we could be inclined to believe the the tail is
the dotted one, the apparent one. The two tails are indeed

essentially indistinguishable for most cases, but the divergence

is evident when we approach H. Hence our transformation.
Notice that '(y) y for very large values of H. This means
that for a very large upper bound, unlikely to be reached as
the world population, the results we get for the tail of Y and
Z = '(Y ) are essentially the same most of the times. This is
exactly what happens with our data, considered that no armed
conflict has ever killed more than 19% of the world population
(the Three Kingdoms, 184-280 AD). But while Y is bounded,
Z is not.
Therefore we can safely model Z as belonging to the
Frchet class, study its tail behavior, and then come back to
Y when we are interested in the first moments, which under
Z could not exist.
III. D ESCRIPTIVE DATA A NALYSIS
Figure 1 showing casualties over time, is composed of four
subfigures, two related to raw data and two related to dual
data7 . For each type of data, using the radius of the different
bubbles, we show the relative size of each event in terms of
victims with respect to the world population and within our
data set (what we call Impact). Note that the choice of the
type of data (or of the definition of size) may lead to different
interpretation of trends and patterns. From rescaled and dual
data, one could be lead to superficially infer a decrease in the
number of casualties over time.
As to the number of armed conflicts, Figure 1 seems to
suggest an increase over time, as we see that most events are
concentrated in the last 500 years or so, an apparent illusionary
increase most likely due to a reporting bias for the conflicts
of antiquity and early Middle Ages.
Figure 3 shows the histogram of war casualties, when
dealing with raw data. Similar results are obtained using
rescaled and dual casualties. The graphs suggests the presence
of a long right tail in the distribution of victims, and the
maximum is represented by WW2, with an estimated amount
of about 73 million victims.
The presence of a Paretian tail is also supported by the
quantile-quantile plot in Figure 4, where we use the exponential distribution as a benchmark. The concave behavior we
can observe in the plot is considered a good signal of fat-tailed
distribution [11], [21].
In order to investigate the presence of a right fat tail, we
can make use of another graphical tool: the meplot, that is,
mean excess function plot.
Let X be a random variable with distribution F and right
endpoint xF (i.e. xF = sup{x 2 R : F (x) < 1}). The
function
R1
(t u)dF (t)
e(u) = E[X u|X > u] = u R 1
,
0 < u < xF ,
dF (t)
u
(2)
is called mean excess function of X (mef). The empirical mef
of a sample X1 , X2 ,..., Xn is easily computed as
Pn
(Xi u)
en (u) = Pi=1
,
(3)
n
i=1 1{Xi >u}
7 Given Equation (1), rescaled and dual data are approximately the same,
hence there is no need to show further pictures.

TAIL RISK WORKING PAPERS

2.0e+07
1.5e+07

Mean Excess

1.0e+07

300
200
100

Frequency

400

2.5e+07

500

Histogram of casualties

5.0e+06

WW2

0e+00

2e+07

4e+07

6e+07

casualties

0e+00

Fig. 3: Histogram of war casualties, using raw data. A long

fat right tail is clearly visible.
that is the sum of the exceedances over the threshold u divided
by the number of such data points. Interestingly, the mean
excess function is a way of characterizing distributions within
the class of continuous distributions [21]. For example, power
law distributions are characterized by the van der Wijks law
[41], that is, by a mean excess function linearly increasing in
the threshold u, while an exponential distribution of parameter
would show a constant mean excess function with value .
Figure 5 shows the the meplot of war casualties, again using
raw data. The graph is easily obtained by plotting the pairs
{(Xi:n , en (Xi:n )) : i = 1, ..., n}, where Xi:n is the i th order
statistic, used as a threshold. The upward trend in Figure 5 is
a further signal of fat-tailed data, as discussed for example in
[7] and [11].

1e+07

2e+07

3e+07

4e+07

Threshold

Fig. 5: Mean Excess Function Plot (meplot) for war casualties,

using raw data. A steep upward trend is visible, suggesting the
presence of a Paretian right tail.
variance is not finite, as is often the case with financial data, it
becomes more problematic to build confidence intervals for the
mean. Similarly, if the third moment (skewness) is not defined,
it is risky to build confidence intervals for the variance.
An interesting graphical tool showing the behavior of moments in a given sample, is the Maximum-to-Sum plot, or
MS Plot. The MS Plot relies on simple consequence of the
law of large numbers [11]. For a sequence X1 , X2 , ..., Xn
of nonnegative i.i.d. random variables, if for p = 1, 2, 3...,
p
p
p
a.s.
E[X p ] < 1, then
0 as n ! 1,
Pn Rn p = Mn /Sn !
p
p
where Sn =
i=1 Xi is the partial sum, and Mn =
p
p
max(X1 , ..., Xn ) the partial maximum.

