Continuous-Time Pension-Fund

Modelling
Andrew J.G. Cairns 1 2
Department of Actuarial Mathematics and Statistics,
Heriot-Watt University,
Riccarton, Edinburgh, EH14 4AS,
United Kingdom
This paper considers stochastic pension fund models which evolve
in continuous time and with continuous adjustments to the con-
tribution rate and to the asset mix. A generalization of constant
proportion portfolio insurance is considered and an analytical solu-
tion is derived for the stationary distribution of the funding level.
In the case where a risk-free asset exists this is a translated-inverse-
gamma distribution.
Numerical examples show that the continuous-time model gives a
very good approximation to more widely used discrete time models,
with, say, annual contribution rate reviews, and using a variety of
models for stochastic investment returns.

1 E-mail: [email protected]
2 WWW: http://www.ma.hw.ac.uk/ andrewc/
1 Introduction
In this paper we consider continuous time stochastic models for
pension fund dynamics. The general form of this simple model is:

dXt = Xt d(t Xt) + (N + D(Xt ) ; B )dt

where Xt = funding level at time t
= Assets/Liabilities at time t
d(t Xt) = real return between t and t + dt
over salary growth
N = normal contribution rate
D(t Xt) = adjustment to the contribution rate
for surplus or de cit
and B = rate of bene t outgo (as a proportion
of the actuarial liability)
(Note that the description of the model given here allows for the dis-
tribution of investment returns to depend upon the funding level.)
Here, it is assumed that the level of bene t outgo is constant
through time relative to the actuarial liability.
Related to the funding level is the target funding level, L, which will
normally be equal to 1 but this need not be the case. This reserve
is related to the normal contribution rate, the level of bene t outgo
and the valuation rate of interest in excess of salary growth, v , in
the following way:
dL =  L + N ; B = 0
dt v
That is, if the experience of the fund is precisely as expected then
interest on the fund plus the normal contribution rate will be pre-
cisely sucient to pay the bene ts. Thus B ; N = v L.
Similar continuous time models have been considered by Dufresne
(1990). A discrete time version version of the model has been con-
sidered in more detail and in various forms by Cairns and Parker
(1996), Dufresne (1988, 1989, 1990) and Haberman (1992, 1994).
This paper will discuss various special cases of the model. The
rst case is where d(t Xt) does not depend upon Xt and, in eect,
reects a static investment policy with independent and identically
distributed returns. This case has previously been considered by
Dufresne (1990) who showed that the stationary distribution of the
fund size was Inverse Gaussian and here we verify his result using
dierent techniques.
The second case will consider Continuous Proportion Portfolio In-
surance. This is a special type of investment strategy which holds a
greater proportion of its assets in low risk stocks when the funding
level Xt is low. Several sub-cases are investigated including one in
which a risk-free asset exists and one in which it does not. The
latter indicates that selling a particular asset class short could be a
problem and as a consequence certain constraints are put in place.
These constraints prevent the fund from going short on the higher
risk assets when the funding level is low and place an upper limit
on the amount by which the fund can go short on low risk assets
when the funding level is high. In all cases, a closed form solution
can be found for the limiting (stationary) density function of Xt.
When there exists a risk-free asset, this distribution is Translated-
Inverse-Gaussian (TIG).
Much of the analysis relies on the following result:
Theorem 1.1
Let the continuous-time stochastic process Xt satisfy the stochastic
dierential equation
dXt = ( + Xt + Xt2)1=2 dZt + dt ; Xt dt
subject to the constraints on the parameters  > 0,  > 0,  2 ;
4  0,  > 0 and  > 0.
(a) If  2 ; 4 < 0, the stationary density function of Xt is
" #
fX (x) = k exp 2a tan x +c b ( + x + x2 ) 1 =
0 ;1 ; ;

 !
for ;1 < x < 1
where a = p41 ;  2   + 2
b = 
2p
c = 4 ;  2
2
where k is a normalizing constant.
0

(b) If  2 ; 4 = 0 and X0 > b, the stationary density function of

Xt is

fX (x) = k( )(x + b)  exp;=(x + b)] for b < x

;

where b =

2 !
 = 2 1+  

 =  + 22

that is, the Translated-Inverse-Gamma distribution with parameters

;b,  ; 1 > 0 and  > 0 (TIG(;b  ; 1 )). (If X  TIG(k   )
then (X ; k) 1  Gamma(  ).)
;

Proof See Cairns (1996).

