Point Processes, and Populations Dynamics
in longevity, credit, HFT modelling
Nicole El Karoui, Alexandre Boumezoued (PhD)
UPMC(Paris VI), LPMAProbability and Random Models Laboratory, UMR-CNRS 7599
Santa Barbara, Septembre 2014
1/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Plan
1 Some Dates with Jean Pierre
2 Motivation to model global population
3 First look at longevity
4 Role of age
5 Macroscopic approximation
2/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Do you know AMART?
You, probably no, but
Jean Pierre, Yes !, and me also, in principle...as his supervisor
3/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
From Amarts to Finance, a long way
• In France, Opening of the First financial Market in 1988
• The first Master program in Mathematical Finance is created
jointly PVI -Ecole Polytechnique in 1990.
6/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Plan
1 Some Dates with Jean Pierre
2 Motivation to model global population
3 First look at longevity
4 Role of age
5 Macroscopic approximation
7/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Why population dynamics ?
Marked Point Process: Renew of interest
Useful tool, under different denominations, for other domains:
I Credit risk Modelling
I Hawkes processes in HFT
I Particular methods in Simulation
I Imaging
I Brain study
I Data Mining
Population dynamics Old problems, Malthus 1798, Verhulst 1838
I First motivation, longevity risk in insurance and finance (
pension fund)
I General Model for population dynamics in Ecology
I By introducing complexity and heterogeite in the population
I The age parameter plays strategic role
I Useful for
Aim: By mixing these different points of view, and starting of
micro evolution at the individual level and demographic patterns in
birth and death to explain macro evolution of the general structure.
8/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
A lot of different denominations
I Agent Based Models (ABM) in Computational Economy
I Microsimulation models of National Agencies (Destinee in
France)
I Demography computational
I Age Period Cohort Analysis
I Interacting particle methods for simulation
Our reference: Individual-Based models in ecology:
I Population structured by traits (i.e. individual characteristics)
(Fournier-Meleard 2004, Champagnat-Ferrire-Mlard 2006)
I Extension to age-structured populations (Tran 2006,
Ferrire-Tran 2009)
9/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Plan
1 Some Dates with Jean Pierre
2 Motivation to model global population
3 First look at longevity
4 Role of age
5 Macroscopic approximation
10/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
First look at Longevity
Developed countries are experiencing a reduction in mortality over
time
I Aging populations: new phenomenon, without past historical
reference
I Societies are facing new different challenges:
new generational equilibrium
role and place of aging population in the society
viability of shared collective systems, in particular (state or
private) pension systems
11/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Aggregate mortality indicators
I Lifetime of an individual: τ
I Life expectancy at birth: E[τ ]
I Death rate d(a) such that P(τ > a) = e−∫ a
0 d(s)s.
I Possible quantity of interest: annual death probability
q(a) = P(τ < a + 1 | τ ≥ a)
I In practice, reduction of mortality over time:
⇒ q(a, t)=probability for an individual aged a at the
beginning of year t to die in the following year.
12/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Mortality tables in Insurance
I q(a, t)= probability for an individual aged a at the beginning
of year t to die in the following year.
I Lexis diagram ”age - period - cohort” :
can be taken into consideration, it proves di�cult to account for censured and truncated data when
assessing national mortality. One may think for instance of migration. There are two di↵erent factors
that can be overlooked when considering migratory impacts. First of all, people leaving the synthetic
cohort are often censored, and thus still taken into account when estimating the period life tables.
Secondly, incoming immigrants can also a↵ect and significantly change the local national mortality
data of the destination country.
Far more relevant details regarding deaths can be gained from within an insured portfolio (e.g. the
cause of death). Although national data by cause of death does exist, it generally lacks consistency,
and is often useless, or of no specific interest in deriving mortality by cause (for example: in the
United States, 11% of deaths are caused by more than four diseases).
Coh
ort
Year
Age
66
65
64
63
200819431942
1
Figure 1.1: Lexis diagram: Age-year-cohort diagram.