0.8
0.6
Rn
0.4
0.2
0.0

0.0

100

200

300

400

500

Fig. 4: Exponential qq-plot of war casualties, using raw data.

The clear concave behavior suggest the presence of a right fat
tail.
A characteristic of the power law class is the non-existence
(infiniteness) of moments, at least the higher-order ones, with
important consequences in terms of statistical inference. If the

0.8

1.0

500

0.0

0.2

0.4

Rn

0.6

0.8
0.6
Rn

6e+07

400

MSplot for p= 4

0.2

4e+07
Ordered Data

300

MSplot for p= 3

0.0

2e+07

200

0.4

1
0
0e+00

100

1.0

Exponential Quantiles

0.2

0.4

Rn

0.6

0.8

1.0

MSplot for p= 2

1.0

MSplot for p= 1

100

200

300
n

400

500

100

200

300

400

500

Fig. 6: Maximum to Sum plot to verify the existence of the

first four moments of the distribution of war casualties, using
raw data. No clear converge to zero is observed, for p = 2, 3, 4,
suggesting that the moments of order 2 or higher may not be
finite.
Figures 6 and ?? show the MS Plots of war casualties for

TAIL RISK WORKING PAPERS

TABLE II: Average inter-arrival times (in years) and their

mean absolute deviation for events with more than 0.5, 1, 2, 5,
10, 20 and 50 million casualties, using raw (Ra) and rescaled
data (Re).

0.8
0.6
Rn
0.4
0.2
0.0

0.0

0.2

0.4

Rn

0.6

0.8

1.0

MSplot for p= 2

1.0

MSplot for p= 1

200

300

400

500

100

200

300

400

MSplot for p= 3

MSplot for p= 4

500

0.8
0.6
Rn
0.4

MAD Ra
35.20
47.73
72.84
113.57
136.88
261.31
325.50

Avg Re
9.63
13.28
20.20
34.88
52.23
73.12
103.97

MAD Re
15.91
20.53
28.65
46.85
63.91
86.19
114.25

0.2

Avg Ra
23.71
34.08
56.44
93.03
133.08
247.14
345.50

0.0

0.0

0.2

0.4

Rn

0.6

0.8

1.0

100

1.0

Thresh.
0.5
1
2
5
10
20
50

100

200

300

400

500

100

200

300

400

500

Fig. 7: Maximum to Sum plot to verify the existence of

the first four moments of the distribution of war casualties,
using rescaled data. No clear converge to zero is observed,
suggesting that all the first four moments may not be finite.
raw and rescaled data respectively. No finite moment appears
to exist, no matter how much data is used: it is in fact clear
that the Rn ratio does not converge to 0 for p = 1, 2, 3, 4, thus
suggesting that the distribution of war casualties has such a
fat right tail that not even the first moment exists. This is
particularly evident for rescaled data.
To compare to thin tailed situations, 8 shows the MS Plot of
a Pareto ( = 1.5): notice that the first moment exists, while
the other moments are not defined, as the shape parameter
equals 1.5. For thin tailed distributions such as the Normal,
the MS Plot rapidly converges to 0 for all p.

0.8
0.6
Rn
0.4
0.2
0.0

0.0

0.2

0.4

Rn

0.6

0.8

1.0

MSplot for p= 2

1.0

MSplot for p= 1

200

300

400

500

100

200

300

400

MSplot for p= 3

MSplot for p= 4

500

0.8
0.6
Rn
0.4
0.2
0.0

0.0

0.2

0.4

Rn

0.6

0.8

1.0

100

1.0

100

200

300
n

400

500

100

200

300

400

500

Fig. 8: Maximum to Sum plot of a Pareto(1.5): as expected,

the first moment is finite (Rn ! 0), while higher moments do
not exists (Rnp is erratic and bounded from 0 for p = 2, 3, 4).
Table II provides information about armed conflicts and
their occurrence over time. Note that we have chosen very
large thresholds to minimize the risk of under-reporting. We

