Portfolio Insurance: determination of a dynamic

CPPI multiple as function of state variables.

H. Ben Ameur* and J.L. Prigent**
* Thema (Universit de Cergy)
** Thema (Universit de Cergy) and ISC (Paris)
December 2006, Preliminary version.
Abstract
This paper examines the CPPI method (Constant Proportion Portfolio Insurance) when the multiple is allowed to vary over time. A quantile
approach is introduced, which can provide explicit values of the multiple
as function of the past asset returns and volatility. These bounds can be
statistically estimated from the behaviour of variations of rates of asset
returns, using for example ARCH type models. We show how the multiple can be chosen to satisfy the guarantee condition, at a given level of
probability and for particular market conditions.
JEL: C 10, G 11.

Introduction

The purpose of portfolio insurance is to give to the investor the ability to limit
downside risk in bearish nancial market, while allowing some participation
in bullish markets. There exist several methods of portfolio insurance: OBPI
(Option Based Portfolio Insurance), CPPI (Constant Proportion Portfolio Insurance), Stop-loss, ... (see for example Poncet and Portait (1997) for a review
about these methods). Here, we are examine the CPPI method introduced by
Black and Jones (1987) for equity instruments and Perold (1986, 1988) for xedincome instruments (see also Black and Rouhani (1987), Roman, Kopprash and
Hakanoglu (1989), Black and Perold (1992)). The CPPI method is based on a
simplied strategy to allocate assets dynamically over time. Basically, two assets are involved: the riskless asset, B, with a constant interest rate r (usually
Treasury bills or other liquid money market instruments) and the risky one, S
(usually a market index).
Usually, the investor, with initial amount to invest V0 , wants to recover a
xed percentage of his initial investment at a given date in the future T . To
obtain a terminal portfolio value VT greater than the insured amount V0 , the
portfolio manager keeps the portfolio value Vt above the oor Ft = :V0 :e r(T t)
at any time t in the period [0; T ]. For this purpose:
The amount et invested in the risky asset is a xed proportion m of the
excess Ct of the portfolio value over the oor.
The constant m is usually called the multiple, et the exposure and Ct the
cushion.
Since Ct = Vt Ft , this insurance method consists in keeping Ct positive
at any time t in the period.
The remaining funds are invested in the riskless asset Bt .
Both the oor and the multiple are functions of the investors risk tolerance.
The higher the multiple, the more the investor will benet from increases in
stock prices. Nevertheless, the higher the multiple, the faster the portfolio will
approach the oor when there is a sustained decrease in stock prices. As the
cushion approaches zero, exposure approaches zero, too. In continuous-time,
when asset dynamics have is no jump, this keeps portfolio value from falling
below the oor. Nevertheless, during nancial crises a very sharp drop in the
market may occur before the manager has a chance to trade. This implies that
m must not be too high (for example, if a fall of 10% occurs, m must not
be greater than 10 in order to keep the cushion positive). Advantages of this
strategy over other approaches to portfolio insurance are its simplicity and its
exibility (see for example De Vitry and Moulin (1994), Black and Rouhani
(1987) and Boulier and Sikorav (1992)). Initial cushion, oor and tolerance can
be chosen according to the own investors objective (see for example Poncet
and Portait (1997) and Prigent (2001)). Banks may bear market risks on the