1.2. Heterogeneity, inter-age dependence and basis risk
It is expected that any given population will display some degree of heterogeneous mortality.
Heterogeneity often arises due to a number of observable factors, which include age, gender, occupation
and physiological factors. As far as longevity risk is concerned, policyholders that are of higher socio-
economic status (assessed by occupation, income or education) tend to experience lower levels of
mortality. However, significant di↵erences also exist within the same socio-economic levels according
to gender. Females tend to outlive males and have lower mortality rates at all ages. In addition, some
heterogeneity arises due to features of the living environment, such as: climate, pollution, nutritional
standards, population density and sanitation (see Section 2.2 for a more detailed discussion).
When considering insured portfolios, insurers tend to impose selective criteria that limits contractual
access to those considered to possess no explosive risk (by level of health and medical history). This
4
13/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Age Pyramid in France
Initial population for N=10 000 in 2008
Age pyramid in 2008 Age pyramid in 2008
100 50 0 50 100
07
1524
3342
5160
6978
8796
107
119
Number of males Number of females
Age
Source: The French National Institute for Statistics and Economic Studies (INSEE
14/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Human Mortality Database
Human Mortality Database. University of California, Berkeley
(USA), and Max Planck Institute for Demographic Research
(Germany).
Available at www.mortality.org or www.humanmortality.de
15/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Mortality structure (France)
0
20
40
6080 1900
1950
2000
0.00
0.05
0.10
0.15
0.20
probabilités de décès (FR)
16/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
National mortality by gender (France)
0
20
40
6080 1900
1950
2000
0.00
0.05
0.10
0.15
0.20
probabilités de décès (femmes, FR)
0
20
40
6080 1900
1950
2000
0.00
0.05
0.10
0.15
0.20
probabilités de décès (hommes, FR)
Femmes Hommes
17/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
National mortality: log q(a,t)
I Looking at log q(a, t) age a in [0,100]
I for different years t (1950,1965,1980,1995,2005)
0 20 40 60 80 100
-10
-8-6
-4-2
0
age a
log(
q(. ,
t,))
19501965198019952005
Figure: Logarithm of annual death probabilities (national population)18/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Aims of microscopic models
I Provide population evolution at the scale of the individual
I Allows to understand patterns of aggregate indicators
Two examples in this talk
1 Impact of aging
2 How individual birth patterns in heterogenous populations can
create artificial mortality changes (”cohort effect”) ?
20/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Individual-based model
Human population extension: Bensusan 2010 (Phd thesis),
Boumezoued & El Karoui (working paper)
I Possibility for an individual to evolve during life (marriage,
divorce, professional evolution,...)
I The population is subject to a stochastic environment
(Yt)t≥0, evolving through time
This environment refers to random changes increasing or
decreasing mortality
Natural catastrophes, pandemics, wars, emerging diseases,...
Medical environment
21/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Individual-based model
Demographic rates: an individual of traits xt ∈ X ⊂ Rd and age
at ∈ [0, a] at time t, (born at time 0)
I Dies at rate d (xt , at , t,Y )
P(Tdeath ≥ t | Y ) = exp(−∫ t
0d(xs , as , s,Y ) ds
)I Gives birth at rate b (xt , at , t,Y )
and the new individual has traits x ′ ∼ Kb(xt , at , x.′)
I Evolves during life at rate e (xt , at , t,Y )
from traits xt to x ′ ∼ K e(xt , at , dx ′)
I Demographic rates depend on characteristics, age, time and
on the stochastic environment Y
I Conditionally on the environment Y , the events for a given
individual are jumps of a counting process
22/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Non homogenous process simulation
Simulation by Thinning of a counting process with rate λt ,
assuming that λt ≤ λ
I Let (Tn) the jump times of a Poisson process Nt with rate λ:
Tn+1 − Tn ∼ Exp(λ)
I Recursively, at time Tn pick a Bernouilli independent r.v. Un
s.t. P(Un = 1) = λTn/λ.