compute the distance, in terms of years, between two timecontiguous conflicts, and use for measure of dispersion the
mean absolute deviation (from the mean). Table II shows the
average inter-arrival times between armed conflicts with at
least 0.5, 1, 2, 5, 10, 20 and 50 million casualties. For a conflict
of at least 500k casualties, we need to wait on average 23.71
years using raw data, and 9.63 years using rescaled or dual
data (which, as we saw, tend to inflate the events of antiquity).
For conflicts with at least 5 million casualties, the time delay
is on average 93.03 or 34.88, depending on rescaling. Clearly,
the bloodier the conflict, the longer the inter-arrival time. The
results essentially do not change if we use the mid or the
ending year of armed conflicts.
The consequence of this analysis is that the absence of a
conflict generating more than, say, 5 million casualties in the
last sixty years highly insufficient to state that their probability
has decreased over time, given that the average inter-arrival
time is 93.03 years, with a mean absolute deviation of 113.57
years! Unfortunately, we need to wait for more time to assert
whether we are really living in a more peaceful era,: present
data are not in favor (nor against) a structural change in
violence, when we deal with war casualties.
Section I asserted that our data set do not form a proper time
series, in the sense that no real time dependence is present
(apart from minor local exceptions).8 Such consideration is
further supported by the so-called record plot in Figure 9.
This graph is used to check the i.i.d. nature of data, and
it relies on the fact that, if data were i.i.d., then records
over time should follow a logarithmic pattern [11]. Given a
sequence X1 , X2 , ..., and defined the partial maximum Mn ,
an observation xn is called a record if Xn = Mn = xn .
For the sake of completeness, if we focus our attention on
shorter periods, like the 50 years following WW2, a reduction
in the number and the size of the conflict can probably be
observed the so-called Long Peace of [15], [24] and [30];
but recent studies suggest that this trend could already have
changed in the last years [20], [27]. However, in our view, this
type of analysis is not meaningful, once we account for the
8 Even if redundant, given the nature of our data and the descriptive analysis
performed so far, the absence of a relevant time dependence can be checked
by performing a time series analysis of war casualties. No significant trend,
season or temporal dependence can be observed over the entire time window.

TAIL RISK WORKING PAPERS

ci = 0.95

50

100

200

500

Fig. 9: Record plot for testing the plausibility of the i.i.d.

hypothesis, which seems to be supported.
extreme nature of armed conflicts, and the long inter-arrival
times between them, as shown in Section III.
IV. TAIL RISK OF ARMED CONFLICTS
Our data analysis in Section III suggesting a heavy right
tail for the distribution of war casualties, both for raw and
rescaled data is consistent with the existing literature, e.g. [8],
[14], [35] and their references. Using extreme value theory,
or EVT [2], [3], [9], [11], [17], [38], we use the Generalized
Pareto Distribution (GPD). The choice is due to the Pickands,
Balkema and de Haans Theorem [1], [28].
Consider a random variable X with distribution function F ,
and call Fu the conditional df of X above a given threshold
u. We can then define the r.v. W , representing the rescaled
excesses of X over the threshold u, so that
F (u + w) F (u)
,
1 F (u)

for 0 w xF u, where xF is the right endpoint of the

underlying distribution F .
Pickands, Balkema and de Haan [1], [28] have showed that
for a large class of underlying distribution functions F , and
a large u, Fu can be approximated by a Generalized Pareto
distribution (hence GPD), i.e. Fu (w) ! G(w), as u ! 1
where
(
1 (1 + w ) 1/ if 6= 0
G(w) =
, w 0. (4)
w
1 e
if = 0
The parameter , known as the shape parameter, is the central
piece of information for our analysis: governs the fatness of
the tails, and thus the existence of moments. The moment of
order p of a Generalized Pareto distributed random variable
only exists if and only if < 1/p [11].
The Pickands, Balkema and de Haans Theorem thus allows
us to focus on the right tail of the distribution without caring
too much about what happens below the threshold u. One
powerful property of the generalized Pareto is the tail stability

u w|X > u) =

Fu (w) = P (X

where Xk,n is the k-th upper order statistics out of a sample

of n observations.

