AUTOCALLABLE STRUCTURED PRODUCTS

Author : Tristan Guillaume

Professional title : Lecturer and research fellow
Mailing address : Université de Cergy-Pontoise, Laboratoire Thema, 33 boulevard du port, F-95011
Cergy-Pontoise Cedex, France
E-mail address : tristan.guillaume@u-cergy.fr
Telephone number : + 33 6 12 22 45 88
Fax number : + 33 1 34 25 62 33

Abstract

In this article, a general, flexible form of autocallable note is analytically valued, and its payoff profile
and risk management properties are discussed. Two models are considered : a two-asset Black-Scholes
one (Black and Scholes, 1973), and a Merton jump-diffusion framework (Merton, 1976) combined
with a Ho-Lee two-factor stochastic yield curve (Ho and Lee, 1986). In the former setting, the general
autocallable structure under consideration includes many popular features : regular coupons, reverse
convertible provision, down-and-out American barrier, best-of mechanism and snowball effect. These
features are more or less fully addressed according to the entailed valuation difficulties. Simpler notes
are easily designed and priced on the basis of this general structure. In the second analytical
framework, the autocallable payoff structure is restricted to the following features : regular coupons, a
reverse convertible provision and possible participation in the growth of a single underlying equity
asset.
The formulae provided in this paper can be expected to be a valuable tool for both buyers and issuers
in terms of risk management. Indeed, they enable investors to assess their chances of early redemption
as well as their expected return on investment as a function of the contract’s specifications, while they
allow issuers to analyze and compute their various risk exposures in an accurate and efficient manner.

1. Introduction

Autocallables, also known as auto-trigger structures or kick-out plans, are very popular in the world of
structured products. They have captured a large part of the market share in recent years. Product
providers use them to offer higher payoffs than those on structured products that automatically run to a
full term. In its standard form, an autocallable is a note that is linked to an underlying risky asset

1
(usually a single stock, a basket of stocks or an equity index) and that has no fixed maturity. What is
referred to as the maturity of the autocallable is actually the maximum duration this product can stay
alive, usually ranging from 2 to 5 years. Several observation dates within the product’s life are
prespecified in the contract, typically on an annual or semi-annual basis. At each observation date, if
the value of the underlying is at or above a prespecified level, usually called the autocall trigger level
or autocall barrier, then the principal amount is paid back by the issuer to the holder of the note, along
with a coupon rate. It is said then that the note autocalls. The prespecified autocall trigger level is often
defined as the level of the underlying asset at the contract’s inception, but it does not have to be. It
may also vary in time. If there is no early redemption, the note proceeds to the next observation date,
where there is again the possibility of early redemption. A lot of plans kick out in year one or two,
leaving investors with the choice to reinvest in rollover substitutes from the same provider, switch to
another offer or opt back into the markets.
Another level can be prespecified for each observation date, below the autocall trigger level, such that
if the note does not autocall but the underlying is above that lower level, usually called the coupon
barrier, then the note pays a coupon rate. Some autocallable notes have a memory function embedded
– also called a snowball effect. With a memory function, the product will pay any coupons that have
not been paid on previous observation dates, if on a subsequent observation date all prerequisites are
met.
Several variants of the standard structure are traded on the markets. One way to allow for yield
enhancement is to include a reverse convertible mechanism : if the note has not been autocalled and if
the underlying asset has fallen below a prespecified conversion level at maturity, which is a kind of
safety threshold from the point of view of the investor, then the principal amount is not redeemed and
shares (or their equivalent cash amount) are paid back instead, which entails a capital loss for the
investor. This amounts to the sale of an out-of-the-money put by the investor to the issuer. This is a
suitable feature in the current market environment. Indeed, low interest rates leave little room for yield
enhancement if investors want full capital protection, as providers must use more of the initial capital
to guarantee its complete return. Moreover, in high volatility markets, volatility has to be sold in order
to generate a higher income. The answer is then to introduce capital-at-risk structures, typically losing
money if the underlying index has fallen 50% or further from its initial level.
Another way to allow for yield enhancement is to make the coupon payments at a given observation
date contingent on the underlying asset not having crossed a prespecified lower level since the latest
observation date. This amounts to the introduction of a down-and-out American barrier (i.e., a
continuous barrier) in a time interval between two observation dates.
Investors may also want to be given the opportunity to participate in potential increases in the value of
the underlying instead of receiving fixed coupons. More specifically, some contracts define a
prespecified upper level such that, if the underlying is above that level at a given observation date,
then the note autocalls and the coupon paid to the holder of the note is a percentage of the spot value