I Then Nt := cardinal{k : Uk = 1, Tk ≤ t} is a counting
process with rate λt
I (Tn) are interpreted as inspection times for the system
I The thinning method makes easier the simulation of counting
processes with complex rates
23/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Simulation algorithm
I Assumption of bounded demographic rates with b, d and e
⇒ use of the Thinning method (∼ Inspection times)
1 Start with N indiv. at T , generate τ ∼ Exp(N(d + b + e)
)The bigger the population, the more it is inspected
2 Select an individual (x I , aI ) uniformly and compute:
p1 = b(x I ,aI ,T+τ,Y )
b+d+e, p2 = d(x I ,aI ,T+τ,Y )
b+d+e, p3 = e(x I ,aI ,T+τ,Y )
b+d+e
Only one individual is checked (not exhaustive)
3 Determine the nature of inspection at time T + τ
I Birth: add a new individual with probability p1
I Death: remove (x I , aI ) with probability p2
I Evolution: change traits of (x I , aI ) with probability p3
I No event with probability 1− p1 − p2 − p3
24/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Measure-valued stochastic process
I At time t, the population is described by the following point
measure on X × [0, a]:
Zt(dx , da) =∑Nt
i=1 δ(x it ,ait)
(dx , da).
A way to describe the boxes, and their evolution
I Computation of various quantities of interest is made by
testing some functions (x , a) 7→ ft(x , a):
〈Zt , ft〉 =∫X×[0,a] ft(x , a)Zt(dx , da) =
∑Nti=1 ft(x i
t , ait).
I Example: population size Nt = 〈Zt , 1〉
I Role of the composition of the initial population at time 0
25/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Plan
1 Some Dates with Jean Pierre
2 Motivation to model global population
3 First look at longevity
4 Role of age
5 Macroscopic approximation
26/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Role of age: a simple example from Alex.B.
Pure birth process with age structure
I each individual has the same birth rate b(a) (d=e=0)
I λt =∑
i∈Zt−b(ai ) = 〈Zt−, b〉
I recursive evolution of the box : individuals in Zt− are all
individual born up to t−
Self-exciting process (Nt) = (〈Zt , 1〉) is a counting process with
intensity
λt = N0b(t) +
∫(0,t)
b(t − s) dNs = N0b(t) +∑Ti<t
b(t − Ti ),
where the (Ti ) are the times of jump of N.
I This leads to an equation on Nt : how to construct it ?
27/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Two examples
Birth rates Hawkes: b(a) = exp−a, Human b(a) = exp−(a− c)2/2
28/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Representation of pure birth process:
Hawkes process
I Let Q(ds, dθ) be a Poisson point measure with intensity
measure ds dθ on R+ × R+
I need the whole population Zt(da) =∑Nt
i=1 δAi (Zt) counting
the birth but also taking track of the ages distribution.
I Representation of the birth process by Thinning:
Zt(da) =
N0∑i=1
δt(da) +
∫ t
0
∫R+
1θ≤〈Zs−,b〉δt−s(da)Q(ds, dθ)
I Note that Zt(da) is Markov but Nt = 〈Zt , 1〉 is not.
29/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Number of events
Hawkes (left) Human (right)
30/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Cohort effect
Birth cohort for the period [t1, t2]: group of individuals born
between t1 and t2.
I Individuals of the same birth cohort share similar demographic
characteristics (”cohort effect”)
I Age, Period, Cohort analysis put a lot of problems in practice,
in different domains, medecine, sociology,...due to the lag in
data,..insurance...
I Huge literature on APC problems
33/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Golden cohort
Golden cohort: generations born between 1925 and 1945Cairns et al. (2009) ra,t = (qa,t−1 − qa,t)/qa,t
The Golden cohort has experienced more rapid improvements than
earlier and later generations.