20
Trials

10

with respect to threshold [11]. Formally, if W GPD(, 1 ),

for W
u1 , then W GPD(, 2 ) for W
u2 > u1 . In
other words: increasing the threshold does not affect the shape
parameter. What changes is only the scale parameter.
We start by identifying the threshold u above which the
GPD approximation may hold. Different heuristic tools can
be used for this purpose, from Zipf plots, i.e. log log plots of
the survival function, to mean excess function plots (as the one
in Figure 5), where one looks for the threshold above which
it is possible to visualize if present the linearity that
is typical of fat-tailed phenomena. Other possibilities like the
Hill plot are discussed for example in [9].
Our investigations suggest that the best threshold for the
fitness of GPD is 25, 000 casualties, when using raw data,
well above the 3, 000 minimum we have set in collecting our
observations. A total of 331 armed conflicts lie above this
threshold (58.6% of all our data).
The use of rescaled and dual data requires us to rescale the
threshold; the value thus becomes 145k casualties. It is worth
noticing that, nevertheless, for rescaled data the threshold
could be lowered to 50k, and still we would have a satisfactory
GPD fitting.
If > 1/2, the parameters of a GPD can be easily
estimated using Maximum Likelihood Estimation (MLE) [9],
while for 1/2 MLE estimates are not consistent and
other methods are better used [11].
In our case, given the fat-tailed behavior, is likely to
be positive. This is clearly visible in Figure 10, showing the
Pickands plot, based on the Pickands estimator for , defined
as
1
Xk,n X2k,n
(P )
k,n =
log
,
log 2
X2k,n X4k,n

6
2

Records

10

Plot of Record Development

20

40

60

80

100

120

140

Fig. 10: Pickands estimator for the shape parameter for

war casualties (raw data). The value 1.5 appears to be a good
educated guess.
(P )

In the Pickands plot, k,n is computed for different values

of k, and the "optimal" estimate of is obtained by looking

TAIL RISK WORKING PAPERS

Rescaled Data

145k

Dual Data

145k

0.8
0.6

(2016)

1.5868

497436

(0.1265)

(2097)

1.5915

496668

(0.1268)

(1483)

Fu(xu)

90620

(0.1233)

0.4

1.4985

0.2

Threshold
25k

0.0

Data
Raw Data

1.0

TABLE III: Maximum likelihood estimates (and standard

errors) of the Generalized Pareto Distribution parameters for
casualties over different thresholds, using raw, rescaled and
dual data. We also provide the number of events lying above
the thresholds, the total number of observations being equal
to 565.

10

50

100

500

1000

5000 10000

x (on log scale)

0.050
0.020
0.005

1F(x) (on log scale)

0.200

0.500

(a) Distribution of the excesses (ecfd and theoretical).

0.001

(P )
at a more or less stable value for k,n in the plot. In Figure
10, the value 1.5 seems to be a good educated guess for raw
data, and similar results hold for rescaled and dual amounts.
In any case, the important message is that > 0, therefore we
can safely use MLE.
Table III presents our estimates of and for raw, rescaled
and dual data. In all cases, is significantly greater than 1, and
around 1.5 as we guessed by looking at Figure 10. This means
that in all cases the mean of the distribution of casualties
seems to be infinite, consistently with the descriptive analysis
of Section III.
Further, Table III shows the similarity between rescaled and
dual data. Looking at the standard errors, no test would reject
the null hypothesis that rescaled = dual , or that rescaled =
dual .
Figure 11 shows the goodness of our GPD fit for dual
data, which is also supported by goodness-of-fit tests for the
Generalized Pareto Distribution [42] (p-value: 0.37). As usual,
similar results do hold for raw and rescaled data. Figure 12
shows the residuals of the GPD fit which are, as can be
expected (see [11]), exponentially distributed.

10

50

100

500

1000

5000 10000

x (on log scale)

(b) Right Tail (empirical and theoretical).

Fig. 11: GPD fitting of war casualties exceeding the 145k

threshold, using dual data.

V. E STIMATING THE SHADOW MEAN

Given the finite upper bound H, we known that E[Y ] must
be finite as well. A simple idea is to estimate it by computing
the sample mean, or the conditional mean above a given
threshold9 , if we are more interested in the tail behavior of
war casualties, as in this paper. For a minimum threshold L
equal to 145k, this value is 1.77 107 (remember that Y
represents the rescaled data).
However, in spite of its boundedness, the support of Y is so
large that Y behaves like a very heavy-tailed distribution, at
least until we do not approach the upper bound H. This makes
the (conditional) sample mean not really reliable, or robust:
one single extreme observation can make it jump, given that
the (conditional) sample mean has a breakdown point equal to
0 [23].
A more robust way of computing what we define the
conditional shadow mean of Y (the true conditional mean not
9 Notice

that the sample mean above a minimum threshold L corresponds

to the concept of expected shortfall used in risk management.