insured portfolios. In that case, banks can use, for example, stress testing since
they may bear consequences of sudden large market decreases. For instance,in
the case of the CPPI method, banks must, at least, provision the dierence on
their own capital if the value of the portfolio drops below the oor. Thus, one
crucial question for the bank which promotes such funds is : what exposure to
the risky asset or, equivalently, what level of the multiple to accept? On one
hand, as portfolio expectation return is increasing with respect to the multiple,
customers want the multiple as high as possible. On the other hand, due to
market imperfections 1 , portfolio managers must impose an upper bound on the
multiple:
First, if the portfolio manager anticipates that the maximal daily historical
drop (e.g. 20 %) will happen during the period, he chooses m 5 which
leads to low return expectation. Alternatively, he may consider that the
maximal daily drop during the management period will never be greater
than a given value (e.g. 10%). A straightforward implication is to choose
m according to this new extreme value (e.g. m 10). Another possibility
is to take account of the occurrence probabilities of extreme events in the
risky asset returns.
Second, he can adopt a quantile hedging strategy. In this case, he determines the multiple as high as possible but so that the portfolio value will
always be above the oor at a given probability level (typically 99%).
The answer to this latter question has important practical implications. The
determination of the multiple according to quantile conditions have been examined for instance in Prigent (2001) in the case of a Lvy process and in Bertrand
and Prigent (2002), using extreme value theory. Hamidi, Jurczenko and Maillet (2006) also consider a modied quantile hedging strategy and shows how a
conditional multiple can be determined.
In this paper, we also determine how the portfolio manager can choose the
value of the multiple according to market uctuations. This allows the introduction of a conditional multiple. To illustrate our approach, we consider several
quantile conditions and introduce a quite general Garch type model. Our results prove that we can determine a conditional multiple as functions of state
variables. These latter ones are the past stock logretruns and volatilities.
The paper is organized as follows. Section 2 presents the basic properties
of the CPPI model. Section 3 introduces the modied CPPI method with a
conditional multiple, based on various quantile conditions. In particular, upper
bounds on the multiple are provided. Section 4 presents some simulations of
our approach.

1 For example, portfolio managers cannot actually rebalance portfolios in continuous time.
Additionally, problems of asset liquidity may occur, especially during stock markets crashes.

2
2.1

The Model
The nancial market

Consider two basic nancial assets: a riskless asset, denoted by B, and a risky
asset denoted by S (a nancial stock or index price).
Changes in asset prices are supposed to occur at discrete times along a whole
period [0; T ].
The riskless asset evolves according to a deterministic rate denoted by r.
The variations of the stock price S between two times tk and tk+1 are dened
by:
Stk+1 = Stk+1 Stk :
Since we search an upper bound on the multiple m (see Proposition 1 in the
following), we have to focus on the left hand side of the probability distribution
St
of Stk+1 . Thus, we introduce the notation :
k

Xk+1 =

Stk+1
St
Stk+1
= k
Stk
Stk

So Xk denotes the opposite of the relative jump of the risky asset at time tk :
In fact, when we want to determine an upper bound on the multiple m, we
have only to consider positive values of X.
Denote by Mn ; the maximum of the nite sequence (Xk )1 k n . We have:
Mn = M ax(X1 ; :::; Xn ):

2.2

The CPPI portfolio

Denote by Vk the value of the portfolio at time tk . As explained in the introduction, the CPPI method is based on the following portfolio insurance condition:
There exists a deterministic oor Fk such that at any time tk , the value
Vk must be above the oor.
The total amount ek invested on the underlying asset S is equal to mCk
where the cushion Ck represents the dierence Vk Fk between the portfolio value and the oor.
The multiple m is a nonnegative constant.
The remaining amount (Vk ek ) is invested on the riskless asset with a
deterministic rate rk for the period [tk 1 ; tk ] :
The higher the multiple m, the higher the amount invested on the risky asset.
Therefore, a speculative investor would choose high values for m. Nevertheless,
in this case, his portfolio is riskier and as shown in what follows, his guarantee

may no longer hold. Indeed, we easily deduce that the portfolio value is solution
of
Vk+1 = Vk ek Xk+1 + (Vk ek )rk+1 ,
from which, we obtain the dynamics of the cushion :
Ck+1 = Ck [1

m Xk+1 + (1

m)rk+1 ]

Since, for all times tk ; the cushion must be positive, we get nally the condition : for all k n,
mXk + (1 m)rk
1
In fact, since rk is relatively small, the previous inequality yields to the following
relation that gives an upper bound on the multiple:
Proposition 1 The guarantee is satised at any time of the management period
with a probability equal to 1 if and only if
8k

n; Xk

1
or equivalently Mn = M ax(Xk )k
m

1
:
m

Since the right end point d of the common distribution F of the variables
Xk is positive, we deduce that the insurance is perfect along any period [0; T ] if
and only if m is smaller than d1 .
For example, if the maximal drop is equal to 20%, then d = 0; 2. Thus m
must be smaller than 5.