2 Model 1: Static investment strategy
This model takes the simplest case possible. In the absence of other
cashows the value of the assets will follow Geometric Brownian
motion. Thus
d(t Xt) = dt = dt + dZt
where Zt is standard Brownian Motion.
In particular investment returns are uncorrelated and do not de-
pend upon the funding level at any point in time. Such a model is
appropriate if the trustees of the fund operate a static asset alloca-
tion strategy: that is, the proportion of the fund invested in each
asset class remains xed.
The de cit at time t is L ; Xt and the adjustment for this de cit
to the contribution rate is
D(Xt) = k:(L ; Xt ):
k = 1=am is the spread factor, and m is the term of amortization.
This method is sometimes referred to as the spread method of amor-
tization (for example, see Dufresne, 1988).
In continuous time this model has been considered by Dufresne
(1990).
The stochastic dierential equation for the fund size is
dXt = (dt+dZt)Xt +(N ;B +k(L;Xt ))dt = ;Xt dt+XtdZt+dt
where  = k ;  and  = (k ; v )L.

2.1 Properties of X
Let X be a random variable with the stationary distribution of Xt.
(Cairns and Parker, 1996, show that such processes are stationary
and ergodic.)
Now Xt falls into the collection of stochastic processes covered in
Theorem 1.1 Thus by Theorem 1.1(b) X has an Inverse Gamma
distribution with parameters  ; 1 and  where  = 2 1 + 2
and  = +242 (that is, X 1  Gamma( ; 1 )). For this to
;

be a proper distribution (that is, one which has a density which

integrates to 1) we require that  > 1. This therefore imposes the
further condition that k >  ; 12 2 . Stronger conditions on k are
required to ensure that X has nite moments.
The stationary distribution of Xt was found by Dufresne (1990),
Proposition 4.4.4, but here we have derived it in a dierent way by
making use of Theorem 1.1.
Let Mj = E (X j ) where j is a non-negative integer. Then it is easy
to show that for j <  ; 1

Mj = ( ; 2)( ; 3):::( ; (j + 1))

j

For j   ; 1, Mj is in nite.
Using these equations we see that
E (X ) = kk ;;v L
E (X 2) = (k ; ()(kk;;v); 1 2 ) L2
2

) V ar(X ) = (k ;(k2)(;kv;) 2 ; 1 2) L2

21 2

Note that it is possible for the process to be stationary but to have

an in nite mean.
Using this information we can calculate, for example, Pr(X > x0 )
where x0 is the government statutory limit of 105% of the actuarial
liability calculated on the UK statutory valuation basis. This gure
gives a guide to the frequency in the long run of breaches of this
upper limit.
2.2 Hitting Times
The problems described below are included as open problems.
Suppose T = inf ft : Xt = xg. Since Xt is stationary it cannot
be true that E s(XT )] = E s(X0)]. If, on the other hand, 0 <
x < X0 < y and T = inf ft : Xt  x or  yg then E s(XT )] =
s(x):Pr(XT = x) + s(y):Pr(XT = y) = E s(X0 )].) Since no closed
form for s(x) exists this problem must be solved numerically.
The problem can be generalized to allow us to gain further infor-
mation about a stopping time T . Suppose we are interested in the
rst time, T , that the process Xt reaches some level x or hits an
upper or a lower bound (y or x). We can at least in principle obtain
the moment generating function for T by generalizing the approach
described in Section 2.1.
Let Yt = f (t Xt) = F (t)G(Xt), which we wish to be a martingale.
Then by Ito's formula we have

dY = FGdt _ + FG dX + 1 FG (dX )2
0 00

2
_ + 1 2 X 2FG ; XFG + FG dt
h i
= FG XdZ + FG
0

2
00 0 0

For Yt to be a martingale we therefore require the :]dt term to be

equal to zero. That is
_
; FF ((tt)) = G(1x)  2x G (x) + (;x + )G (x) = 
h 2 2 i
00 0

) F (t) = F0 exp(;t)
and G(x) satis es:

x2 G (x) + (;x +  )G (x) ; G(x) = 0

00 0

where  = 2=2 ,  = 2=2 and  = 2=2.