visible from data) is therefore to use the log-transformation of

Equation (1).
Let F and G be respectively the distribution of Y and Z.
With f and g, we indicate the densities.
In the previous section we have seen that, when using the
dual version of rescaled data, G GPD(, ) for Z
L=
145000 (even 50k gives good results, but 145k is comparable
to the 25k threshold of raw data). Moreover we know that
L Z
Z = '(Y ), so that Y = ' 1 (Z) = (L H)e H + H.
This implies that the distribution of Y , for Y
obtained from the distribution of G.
First we have
Z 1
L

g(z) dz =

'

(1)

f (y) dy.
L

L, can be

(5)

TAIL RISK WORKING PAPERS

10

their ratio for different minimum thresholds L, using rescaled

data. As already said, our GPD fit is already satisfactory for
L = 50000, that is why we start from that threshold.

Residuals

The ratios in Table IV show how the sample mean underestimates the shadow mean, especially for lower thresholds,
for which the shadow mean is almost 3 times larger than the
sample mean. For these reasons, a "journalistic" reliance on
sample mean would be highly misleading, when considering
all conflict together, without setting large thresholds. There
would be the serious risk of underestimating the real lethality
of armed conflicts.

TABLE IV: Conditional shadow mean, conditional sample

mean and their ratio for different minimum thresholds. In bold
the values for the 145k threshold. Rescaled data.

50

100

150

200

250

300

Ordering

(a) Scatter plot of residuals and their smoothed average.

Sample107
1.2753
1.5171
1.7710
2.2639
2.8776
3.9481

Ratio
2.88
2.43
2.08
1.95
1.96
1.85

Shadow107
3.6790
3.6840
3.6885
4.4089
5.6542
6.5511

A similar procedure can be applied to raw data. Fixing H

equal to 7.2 billion and L = 25000, we can define V = '(X)
and study its tail. The tail coefficient V is 1.4970. Even in
this case the sample mean underestimates the shadow mean
by a factor of at least 1.6 (up to a maximum of 7.5!).

Exponential Quantiles

Thresh.103
50
100
145
300
500
1000

Ordered Data

(b) QQ-plot of residuals against exponential quantiles.

Fig. 12: Residuals of the GPD fitting: looking for exponentiality.

And we know that

g(z; , ) =

1+

z 2 [L, 1).

This implies, setting = 1 and k = /H,

1
1 log(H y)klog(H L)
f (y; , k) =
,
(H y)k

(6)

y 2 [L, H].
(7)

We can then derive the shadow mean of Y as

Z H
E[Y ] =
y f (y) dy,

(8)

obtaining

E[Y ] = L + (H

L)ek (k) (1

, k) .

(9)

Table IV provides the shadow mean, the sample mean and

These results tell us something already underlined by [12]:

we should not underestimate the risk associated to armed
conflicts, especially today. While data do not support any
significant reduction in human belligerence, we now have both
the connectivity and the technologies to annihilate the entire
world population (finally touching that value H we have not
touched so far). In the antiquity, it was highly improbable for
a conflict involving the Romans to spread all over the world,
thus also affecting the populations of the Americas. A conflict
in ancient Italy could surely influence the populations in Gaul,
but it could have no impact on people living in Australia at
that time. Now, a conflict in the Middle East could in principle
start the next World War, and no one could feel completely
safe.
It is important to stress that changes in the value of H do not
effect the conclusions we can draw from data, when variations
do not modify the order of magnitude of H. In other terms,
the estimates we get for H = 7.2 billion are qualitatively
consistent to what we get for all values of H in the range
[6, 9] billion. For example, for H = 9 billion, the conditional
shadow mean of rescaled data, for L = 145000, becomes
4.0384107 . This value is surely larger than 3.6885107 , the
one we find in Table IV, but it is still in the vicinity of 4107 ,
so that, from a qualitative point of view, our conclusions are
still completely valid.

TAIL RISK WORKING PAPERS

11

1.0

6000

(a) Raw data.

5000
4000
Frequency

2000

Ratio
2.8444
2.8843
3.0422
2.9843
3.0412
3.1332

2.5

1000

Sample108
0.8574
0.9338
1.0120
1.1363
1.2774
1.4885

2.0

Shadow108
2.4388
2.6934
3.0788
3.3911
3.8848
4.6639

1.5

TABLE V: Shadow standard deviation, sample standard deviation and their ratio for different minimum thresholds. Rescaled
data.
Thresh.103
50
100
145
300
500
1000

1500
0

500

1000

Frequency

2000

2500

3000

(by selecting across the various estimates shown in Table

I) the effect does not go beyond the smaller decimals.
We can conclude that our estimates are remarkably robust to
imprecisions and missing values in the data10 .