2 DATA 7
1970 1980 1990 2000
2040
6080
Year
Age
−2%
−1%
0%
1%
2%
3%
4%
Annu
al im
prov
emen
t rat
e (%
)
Figure 3: Improvement rates in mortality for England & Wales by calendar yearand age relative to mortality rates at the same age in the previous year. Red cellsimply that mortality is deteriorating; green small rates of improvement, and blueand white strong rates of improvement. The black diagonal line follows the progressof the 1930 cohort.
2.1.2 The cohort eÆect
Some of the models we employ incorporate what is commonly called the “cohorteÆect”. The rationale for its incorporation lies in an analysis of the rates at whichmortality has been improving at diÆerent ages and in diÆerent years. Rates ofimprovement are plotted in Figure 3 (see, also Willets, 2004, and Richards et al.,2006). A black and white version of this graph can be found in the Appendix, Figure38.
In line with previous authors (see, for example, Willets, 2004, Richards et al., 2006)we can note the following points. In certain sections of the plot, we can detectstrong diagonals of similar colours. Most obviously, cohorts born around 1930 havestrong rates of improvement between ages 40 and 70 relative to, say, cohorts born10 years earlier or 10 years later. The cohort born around 1950 seems to have worsemortality than the immediately preceeding cohorts.
There are other ways to illustrate the cohort eÆect and these can be found in Ap-pendix A.
34/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Analysis of R. C. Willets, 2004
Some possible explanations:
I Impact of World War II on previous generations,
I Changes on smoking prevalence: tobacco consumption in next
generations,
I Impact of diet in early life,
I Post World War II welfare state,
I Patterns of birth rates
35/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Cohort effect (UK)
”One possible consequence of rapidly changing birth rates is that
the average child is likely to be different in periods where birth
rates are very different. For instance, if trends in fertility vary by
socio-economic class, the class mix of a population will change.”
The Cohort Effect: Insights And Explanations, 2004, R. C. Willets
37/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Simple toy model
I Reference death rate d(a) = A exp(Ba)
I Parameters A ∼ 0.0004 and B ∼ 0.073 estimated on French
national data for year 1925 to capture a proper order of
magnitude
I ”Upper class”: time independent death rate d1(a) = d(a) and
birth rate b1(a) = c1[20,40](a) (c=0.1)
I ”Lower class”: time independent death rate d2(a) = 2d(a)
but birth rate
b2(a, t) = 4c1[20,40](a)1[0,t1]∪[t3,∞)(t) + 2c1[20,40](a)1[t2,t3](t)
Constant death rates but reduction in overall fertility between
times t1 (=10) and t2 (=20)
I Aim: compute standard demographic indicators
38/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Aggregate fertility
I One trajectory with 20000 individuals (randomly) splitted
between groups. Estimation of aggregate fertility
0 5 10 15 20 25 30
0.0
0.1
0.2
0.3
0.4
year
num
ber o
f chi
ldre
n pe
r ind
ivid
ual a
ged
[20,
40]
39/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Life expectancy by year of birth
I ”Cohort effect” for aggregate life expectancy
0 5 10 15 20 25 30
5556
5758
5960
year of birth
life
expe
ctan
cy (a
t birt
h)
40/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
0 5 10 15 20 25 30
0.0
0.1
0.2
0.3
0.4
year
num
ber o
f chi
ldre
n pe
r ind
ivid
ual a
ged
[20,
40]
0 5 10 15 20 25 30
5556
5758
5960
year of birth
life
expe
ctan
cy (a
t birt
h)
Figure: Observed fertility (left) and estimated life expectancy by year of
birth (right)
I Death rates by specific group remain the same
I But reduction in fertility for ”lower class” during 10-20
modifies the generations composition
⇒ ”upper class” is more represented among those born
between 10 and 20
Wealth and longevity
The role of heterogenity
I by birth or emigration
I by evolution (swap-mutation) often faster than the
demographic events
I by the environment
Identification problem
I In spatial Birth Death Point process (Huber 2010), swapping
is used to accelerate the convergence, and shortening the
mixing time of the chain.