Again, no general form for G(x)() can be found, so numerical solu-
tions seems to provide the way forward. However, it may be possible
to prove qualitative results regarding the shape of the distribution
of T .

3 Model 2: Continuous Proportion Port-

folio Insurance
Black and Jones (1988) and Black and Perold (1992) discuss an
investment strategy called Continuous Proportion Portfolio Insur-
ance (CPPI) which is appropriate for funds which have some sort
of minimum funding constraint imposed by either by law or by the
trustees of the fund.
When the funding level is low (A=L < M ) all assets should be
invested in a low risk portfolio (relative to the M ). As A=L rises
above M any surplus and, perhaps more, should be invested in
higher risk assets.
This is in contrast to the static investment strategy discussed in
Section 2 which rebalances the portfolio continuously to retain the
same proportion of assets in each asset class.
Suppose that we have two assets in which we can invest. Asset 1 is
risk free and oers an instantaneous rate of return of 1 . Asset 2 is
a risky asset with d2(t) = 2 dt + 2 dZt. 1 < 2 and 22 > 0 (with
22 > 0). Since asset 2 is risky we have 2 > 1 .
Let p(t) be the proportion of assets at time t which are invested
in asset 2 and let Xt be the funding level at time t. Under the
static investment strategy p(t) = p for all t. Under the CPPI
strategy p(t) depends on Xt only: p(t) = 0 whenever Xt < M 
and p(t) = p(Xt ) > 0 when Xt > M . A strategy which results in
p(t) > 1 for some values of Xt allows for the risk-free asset to be
sold short.
We consider the case p(t) = (Xt ; M )=Xt . Then
 dXt = p(t)Xtd2 (t) + (1 ; p(t))Xt 1 dt + (k ; v )Ldt ; kXtdt
= (Xt ; M )(2 dt + 2 dZt) + M1 dt + (k ; v )Ldt
;k(Xt ; M )dt ; kMdt
Hence

d(Xt ; M )= c:dt ; a(Xt ; M )dt + 2 (Xt ; M )dZt

where a = k ; 2
c= (k ; v )L ; (k ; 1 )M
Xt ; M  Inverse-Gamma( ; 1  )
where  = 2(1 + a2 )
2
 = 2c2
2

) E Xt ; M ] =  ; 2
 2
V ar(Xt ; M ) = ( ; 2)2( ; 3)

Therefore we have

E Xt ] = M + (k ; v )Lk ;
;
(k ; 1 )M
2
" #
V arXt ] = (k ; v )L ; (k ; 1 )M 2 22
k ; 2 2(k ; 2 ) ; 22
provided k ; 2 > 12 22
From these equations, we see that we require c > 0 to ensure that
Xt > M for all t almost surely (that is, the risk-free interest plus
the amortization eort must be sucient to keep the funding level
above M ). We also require a > 0 (that is, k > 2) to ensure that
Xt does not tend to in nity almost surely. Finally we can see that
the variance will be in nite if k ; 2  12 22 .