3000

Other shadow moments

As we did for the mean, we can use the transformed distribution to compute all moments, given that, since the support
of Y is finite, all moments must be finite. However, because of
the wide range of variation of Y , and its heavy-tailed behavior,
one needs to be careful in the interpretation of these moments.
As discussed in [38], the simple standard deviation (or the
variance) becomes unreliable when the support of Y allows
for extremely fat tails (and one should prefer other measures of
variability, like the mean absolute deviation). It goes without
saying that for higher moments it is even worse.
Deriving these moments can be cumbersome, both analytically and numerically, but it is nevertheless possible. For
example, if we are interested in the standard deviation of Y
for L = 145000, we get 3.08 108 . This value is very large,
three times the sample standard deviation, but finite. Other
values are shown in Table V. In all cases, the shadow standard
deviation is larger than the sample ones, as expected when
dealing with fat-tails [13]. As we stressed that care must be
used in interpreting these quantities.
Note that our dual approach can be generalized to all
those cases in which an event can manifest very extreme
values, without going to infinity because of some physical or
economical constraint.

VI. D EALING WITH MISSING AND IMPRECISE DATA

An effective way of checking the robustness of our estimates
to the "quality and reliability" of data is to use resampling
techniques, which allow us to deal with non-precise [43]
and possibly missing data. We have performed three different
experiments:
Using the jackknife, we have created 100k samples, by
randomly removing up to 10% of the observations lying
above the 25k thresholds. In more than 99% of cases
>> 1. As can be expected, the shape parameter goes
below the value 1 only if most of the observations we
remove belong to the upper tail, like WW1 or WW2.
Using the bootstrap with replacement, we have generated
another set of 100k samples. As visible in Figure 13,
1 in a tiny minority of cases, less than 0.5%. All other
estimates nicely distribute around the values of Table III.
This is true for raw, rescaled and dual data.
We tested for the effect of changes in such large events as
WW2 owing to the inconsistencies we mentioned earlier

1.0

1.5

2.0

2.5

3.0

3.5

4.0

(b) Rescaled data.

Fig. 13: Distribution of the estimate (thresholds 25k and

145k casualties for raw and rescaled data respectively) according to 100k bootstrap samples.

VII. F REQUENCY OF ARMED CONFLICTS

Can we say something about the frequency of armed conflicts over time? Can we observe some trend? In this section,
we show that our data tend to support the findings of [18]
and [32], contra [24], [30], that armed conflict are likely to
follow a homogeneous Poisson process, especially if we focus
on events with large casualties.
10 Thats why, when dealing with fat tails, one should always prefer to
other other quantities like the sample mean [38]. And also use in corrections,
as we propose here.

TAIL RISK WORKING PAPERS

11 When considering inter-arrival times, we deal with integers, as we only

record the year of each event.

4
3
0

Exponential Quantiles

have to admit.

10

Ordered Data

ACF

0.2

0.4

0.6

0.8

1.0

(a) Exponential QQ-plot of gaps (notice that gaps are expressed in

years, so that they are discrete quantities).