2.3

Quantile conditions

The previous condition, which is rather strong, can be modied if a quantile

hedging approach is adopted, like for the Value-at-Risk (see Fllmer and Leukert
(1999) for recent application of this notion in nancial modelling) : for example,
an auxiliary oor can be chosen above the initial oor and the new condition
is to guarantee that the portfolio value will be always above this new oor at a
given probability 1
: This gives the following relation for a period [0; T ]:
P[Ct

0; 8t 2 [0; T ]]

or, equivalently, if MT denotes the maximum of the Xk for times tk in [0; T ] :

P[8tk 2 [0; T ] ; Xk

1
] = P[MT
m

1
]
m

1
Note that, since m is nonnegative, the condition Xk
m is equivalent to
1
Xk I0 Xk
m.
Then we can detail the quantile hedging condition. For this, introduce the
function FMT1 dened as the inverse of the distribution function F which is
assumed to be strictly increasing. Then, we get (see Prigent (2001)) :

Proposition 2 For small rk ;we have approximately the following condition:

m

FM T

1
((1

))

Additionally, if the sequence (Xk )k is i.i.d. with common cdf FMT , then we
have:
1
:
m
1
F 1 (1
)T
This condition gives an upper limit on the multiple m which is obviously
higher than the standard limit d1 .

3
3.1

CPPI with a conditional multiple

Conditional multiple

The simplicity and exibility of the CPPI method allows the introduction of
several extensions:
First, we can introduce a stochastic oor, in particular to keep past prots
from rises in the stock market. For example, Estep and Kritzman (1988)
have introduced the Time Invariant Portfolio Protection (TIPP). This
method is based on the following guarantee condition:
Vt

max(Pt ; sup Vs );
s t

where k is an exogenous parameter which lies between 0 and 1. In this

case, the investor does not want to lose more than a given percentage of
the maximum of his past portfolio values. This strategy is of CPPI type
but with a stochastic oor, given by k max(Pt ; sups t Vs ):
Second, the multiple can be no longer xed but determined from market
and portfolio dynamics:
For instance, Prigent (2001) introduces a more general exposure function
e(t; X) dened on [0; T ] R+ ; positive and continuous.
This new exposure e has the following form:
et = e(t; Ct ):
Then, conditions on e(t; X) must be imposed such that the cushion
always is positive.
1) If the cushion is nil, then the exposure must be equal to 0:
e(t; 0) = 0:
6

2) If the relative sizes of jumps SS have a lower bound d (negative),

then, for any (t; X); we must have:
1
X:
d

e(t; X)

Another possibility is to chose the multiple according to asset values, turbulence, and market volatility, as proposed by Hamidi et al. (2006). In particular,
they use a regression quantile method, as introduced by Engel and Manganelli
(2004).
In what follows, we examine the problem of the determination of a conditional multiple, using a quantile condition. To illustrate our approach, we
assume that the risky asset logreturn follows an Arch type model.

3.2

Quantile Hedging and conditional multiple

3.2.1

Quantile condition

The quantile condition is dened by:

P[8t; Ct > 0]

(1)

This condition can be deduced for su cient assumptions. For instance, we

have the following result:
Proposition 3 Consider the ltration (Ftk )k generated by the observations of
the arithmetical returns ( Xk )k .
Introduce the assumption (A1):
8k; P[Ctk > 0 Ct1 > 0; :::; Ctk

> 0]

(1

)1=T :

Consider also the assumption (A2):

8k; PFtk

[Ctk > 0]

(1

)1=T :

(1

)1=T ;

Finally, introduce the condition (A3):

8k; PGtk

[Ctk > 0]

where Gtk 1 is the

algebra generated by Ftk 1 and the random event
Ctk 1 > 0 .
Each of these three assumptions implies Property (1).
Proof. Suppose (A1) is satised. Then, using the equality:

:::

P[Ct1
P[CT

> 0; :::::CT > 0] = P[Ct1 > 0]

> 0 jC1 > 0; ::::CT 1 > 0 ];

we deduce the result.

7

P[Ct2 > 0 jCt1 > 0 ]

:::

Suppose now that (A2) is satised.