4 Comparing models 1 and 2

Models 1 and 2 describe two quite dierent asset allocation strate-
gies and it is, therefore, useful to be able to compare them and to
decide which strategy is better and when. The following theorem
answers this to a certain extent.
Theorem 4.1
Suppose that we have a risk-free asset (with d1 (t) = 1 :dt) and a
risky asset (with d2 (t) = 2 :dt + 2 dZ (t)).
Under CPPI the mean funding level is  = E Xt ] and its variance
is C2 = V arXt ]. Under a static investment strategy we invest a
proportion p is the risky asset and 1 ; p in the risk-free asset.
There exists p such that under the static investment strategy E Xt] =
 (as with CPPI) and V arXt ] = S2 < C2 .
Proof See Cairns (1996) but note that the appropriate value of p
is ( ; M )=.
Interpretation: In the variance sense, the static strategy is more
ecient than CPPI: that is, given a CPPI strategy we can always
nd a static strategy which delivers the same mean funding level
but a lower variance.
One example illustrating this result is plotted in Figure 1.
Under CPPI we have 1 = 0:02 2 = 0:05 v = 0:015 22 = 0:152,
L = 1 M = 0:7 and k = 0:1. This gives rise to E Xt] = 1:28,
V arXt ] = 0:3132. The mean is relatively high because the valua-
tion rate of interest v appears very cautious. However, the use of
such a cautious basis is not necessarily too far from regular practice.
Under the equivalent static strategy which has E Xt] = 1:28 we
invest 45.3% of the fund in asset 2 and 54.7% of the fund in the
 2.0
CPPI

Probability Density
1.5
Static
1.0
0.5
0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0

Funding Level

Figure 1: Comparison of the stationary densities for the Static and

CPPI asset allocation strategies. Static (solid curve): E Xt] = 1:28,
V arXt ] = 0:2432. CPPI (dotted curve): E Xt ] = 1:28, V arXt] =
0:3132.
risk-free asset 1. The stationary variance of the fund size is then
found to be V arXt] = 0:2432.
The static variance is signi cantly less than that for CPPI. This is
not too evident from Figure 1, but arises out of the fact that the
CPPI density has a much fatter right hand tail. CPPI also gives a
much more skewed distribution.
Now there are various reasons for why we may prefer CPPI to the
static strategy. Principally this will happen when the objective
of the pension fund is more than just to minimize the variance
of the contribution rate. For example, there may be a penalty
attached to a funding level which is below some minimum. In the
example above, if this is anything below about 0.9 then CPPI may
be favoured. More generally some utility functions may result in a
higher expected utility for CPPI (in particular, those which penalize
low funding levels).
Conversely there exist utility functions which result in optimal
strategies which are the exact opposite of CPPI. For example,
Boulier et al. (1995) maximize the function
Z
V= exp(;s)C (t)2 ds
1

where C (t) = N + D(t Xt) is the contribution rate at time t.

They found that the optimal strategy was to invest in risky assets
when the funding level is low and to move into toe risk-free asset
as the funding level increases. The rationale behind this is that if
there is no minimum funding constraint then: (a) one should try
to reach a high funding level as quickly as possible, no matter how
risky the strategy and (b) when a high funding position is reached
then this should be protected. Investing in a low risk strategy when
the funding level is high will do two things: (a) protect the low
contribution rate and (b) reduce the risk that if too much surplus
is generated then the bene ts will have to be improved.
In practice, one may wish to combine these two extremes by having
a bell shaped asset allocation: that is, one which moves into the
risk-free asset if the funding level approaches the minimum or if
the funding level gets quite high and into more risky assets if the
funding level lies between these two extremes.

5 Model 4: A generalization of CPPI

Section 3 described CPPI in its most basic form. Portfolio A was
considered to be risk free for the purposes of minimum funding,
while Portfolio B was a more risky portfolio oerring higher ex-
pected returns. If the funding level (the A/L ratio) according to
some prescribed basis lies below some minimum M then all assets
would be invested on the low-risk asset A. If the A/L ratio is above
M then a multiple c of the surplus assets over this minimum would
be invested in the risky asset B. Given the existence of a risk-free
asset A, and provided the level of adjustment for surplus or de cit is
high enough then such a strategy ensures that the A/L ratio never
falls below M , provided it starts above this level.
Here we will generalize this strategy to take account of the fact that
often it is not possible to construct a completely risk-free portfolio
(since the nature of the liabilities means that it is rarely possible
for us to match them with appropriate assets).
Suppose that we may invest in a range of n assets. The values
of these assets all follow correlated Geometric Brownian Motion.
Thus asset j produces a return in the time interval t, t+dt) of
n
X
dj (t) = j dt + cjk dZk (t)
k=1

where Z1(t) : : :  Zn(t) are independent standard Brownian Motions.