0.0

The good GPD approximation allows us to use a wellknown model of extreme value theory: the Peaks-overThreshold, or POT. According to this approach, excesses over
a high threshold follow a Generalized Pareto distribution, as
stated by the Pickands, Balkema and de Haans Theorem
[1], [28], and the number of excesses over time follows
a homogeneous Poisson process [13]. If the last statement
were verified for large armed conflicts, it would mean that
no particular trend can be observed, i.e. that the propensity
of humanity to generate big wars has neither decreased nor
increased over time.
In order to avoid problems with the armed conflicts of
antiquity and possible missing data, we here restrict our
attention on all events who took place after 1500 AD, i.e.
in the last 515 years. As we have shown in Section IV,
missing data are unlikely to influence our estimates of the
shape parameter , but they surely have an impact on the
number of observations in a given period. We do not want
to state that we live in a more violent era, simply because we
miss observations from the past.
If large events, those above the 25k threshold for raw data
(or the 145k one for rescaled amounts), follow a homogeneous
Poisson process, in the period 1500-2015AD, their inter-arrival
times need to be exponentially distributed. Moreover, no time
dependence should be identified among inter-arrival times, for
example when plotting an autocorrelogram (ACF).
Figure 14 shows that both of these characteristics are satisfactorily observable in our data set.11 This is clearly visible in
the QQ-plot of Subfigure 14a, where most inter-arrival times
tend to cluster on the diagonal.
Another way to test if large armed conflicts follow a
homogeneous Poisson process is by considering the number
of events in non-overlapping intervals. For a Poisson process,
given a certain number of events in a time interval, the
numbers of events in non-overlapping subintervals follow a
Multinomial distribution. If the Poisson process is homogeneous, then the Multinomial distribution is characterized by
the same probability of falling in any of the sub-intervals, that
is an equiprobability. It is not difficult to verify this with our
data, and to see that we cannot reject the null hypothesis of
equiprobable Multinomial distribution, over the period 15002015, in which 504 events took place, choosing a confidence
level of 10%.
Once again, the homogeneous Poisson behavior is verified
for raw, rescaled and dual data. Regarding the estimates of ,
if we restrict our attention on the period 1500-2015, we find
1.4093 for raw data, and 1.4653 for rescaled amounts.
To conclude this section, no particular trend in the number
of armed conflicts can be traced. We are not saying that this
is not possible over smaller time windows, but at least in the
last 500 years humanity has shown to be as violent as usual.
To conclude our paper, one may perhaps produce a convincing theory about better, more peaceful days ahead, but
this cannot be stated on the basis of statistical analysis this
is not what the data allows us to say. Not very good news, we

12

10

15

20

Lag

(b) ACF of gaps. Notice that the first lag has order 0.

Fig. 14: Two plots to verify whether armed conflicts (raw

data) over the 25k casualties threshold follow a homogeneous
Poisson Process

ACKOWLEDGEMENTS
Captain Mark Weisenborn engaged in the thankless and
gruesome task of initially compiling the data, checking across
sources and linking each conflict to a narrative. We also
benefited from generous help on social networks where we
put data for scrutiny, as well as advice from historians such
as Ben Kiernan.
We thank Raphael Douady, Yaneer Bar Yam, Joe Norman,
Alex(Sandy) Pentland and his lab for discussions and comments.
Pasquale Cirillo acknowledges the support of his Marie
Curie CIG Multivariate Shocks" (PCIG13-GA-2013-618794).