We use the following lemma: for any random variables X and Y;
PY [X < 0]

a ) PY 2B [X < 0]

Rk

a; 8B

Let:
X = Ctk and Y = (Xt1 ; :::::::::::::Xtk 1 ):
Introduce the set B dened by:
B = Ct1 > 0; ::::::::::::::Ctk

>0 :

Then, we have:
8k; PCt1 >0;::Ctk

>0

[Ctk > 0]

)1=T ;

(1

from which, we deduce that assumption (A1 ) is satised, which implies Property
(1).
In what follows, we focus on Conditions (A2 ): PFtk 1 [Ctk > 0] (1
)1=T :
Recall that the cushion value is determined from the following relation2
Ctk = Ctk

(1

mtk

Xtk ) > 0:

The guarantee condition at any time tk is: Ctk > 0:

Case 1: (Ctk 1 > 0)
Ctk > 0 () 1
Case 2: (Ctk

mtk

Xtk > 0:

mtk

Xtk < 0:

< 0)
Ctk > 0 () 1

As it can be seen, at any time tk 1 ; we have to choose a value of the multiple

mtk 1 such that, at time tk ; we still have Ctk > 0:
Consequently, we can search mtk 1 of the following type:
mtk

= gtk

ICtk

1 >0

+ htk

ICtk

1 <0

where both are random variables which are Ftk 1 measurable.

Condition (A3) implies condition (A1). Thus, it also implies Property (1).
2 Since

we consider small time intervals (for example, daily rebalancing), we set rk = 0.

3.2.2

Determination of the multiple

In what follows, we search for explicit forms of the random variables gtk 1
and htk 1 : We assume that the risky asset logreturn Y follows a Garch(p,q)
model. As it well-known, this kind of dynamics is quite suitable to describe asset
uctuations in a discrete-time setting. The ARCH (Autoregressive Conditionally
Heteroscedastic) models, introduced by Engle (1982), are specic non-linear
time series models. They can describe quite exhaustive set of the underlying
dynamics. They have been largely applied on macroeconomics and statistical
theory.
The GARCH is dened in the following way. Consider the logreturn Y :
Ytk = ln

Stk
Stk 1

Stk

Stk
Stk

= exp(Yk )

1:

Consider the systems of auto regressive equations:

Pp
Yk = 0 + i=1 i Ytk i + k
k;
( k ) = + C0 ( k 1 ) + C1 ( k 1 )
( k

1 );

where k denotes the volatility, the sequence ( k )k is i.i.d with common pdf
f > 0 and , C0 (:); and C1 (:) are deterministic functions. In particular, we
assume that the function : R+ ! R is strictly increasing.
The information delivered by the observation of risk asset returns is generated by the ( t1 ; ::::: tk 1 )k :
We have:
Ftk 1 =
algebra ( t1 ; ::::: tk 1 ):
(

Thus the random variables gtk 1 and htk 1 are deterministic functions of
t1 ; ::::: tk 1 ):
Therefore, we have to search for a multiple mtk which has the following form:
mtk

= g(tk

+h(tk

1 ; Yt1 ; ::::::; Ytk

1 ; Yt1 ; ::::::; Ytk

t1 ; :::::: tk )

t1 ; :::::: tk )

ICtk

ICtk
1 <0

1 >0

Remark 4 Note also that the random variables Ytk 1 are deterministic functions ( t1 ; ::::: tk 1 ): But, for a better nancial interpretation of the multiple, we
explicitly introduce functions of (Yt1 ; ::::::; Ytk 1 ) itself.
Indeed, we examine how the conditional multiple depends on both the volatility levels on the past logreturns. These two kinds of variables are here the state
variables.
In order to determine the conditional multiple, two cases have to be distinguished3 :
The cushion at time

tk

is non negative: Ctk

The cushion at time

tk

is non positive: Ctk

3 Since we assume that the pdf of the random variables

of the event Ctk 1 = 0 is null.

> 0:
< 0:

is strictly positive, the probability

Case 1: Ctk

> 0 The quantile condition (A2) is the following:

PFtk

Stk

[1 + mtk

Stk

> 0]

)1=T ;

(1

which is equivalent to:

PFtk

[mtk 1 (exp(Ytk )

1) >

1]

)1=T

(1

At time tk , we have two cases:

exp(Ytk ) < 1 : Stk < Stk
exp(Ytk ) > 1 : Stk > Stk

(S decreases),
(S increases).