At all times portfolio A invests a proportion jA in asset j for j =
1 2 : : :  n, with the portfolio being continually rebalanced to ensure
that the proportions invested in each asset remain constant.
Portfolio B follows the same strategy but has a dierent balance
of assets fjB gnj=1. Portfolio B invests in what may be regarded as
more risky assets than does portfolio A.
For portfolio A the return in the time interval t, t+dt) is
n  n !
X X
dA(t) = A j j dt + cjk dZk (t)
j =1 k=1

similarly for portfolio B the return in the time interval t, t+dt) is
n  n !
X X
dB (t) = B j j dt + cjk dZk (t)
j =1 k=1

The matrix C = (cjk ) is somewhat arbitrary but has the constraint

that CC T = V where V is the symmetric convariance matrix for
the n assets.
These equations can be condensed into the following forms:
 dA(t) = Adt + AAdZA(t) + AB dZB (t)
dB (t) = B dt + BA dZA(t) + BB dZB (t)
Xn
where A = jAj
j =1
X n
B = jB j
j =1
 !
and if S = AA AB
BA BB
 !
then SS T = AT VV A AT VV B
T T
B A B B

Thus without loss of generality we may work with two assets 1 and
2 instead of the two portfolios A and B.
At any time a proportion of the fund p(t) is invested in asset 2.
Thus the return in the time interval t, t+dt) is

d(t) = (1 ; p(t))d1(t) + p(t)d2(t)

where d1(t) = 1 dt + 11 dZ1(t) + 12 dZ2(t)
d2(t) = 2 dt + 21 dZ1(t) + 22 dZ2(t)
In a continuous time stationary pension fund model there is a con-
tinuous inow of contribution income C (t) and a continuous outow
of bene t payments B . The contribution rate is made up of two
parts: the normal contribution rate N  and an adjustment for the
dierence between the funding level X (t) and the target level of L.
Thus C (t) = N + k(L ; X (t)).
The stochastic dierential equation governing the dynamics of the
fund size is therefore
dX (t) = X (t)d(t) + N ; B + k(L ; X (t))]dt
Note that if v is the valuation force of interest then N , B and L
are related by the balance equation 0 = dL = v Ldt + (N ; B )dt
which implies that N ; B = ;v L. Hence
dX (t) = X (t)d(t) + (k ; v )L ; kX (t)]dt
Generalizing the formulation of Black and Jones (1988) we suppose
that
p(t) = p0 +Xp(1tX) (t)

Then (abbreviating X (t) by X and dX (t) by dX etc.) we have

dX = (;p0 + (1 ; p1)X )f1 dt + 11 dZ1 + 12 dZ2g

+(p0 + p1X )f2dt + 21 dZ1 + 22 dZ2g
;kXdt + (k ; v )Ldt
= p0(21 ; 11 ) + ((1 ; p1)11 + p1 21 )X ]dZ1
+p0 (22 ; 12 ) + ((1 ; p1)12 + p122 )X ]dZ2
+p0(2 ; 1 ) + (k ; v )L]dt
+(1 ; p1)1 + p12 ; k]X dt
= ( + X + X 2)1=2 dZ3 + dt ; Xdt
where Z3 (t) is a standard Brownian Motion and

 = p20(h21 ; 11 )2 + (22 ; 12 )2]

 = 2p0 (21 ; 11 )((1 ; p1)11 + p1 21 )
i
+(22 ; 12 )((1 ; p1 )12 + p122 )
 = (1 ; p1 )11 + p121 ]2 + (1 ; p1)12 + p1 22 ]2
 = p0(2 ; 1) + (k ; v )L
 = k ; (1 ; p1 )1 ; p12
This stochastic dierential equation for X (t) is therefore in the
correct form for Theorem 1.1. Thus the stationary distribution of
X (t) is
" #
fX (x) = k exp 2a tan x +c b ( + x + x2 ) 1 =
0 ;1 ; ;