TAIL RISK WORKING PAPERS

R EFERENCES
[1] A. A. Balkema, L. De Haan (1974). Residual life time at great age.
Annals of Probability 5, 792-804.
[2] A. A. Balkema, L. De Haan (1978). Limit distributions for order
statistics I. Theory of Probability & Its Applications 23:1, 77-92.
[3] A. A. Balkema, L. De Haan (1979). Limit distributions for order
statistics II. Theory of Probability & Its Applications 23:2, 341-358.
[4] BBC (2012). In our times: The An Lushan Rebellion. http://www.bbc.
co.uk/programmes/b01by8ms.
[5] D. Berlinski (2009). The Devils Delusion: Atheism and its Scientific
Pretensions. Basic Books.
[6] L.-E. Cederman (2003). Modeling the size of wars: from billiard balls
to sandpiles. The American Political Science Review 97, 135-150.
[7] P. Cirillo (2013). Are your data really Pareto distributed? Physica A:
Statistical Mechanics and its Applications 392, 5947-5962.
[8] A. Clauset, R. Woodard (2013). Estimating the historical and future
probabilities of large terrorist events. Annals of Applied Statistics 7,
1838-1865.
[9] L. de Haan, A. Ferreira (2006). Extreme Value Theory: An Introduction.
Springer.
[10] U. Einmahl, D. Li (2005). Some results on two-sided LIL behavior.
Annals of Probability 33, 1601-1624.
[11] P. Embrechts, C. Klppelberg, T. Mikosch (2003). Modelling Extremal
Events. Springer.
[12] R. Epstein (2011). Book Review: The Better Angels of Our Nature:
Why Violence Has Declined. Scientific American, October 7.
[13] M. Falk, J. Hsler J, R. D. Reiss R-D (2004). Laws of small numbers:
extremes and rare events, Birkhuser.
[14] J. A. Friedman (2014). Using power laws to estimate conflict size.
Journal of Conflict Resolution, 1-26.
[15] J.L. Goldstein (2011). Winning the war on war: the decline of armed
conflict worldwide. Penguin.
[16] J. Gray (2015). Steven Pinker is wrong about violence and war. The
Guardian, Friday 13 March. http://www.theguardian.com/books/2015/
mar/13/john-_gray-_steven-_pinker-_wrong-_violence-_war-_declining
[17] E. J. Gumbel (1958). Statistics of Extremes. Cambridge University Press.
[18] B. Hayek (2002). Statistics of Deadly Quarrels. American Scientist 90,
10-14.
[19] A. Horne (2006). A Savage War of Peace. New York Review Books
Classics
[20] N. Inkester, ed. (2015). Armed Conflict Survey 2015. IISS.
[21] C. Kleiber, S. Kotz (2003). Statistical Size Distribution in Economics
and Actuarial Sciences. Wiley.
[22] K. Klein Goldewijk, G. van Drecht (2006). HYDE 3.1: Current and
historical population and land cover. In A. F. Bouwman, T. Kram, K.
Klein Goldewijk. Integrated modelling of global environmental change.
An overview of IMAGE 2.4. Netherlands Environmental Assessment
Agency.
[23] R. Maronna, R.D. Martin, V. Yohai (2006). Robust Statistics - Theory
and Methods. Wiley.
[24] J. E. Mueller (1989). Retreat from Doomsday: the obsolescence of major
war. Basic Books.
[25] Necrometrics (2015). http://necrometrics.com.
[26] J. Nelehov, P. Embrechts, V. Chavez-Demoulin (2006). Infinite mean
models and the LDA for operational risk. Journal of Operational Risk
1, 3-25.
[27] R. Norton-Taylor (2015). Global armed conflicts becoming more
deadly, major study finds. The Guardian, Wednesday 20 May.
http://www.theguardian.com/world/2015/may/20/armed-_conflict-_
deaths-_increase-_syria-_iraq-_afghanistan-_yemen
[28] J. Pickands (1975). Statistical inference using extreme order statistics.
the Annals of Statistics, 119-131.
[29] C. Phillips, A. Axelrod (2004). Encyclopedia of Wars, 3-Volume Set.
Infobase Pub.
[30] S. Pinker (2011). The Better Angels of Our Nature: Why Violence Has
Declined. Viking Adult.
[31] L. F. Richardson (1948). Variation of the Frequency of Fatal Quarrels
with Magnitude. Journal of the American Statistical Association 43,
523-46.
[32] L. F. Richardson (1960). Statistics of Deadly Quarrels. Chicago Quadrangle Book.
[33] L. A. Cox, M. R. Greenberg, eds. (2011). Special Issue: Advances in
Terrorism Risk Analysis. Risk Analysis, online ISSN 1539-6924.
[34] C.Robert, G. Casella (2005). Monte Carlo Statistical Methods. Springer.

13

[35] A. Scharpf, G. Schneider, A. Noh, A. Clauset (2014). Forecasting of the

risk of extreme massacres in Syria. European Review of International
Studies 1, 50-68.
[36] T. B. Seybolt, J. D. Aronson, B. Fischhoff, eds. (2013). Counting
Civilian Casualties, An Introduction to Recording and Estimating Nonmilitary Deaths in Conflict. Oxford University Press.
[37] M. Spagat, A. Mack, T. Cooper, J. Kreutz (2009). Estimating War
Deaths: An Arena of Contestation. Journal of Conflict Resolution 53,
934-50.
[38] N. N. Taleb (2015). Silent Risk: Lectures on Probability, Vol 1. SSRN.
[39] United Nations - Department of Economic and Social Affairs (1999).
The World at Six Billion. UN Press.
[40] United Nations - Department of Economic and Social Affairs (2015).
2015 Revision of World Population Prospects. UN Press.
[41] J. van der Wijk (1939). Inkomens- en Vermogensverdeling. Publication
Nederlandsch Economisch Instituut 26.
of thOe
[42] J.-A. Villasenor-Alva, E. Gonzalez-Estrada (2009). A bootstrap goodness
of fit test for the generalized Pareto distribution. Computational Statistics
and Data Analysis 53, 3835-3841.
[43] R. Viertl (1995). Statistical Methods for Non-Precise Data. CRC Press.
[44] P. Wallensteen, M. Sollenberg (2001). Armed Conflict 1989-2000.
Journal of Peace Research 38, 629-644.
[45] M. White (2011). The Great Big Book of Horrible Things. W.W. Norton
and Company.
[46] M. White (2013). Atrocities. W.W. Norton and Company.