1
1

We assume that the multiple mtk 1 is non negative (it is the standard assumption. It corresponds to a long position on the risky asset).
Then, we deduce:
PFtk
PFtk
P
=P

Ftk

[mtk 1 (exp(Yk )

[mtk 1 (exp(Yk )

Ftk

1) >

[mtk 1 (exp(Yk )

[exp(Yk ) 1 > 0]+P

Ftk

1) >

1) =

1 \ (exp(Yk )

1) >

1) > 0]+

1 \ (exp(Yk )

[mtk 1 (exp(Yk ) 1) >

1) < 0]
1\(exp(Ytk )

1) < 0]

Since we have:
mtk 1 (exp(Yk )

1)

1 ) Ytk

ln(1

);

mk

then
PFtk

1) = PFtk

[mtk 1 (exp(Yk ) 1) >

= PFtk

1
mtk

[Ytk > ln(1

[Ytk > 0]+PFtk

)] = 1
1

Therefore, the quantile condition

PFtk

1 ;Ctk

1 >0

[Ctk > 0]

[0 > Ytk > ln(1

1
F Fk 1 [ln(1
mk
{z
|
)1=T ;

(1

1
mk

)] :
}

is equivalent to:
1

F zk 1 [ln(1

1
mtk

)]

(1

)1=T ;

and also to:

ln(1

1
mtk

(F zk 1 )

[1

(1

)1=T ]

and nally, we have:

1-1) If exp[(F ztk 1 ) 1 [1 (1
)1=T ]] > 1; since mtk
non negative, the quantile condition is
10

is assumed to be

)];
1

already satised.
1-2) If exp[(F ztk
following constraint:

[1

)1=T ]] < 1; then mtk 1 must satisfy the

(1

mtk

exp[(F ztk 1 )

[1

)1=T ] ]]

(1

Let us examine the two previous subcases. They are based on the sign of
the term:
)1=T ] :
(F ztk 1 ) 1 [1 (1
Consequently, we must study the conditional cdf F ztk
assume (GARCH(p,q) model):
PP
Yk = 0 + i=1 i Ytk
( k ) = + C0 ( k 1 ) + C1 (

: Recall that we

k;

k 1)

k 1 ):

Lemma 5 For any real numbers a and b (b > 0) and for any random variable
with pdf f , we have:
P[a + b

x] = P[ <

a
b

]=

a
b

f (u)du:
1

Thus the pdf of a + b is given by:

f (x) =

1
b

f(

a
b

):

We apply this lemma by setting:

8
a + b = Ytk ;
>
>
<
= ;
PP tk
a
=
+
Ytk i ;
>
0
i=1 i
>
:
b = k > 0:
We get:

z
fYkk

(y) =

z
FYkk

( +

tk

(y) = F

PP

i=1

Ytk i )

tk

( +

PP

i=1 Ytk

tk

where F denotes the common cdf of the random variables

Thus, we deduce:

11

k:

The condition in subcase (1-1) is satised if and only if

0

P
X

Ytk

(1

)1=T ) > 0:

(1

i=1

The condition in subcase (1-2) is satised if and only if

0

P
X

Ytk

(1

)1=T ) < 0;

(1

i=1

in which case, the conditional multiple must satisfy:

mtk

exp[(

PP

Ytk i ) +

i=1

1 (1

)1=T )]

(1

Remark 6 The previous result shows that, if at time tk 1 , the auto regressive
PP
terms 0 + i=1 i Ytk i and the conditional volatility k are su ciently high,
then the logreturn Ytk has a su ciently high probability to be positive and thus,
there is no constraint on the multiple at time tk 1 4 :
Case 2: Ctk
given by:

< 0 and mtk

PFtk

> 0 As previously, the quantile condition is

[1 + mtk

Stk
< 0]
Stk 1

(1

)1=T ;

which is equivalent to:

PFtk

[mtk 1 (exp(Ytk )

1) <

1]

(1

)1=T

At time tk , we have two cases:

exp(Ytk ) < 1 : Stk < Stk
exp(Ytk ) > 1 : Stk > Stk
We assume that the multiple mtk

1
1

(S decreases),
(S increases).

is non negative.

PFtk

[mtk 1 (exp(Ytk )

1) <

1] =

Ftk

[mtk 1 (exp(Ytk )

1) <

1 \ (exp(Ytk )

1) > 0]

+P

Ftk

[mtk 1 (exp(Ytk )

1) <

1 \ (exp(Ytk )

1) < 0]:

Therefore, we have:
PFtk

[mtk 1 (exp(Ytk )

1) <

1] =

Ftk

[mtk 1 (exp(Ytk )

1) <

1 \ (exp(Ytk )

Ftk

[mtk 1 (exp(Ytk )

1) <

1]

4 Note that F 1 (1
(1
the quantile at the level (1

1) < 0] =

)1=T ) is a particular value of the random variable

(1
)1=T )).