42/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Wealth and longevity: complex dependence
43/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Plan
1 Some Dates with Jean Pierre
2 Motivation to model global population
3 First look at longevity
4 Role of age
5 Macroscopic approximation
44/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Martingale problem
(No evolution for simplicity of notations)
Conditionally on the complete path of Y , in conditionally
independency framewor
Mt(f ) =< Zt , ft > − < Z0, f0 > −∫ t
0
ds
∫Zs(dx , da)
[(∂fs∂a
+∂fs∂s
)(x , a)
+b(x , a, s,Y )
∫χ
fs(x ′, 0)K b(x , a, dx ′)− d(x , a, s,Y )fs(x , a)]
(1)
is a square integrable martingale with quadratic variation
< M(f ) >t=
∫ t
0
ds
∫χ×[0,a]
Zs(dx , da)×
[b(x , a, s,Y )
∫χ
f 2s (x ′, 0)K b(x , a, dx ′) + d(x , a, s,Y )f 2
s (x , a)]
(2)(3)
46/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Macroscopic approximation
Assumptions
I Renormalization: pop. described by the measure
Znt (dx , da) = 1
n
∑Nnt
i=1 δ(x it ,ait)
(each individual has weight 1/n)
I Weak convergence of the initial population as n→ +∞:
Zn0 (dx , da)⇒ g0(x , a)γ(dx)da (initial size is of order n)
I By homogeneity, the quadratic variation of Mn(f ) is of order1n and so goes to 0
I Cv in distribution on the canonical space of cadlag measure
valued process of the process (Znt ))
47/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Conditionnaly to Y , limit PDE
I Law of Large Numbers (Heuristic): the noise vanishes as the
size n of the initial population goes to infinity
⇒ deterministic behavior in time for large populations
Limit PDE
I Limit process as the size → +∞I Weak convergence of (Zn
t (dx , da))t≥0 to the solution(gt(x , a)γ(dx , da
)t≥0
of conditional (wrt Y) deterministic
PDE
Link between two description in a given environment:
I microscopic: stochastic behavior of each individual
I macroscopic: deterministic evolution of the whole population
48/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Deterministic equations in demography
I Malthus (1798), Verhulst (1838): pop. structured by traits
g(x , t): density of individuals of trait x at time t
∂g
∂t(x , t) =
∫χ
g(y , t)b(y)kb(y , x)dy − d(x)g(x , t),
g(x , 0) = g0(x).
I McKendrick (1926), VonFoerster (1959): structured by age
g(a, t): density of individuals of age a at time t
∂g
∂t(a, t) +
∂g
∂a(a, t)︸ ︷︷ ︸
transport
= −d(a)g(a, t), g(0, t) =
∫ +∞
0b(a)g(a, t)da︸ ︷︷ ︸
renewal
g(a, 0) = g0(a).
49/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Large population limit
PDE for the population density g(x , a, t): approximation (a.s) for
large populations with
I stochastic environment Y
I evolution during life(∂g
∂t+∂g
∂a
)(x , a, t) = −d
(x , a, t,Y
)g(x , a, t)
−e(x , a, t,Y
)g(x , a, t) +
∫χ
e(x′, a, t,Y )ke(x
′, a, x)g(x
′, a, t)γ(dx
′)
g(x , 0, t) =
∫X×[0,a]
b(x
′, a, t,Y
)kb(x
′, a, x)g(x
′, a, t)γ(dx
′)da
g(x , a, 0) = g0(x , a)
I Take advantage of the impact of pure environment noise
50/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014
Conclusion and Perspectives
I Empirical aggregate mortality is the result of complex
demographic mechanisms
I Intrinsic mortality and fertility (+evolution) for different
characteristics leads to a dynamic heterogeneity: the
composition of the population changes over time
In progress or further research:
I Test many assumptions on demographic patterns
I Calibration and estimation of the intensities, and identifcation
isssues
I Improve numerical efficiency, based on competences of the
other actors in these aera
Conclusion: Tool for hypotheses testing for a better understanding
of demographic patterns
51/52 Nicole El Karoui, In honor of J.P.Fouque Santa Barbara 2014