 ! for ;1 < x < 1

where a = p 1  + 2
4 ;  2 
b = 2
p4 ;  2
c = 2
This is true provided that it is not possible to synthesize a risk-free
asset out of the two portfolios. If that is the case then we will have
4 ;  2 = 0.
An example of this is given in Figure 2. Here we have 1 = 0:02,
2 = 0:05, v = 0:02, k = 0:1, L = 1, 11 = 0:04, 12 = 21 = 0:08
and 22 = 0:15. The asset allocation strategy uses p0 = ;0:8 and
p1 = 1. This gives E Xt ] = 1:11 and V arXt] = 0:4392. Figure
2 also plots the density for the equivalent static strategy. This
strategy used a linear combination of portfolios 1 and 2 (with p =
0:275) and gives E Xt ] = 1:11 and V arXt ] = 0:3432. We see that
generalized CPPI appears to have a similar eect to the more basic
form: that is, the distribution has lower probabilities of low funding
levels, a fat tail and is more skewed than the static strategy.
It should be noted that below a funding level of M = ;p0=p1 the
new CPPI strategy goes short in asset 2 and long in asset 1. Fur-
thermore, there is nothing to stop the funding level going negative
(although the probability that this happens in any one year is very
small). This is because at that point the fund is long in asset 1
and short in asset 2. If asset 2 performs much better than asset 1
then the funding level will continue to move in a negative direction.
 1.5
Gen.CPPI
Probability Density
1.0
Static
0.5
0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0

Funding Level

Figure 2: Comparison of the stationary densities for the Static and

Generalized CPPI asset allocation strategies. Static (solid curve):
E Xt] = 1:11, V arXt] = 0:3432. Generalized CPPI (dotted curve):
E Xt] = 1:11, V arXt ] = 0:4392.

In eect, when the funding level goes below M , the level of risk
increases again. To avoid this problem, Cairns (1996) considers the
case
( p +p X )
p(t) = 0
0 1 t
Xt whenXt  ;p0 =p1
whenXt > ;p0 =p1
This strategy remains wholly in asset 1 below the minimum and
means that Xt will remain positive with probablity 1.
6 References
Black, F. and Jones, R. (1988) Simplifying portfolio insurance
for corporate pension plans. Journal of Portfolio Management
14(4), 33-37.
Black, F. and Perold, A. (1992) Theory of constant proportion
portfolio insurance. Journal of Economic Dynamics and Control
16, 403-426.
Boulier, J-F., Trussant, E. and Florens, D. (1995) A dynamic
model for pension funds management. Proceedings of the 5th
AFIR International Colloquium 1, 361-384.
Cairns, A.J.G. (1996) Continuous-time stochastic pension fund
models. In preparation.
Cairns, A.J.G. and Parker, G. (1996) Stochastic pension fund
modelling. Submitted.
Dufresne, D. (1988) Moments of pension contributions and fund
levels when rates of return are random. Journal of the Institute of
Actuaries 115, 535-544.
Dufresne, D. (1989) Stability of pension systems when rates of
return are random. Insurance: Mathematics and Economics 8,
71-76.
Dufresne, D. (1990) The distribution of a perpetuity, with applica-
tions to risk theory and pension funding. Scandinavian Actuarial
Journal 1990, 39-79.
Haberman, S. (1992) Pension funding with time delays: a stochas-
tic approach. Insurance: Mathematics and Economics 11, 179-
189.
Haberman, S. (1994) Autoregressive rates of return and the vari-
ability of pension fund contributions and fund levels for a de ned
bene t pension scheme. Insurance: Mathematics and Economics
14, 219-240.
Oksendal, B. (1992) Stochastic Dierential Equations. Springer
Verlag, Berlin.
Parker, G. (1994) Limiting distribution of the present value of a
portfolio of policies. ASTIN Bulletin 24, 47-60.