12

(since it is

Consequently, we get the equivalence between the two following conditions:

PFtk

PFtk

Stk

[1 + mtk
1

[Ytk

Stk

< 0]

ln(1

)1=T ;

(1

mk

)]

)1=T :

(1

Thus, the quantile condition is also equivalent to:

F ztk 1 [ln(1
1

ln(1
1

(1

mk

exp((F ztk 1 )

1 ) 1 [(1
2-1) If exp[(F
ztk 1
2-2) If exp[(F
) 1 [(1
constraint:

)1=T ;

(1

ztk

mtk

)]

mk

(F ztk 1 )

mk

(1

(1

i
)1=T ;

i
)1=T ):

)1=T ]] > 1; there is no positive solution for mtk 1 :

)1=T ]] < 1; then mtk 1 must satisfy the following
1
exp[(F ztk 1 ) 1 [(1

)1=T ] ]]

Thus, we deduce:
The condition in subcase (2-1) is satised if and only if

P
X

Ytk

)1=T ) > 0:

((1

i=1

The condition in subcase (2-2) is satised if and only if

P
X

Ytk

)1=T ) < 0;

((1

i=1

in which case, the conditional multiple must satisfy:

mtk

exp[(

PP

i=1

13

Ytk i ) +

1 ((1

)1=T )]

Case 3: Ctk

< 0 and mtk

PFtk

< 0 The quantile condition is:

[1 + mtk

Stk
1

Stk

< 0]

)1=T ;

(1

which is equivalent to:

PFtk

[mtk 1 (exp(Ytk )

1) <

1]

)1=T

(1

We have:
PFtk

[mtk 1 (exp(Ytk )

1) <

1] =

PFtk

[mtk 1 (exp(Ytk )

1) <

1 \ (exp(Ytk )

1) > 0]

Ftk

[mtk 1 (exp(Ytk )

1) <

1 \ (exp(Ytk )

1) < 0]

+P

= PFtk

[mtk

Ftk

[exp(Ytk ) > 1

= P

1 (exp(Ytk )

1) <
1
]

mtk

1]

Therefore the quantile condition is equivalent to:

1

F ztk 1 [ln(1

mtk

)]

)1=T ;

(1

which is also:
1

(1

mk

ztk

3-1) If exp[(F
lowing constraint:

mtk

exp((F ztk 1 )

[1

(1

(1

i
)1=T ):

)1=T ]] > 1; then mtk 1 must satisfy the fol1

3-2) If exp[(F ztk 1 )

mtk 1 :
Thus, we deduce:

exp[(F ztk 1 )

1
1

[1

[1

)1=T ] ]]

(1

)1=T ]] < 1; there is no negative solution for

(1

The condition in subcase (3-1) is satised if and only if

P
X

Ytk

i=1

14

[1

(1

)1=T ] > 0:

The condition in subcase (3-2) is satised if and only if

P
X

Ytk

[1

(1

)1=T ] < 0;

i=1

in which case, the conditional multiple must satisfy:

mtk

exp[(

0+

PP

Ytk i ) +

i=1

To summarize, we getPthe following result.

P
Denote Ztk 1 = 0 + i=1 i Ytk i + k F
PP
Ytk i + k F 1 ((1
)1=T ):
0+
i=1 i

(1 (1

[1

(1

)1=T ] ]

)1=T ) and Wtk

Proposition 7 The quantile condition (A2) at time tk 1 can be satised as

soon as the cushion value at time tk 1 is non negative. In this case, if Ztk 1 >
0, then the multiple can take any positive value, and, if Ztk 1 < 0, then the
conditional multiple must satisfy:
mtk

1
:
exp[Ztk 1 ]

If the cushion value at time tk 1 is non positive, then:

- If Wtk 1 < 0; then there exist positive solutions mtk 1 ;which must satisfy
the following constraint:
mtk

1
:
exp[Wtk 1 ]]

- If Ztk 1 > 0; then there exist negative solutions mtk 1 , which must satisfy
the following constraint:
mtk
- If Wtk 1 > 0 or Ytk
no negative solution.

1
:
exp[Ztk 1 ]]

> 0, there exist respectively no positive solution and

Corollary 8 The quantile condition (A3) at time tk 1 can always be satised.

If Ztk 1 > 0, then the multiple can take any positive value, and, if Ztk 1 < 0,
then the conditional multiple must satisfy:
mtk

1
:
exp[Ztk 1 ]

Remark 9 When the cushion is positive at time tk 1 ; the choice of the multiple is very exible. Thus, within the quantile condition at time tk 1 ; we can

15

add some other conditions on the multiple to better benet from market conditions. When the cushion is negative (which happens with a small probability),
the quantile condition generally cannot be satised, except for small values of
(1
): But, in this case, this is not a true insurance condition. Therefore,
a possible strategy is to adopt the previous condition when the cushion is positive and to invest the whole portfolio value on the riskless asset, as soon as the
cushion is negative.

Simulations

4.1

First case: i.i.d. returns and xed multiple

We can determine the value of the multiple according to assumptions on logreturn process.
For example, suppose that St follows a geometric Brownian motion:
St+1 = St

1
2

exp((

)t +

Zt );

where, in particular, Zt has a standard Gaussian distribution N (0; 1): Denote

by N its cdf.
In that case, the previous conditions are made on Ztk 1 and Wtk 1 which
now are constant:
Ztk

1
2

=(

) t+

(1

(1

)1=T );

and

p
1
2
) t+
t N 1 ((1
)1=T ):
2
- Consider for instance the following parameter values:
Wtk

=(

= 0:1;
= 0:3:
Proposition 10 Since the xed multiple m is positive, then the probability that
there exists a time t such that the cushion Ct is negative is given by:
1

P [8t; Ct > 0] = 1

where
F (x) = N

ln 1

p2

) t

1
m

In what follows, we determine the correspondence between the level of probability (1

) and the multiple m.
For instance, we x the value of the instantaneous rate of return: = 5%
and consider a varying volatility.
16

m(
(1
m(
(1
m(
(1
m(
(1
m(
(1

= 0; 2)
) en %
= 0; 25)
) en %
= 0; 30)
) en %
= 0; 35)
) en %
= 0; 40)
) en %

19
100
13
100
12
100
9
100
7
100

20
99; 6
14
100
13
100
10
100
8
99:8

21
99
15
100
14
99:8
11
99:8
9
99:6

22
98:5
16
99:8
15
99
12
99:4
10
99

23
97:8
17
99:4
16
97:8
13
97
11
98:7

24
96:2
18
98:9
17
96
14
96
12
98

25
95
19
95
19
93:8
15
94
13
95:4

26
93
20
94:5
20
90:5
16
90
14
91

As shown in previous table, the volatility levels determine the multiple values. This is also illustrated by the next gure.

Note that the higher the volatility , the smaller the multiple value m for
xed probability level (1 ").
p
For example, if T = 1 year, for daily volatilities equal to = 0; 40
1=250;
the conditional multiple varies between 7 and 14.
Recall that the usual values of the non conditional multiple are in f6; 10g :

17

4.2

Dependent returns (ARCH type models)

Consider for instance the Garch(1,1) model with parameter values such as in
Nelson (1990).
p
1 2
t
;
Ytk = (
) t + tk
tkp
2
=
+
(a
)
t
+
t:
tk
tk 1
tk 1
tk
4.2.1

Convergence of the Garch(1,1) model to an Ornstein-Uhlenbeck

process

The continuous-time model is given by:

d

= (a

t )dt

+ dWt ;

where is the speed of convergence to the long term value of the volatility a:
The parameter denotes the volatility of the volatility. We get:
Z t
a 0 exp( t) +
exp( (s t) dWs :
t =a
0

4.2.2

Determination of the conditional multiple from Garch (1,1)

model.

Consider the following parameter values:

= 0:2; = 0:07; = 5; T = 250;
a = 0:2
1) Next gure illustrate the paths of the risky asset, under the Garch assumptions.

18

2) Paths of the portfolio value are as follows (recall that we take V0 = 104 ).

3) We obtain daily logreturns whosevalues are most of the time between

0:02% and 0:02%:

4) The volatility levels are indicated in the following gure. For the parametric specication that we consider, the volatility values converge to the long
term value equal to 2%:
19

5) The cdf of the daily logreturns is gaussian and illustrated below:

6) For the parameter values of this example, the variations of the conditional
multiple are shown in next gure.

20

The previous values of the conditional multiple imply the insurance quantile
condition. Since these values are higher than for the standard case, better
performances can be observed when the risky asset increases. However, other
conditions can be imposed to limit downside risk, while allowing the quantile
condition.

Conclusion

As shown in this paper, it is possible to choose higher multiples for the CPPI
method if quantile hedging is used. Upper bounds can be calculated for each
level of probability and according to state variables. This allows the introduction
of a conditional multiple. This new multiple can be determined according to the
distributions of the risky asset logreturn and volatility. Other conditions can
be imposed on this multiple, while the quantile hedging constraint is satised.
The dierence with the standard multiple is signicant. Further extensions
may allow to better take account of the potential losses, when nancial asset
prices decrease. For example, criterion such as the Expected Shortfall can be
introduced. Other state variables can also be considered, such as exogenous
macro economic factors. Finally, the impact of transaction costs can also be
examined.

21

References
[1] Bertrand, P., and Prigent, J-L. (2002). Portfolio insurance: the extreme
value approach to the CPPI method. Finance.
[2] Black, F. and Jones, R. (1987). Simplifying portfolio insurance. The Journal
of Portfolio Management, 48-51.
[3] Black, F. and Perold, A.R. (1992). Theory of constant proportion portfolio
insurance. The Journal of Economics, Dynamics and Control, 16, 403-426.
[4] Black, F.and Rouhani, R. (1989). Constant proportion portfolio insurance
and the synthetic put option : a comparison, in Institutional Investor focus
on Investment Management, edited by Frank J. Fabozzi. Cambridge, Mass.
: Ballinger, pp 695-708.
[5] Bookstaber, R. and Langsam, J.A. (2000). Portfolio insurance trading rules.
The Journal of Futures Markets, 8, 15-31.
[6] Boulier, J.F. and J. Sikorav, (1992), Portfolio insurance: the attraction of
security, Quants, 6, CCF, Paris.
[7] Bremaud, P., (1981), Point processes and Queues: martingale dynamics.
Springer Verlag, Berlin.
[8] De Vitry, T. and S. Moulin, (1994), Aspects thoriques de lassurance de
portefeuille avec plancher, Banque et Marchs, 11, 21-27.
[9] Engle, R., (1982), Autoregressive Conditional Heteroscedasticity with estimates of the variance of U.K. ination, Econometrica, 50, 987-1008.
[10] Engel, R., and Manganelli, S. (2004), CAViaR: Conditional Auto Regressive
Value-at-Risk by Regression Quantiles, Journal of Business and Economic
Statistics, 22, 367-381.
[11] Ested T. and M. Kritzman,(1988), TIPP: insurance without complexity ,
Journal of Portfolio Management.
[12] Fisher, R.A. and L.H.C.Tippett, (1928), Limiting forms of the frequency distribution of the largest and smallest member of a sample.
Proc.Cambridge Philos.Soc., 24, 180-190.
[13] Fllmer H. and P. Leukert, (1999), Quantile Hedging, Finance and Stochastics, 3, 251-273.
[14] Hamidi, B., Jurczenko, E., and Maillet, B. (2006): Dun multiple conditionnel en assurance de portefeuille, Working paper, A.A.Advisors-QCG
(ABN Amro Group).
[15] Last G. and A. Brandt, (1995), Marked point processes on the real line,
Springer-Verlag, Berlin. 383-408.
22

[16] Nelson D.B. (1990), Arch models as diusion approximations, Journal of

Econometrics, 45, 7-39.
[17] Perold, A.R., (1986), Constant Proportion Portfolio Insurance, Harvard
Business School, Working paper.
[18] Perold, A.R. and W. Sharpe , (1988), Dynamic Strategies for Asset Allocations, Financial Analysts Journal, 16-27.
[19] Poncet, P. and R. Portait , (1997), Assurance de Portefeuille in Y. Simon
ed. Encyclopdie des Marchs Financiers, Economica,140-141.
[20] Prigent, J.L., (2001a), Option pricing with a general marked point process,
Mathematics of Operation Research, 26, 50-66.
[21] Prigent, J.L., (2001b), Assurance du portefeuille: analyse et extension de
la mthode du coussin, Banque et Marchs, 51, 33-39.
[22] Roman E., Kopprash R. and E. Hakanoglu , (1989), Constant Proportion
Portfolio Insurance for xed-income investment, Journal of Portfolio Management.

23