2
of the underlying. Finally, some autocallable notes include best-of (worst-of) features, which usually
consist in providing investors with a percentage of the maximum (minimum) between two or more
assets in case the contract allows for a participation in the market upward potential.
Despite their widespread use in the markets, there are not many academic studies on
autocallable products. While Bouzoubaa and Osseiran [2010] describe a variety of payoffs and analyse
risk management issues, all the other contributions focus on numerical pricing schemes. Fries and
Joshi [2008] study a product specific variance reduction scheme for Monte Carlo simulation purposes.
Deng et al. [2011] discuss finite difference methods in a numerical partial differential equation
framework. Kamtchueng C. [2011] discusses smoothing algorithms for the Monte Carlo simulation of
the Greeks. Alm et al. [2013] develop a Monte Carlo algorithm that allows stability with respect to
differentiation. No research article has yet come up with an analytical solution to the valuation
problem raised by the autocallable structured products traded in the markets, i.e. autocallables with
discrete observation dates. The main purpose of this paper is to fill this gap by providing a closed form
formula for a flexible autocallable structure with discrete observation dates. In a Black-Scholes
framework (Black and Scholes, 1973), our formula can capture the following features : regular
coupons, reverse convertible provision, down-and-out American barrier, best-of mechanism and
snowball effect. The way in which all these features can be embedded into an analytical formula is
more or less restrictive according to the entailed valuation difficulties : while regular coupons, the
reverse convertible provision and the snowball effect can be fully tackled, the American barrier and
best-of features are only partially covered, in the sense that the American barrier is not alive during the
entire product life and the best-of mechanism is limited to two assets. The advantage of using a Black-
Scholes model is that it enables to value sophisticated payoffs and it provides a simple hedging
method. However, it is well known that the model assumptions are flawed. In particular, it has long
been known that equity prices are not purely continuous and exhibit jumps. The latter are a way to
account for the skew observed in the options market, especially the steeper skew for short expirations
(Gatheral, 2006). Secondly, the assumption that interest rates are constant is particularly spurious for
the valuation of autocallable notes. Indeed, the latter are a combination of fixed income and equity
components which are usually long-dated. Moreover, the correlation between equity and interest rate
sources of randomness has a significant impact. When the stock market goes up, the duration of the
autocallable structure goes down. If there is a positive correlation between equity and interest rate,
sellers of the note make losses while hedging their interest rate exposure, whether equity increases – as
they have to sell longer-dated zero coupon bonds and buy more short term zero coupon bonds under
higher interest rates, or decreases – as they need to sell short term bonds and buy more longer-dated
bonds under lower interest rates. Conversely, if there is a negative correlation between equity and
interest rate, sellers of the notes make a net profit while hedging their interest rate exposure, whether
equity goes up and down because of the opposite directions of equity and interest rate. As a
consequence, pricing models that do not take the correlation between equity and interest rate into

3
account will underprice the autocallable structure when this correlation is positive and overprice it
when that correlation is negative.
That is why a second model is considered, which consists of a combination of a Merton-type jump-
diffusion (Merton, 1976) and a Ho-Lee two-factor stochastic yield curve (Ho and Lee, 1986). As a
result of the increased sophistication of the model assumptions, fewer payoff specifications can be
tackled, namely : regular coupons, a reverse convertible provision and possible participation in the
growth of a single underlying equity asset.
In both models, the way to achieve an analytical solution is to price the structure as a whole in the
form of an option valuation formula that is a function of the contract’s specifications and of fixed
levels of volatility. The obtained formulae can be expected to be a valuable tool both for investors and
issuers in terms of risk management, as they allow to analyze the influence of each variable as well as
to study the global properties of the structure in an accurate and efficient manner. They also provide a
sensible, fast approximation of the structure’s fair price before more general numerical schemes may
be tested, allowing for more stochastic factors.
The article is organized as follows : in Section 2, the autocallable structure priced under the Black-
Scholes model is described in detail; in Section 3, risk management issues are analysed; in Section 4,
the formula under the Black-Scholes model is stated; mathematical proof of this formula is not
reported because it is cumbersome; in Section 5, the formula under the jump-diffusion model with a
stochastic yield curve is stated and the proof of this formula is provided. An appendix discusses the
numerical implementation of the formulae.

2. A flexible autocallable structure

Let us begin by describing in detail the payoff structure that will be valued in a Black-Scholes setting

in Section 4. Several observation dates t1, t2,..., tn  T are set within the product life t 0  0,T  ,
 
where T is the maximum duration of the product (its “maturity”). Two risky assets, S1 and S2 , are

picked. At each observation date tm , 1  m  n , prior to expiry, autocall may occur in two ways :

(i) If the value of S1 at time tm , denoted by S 1 tm  , lies within the range Dm ,U m  , Dm  U m ,
 
then the investor’s initial capital or notional N is redeemed at time tm , along with a coupon rate ym .

Both the range Dm ,U m  and the coupon rate ym are prespecified in the contract. The autocall trigger
 
level Dm is typically the value of the underlying at inception, i.e. S1 t 0  .

4
(ii) If S 1 tm  is greater than the level U m , the note autocalls too, but, instead of yielding a

prespecified coupon rate, it provides the investor with a percentage 1 of the return ratio

S1 tm  / S 1 t 0  if this ratio is greater than the return ratio S 2 tm  / S 2 t 0  or with a percentage 2

of S 2 tm  / S 2 t 0  if the latter is greater than S1 tm  / S 1 t 0  . Thus, the asset S2 was introduced

in the first place to allow for a best-of mechanism in the event of early redemption with a participation
rate.

Besides, if S 1 tm  lies within a range C m , Dm  , C m  Dm , then the note does not autocall but a
 
prespecified coupon rate zm is paid out to the investor; furthermore, all prior coupons that may have

been lost because the coupon barriers C k , 1  k  m , had not been crossed, are also paid out to the

investor at time tm . This amounts to embedding a memory function or snowball effect into the note.

Thus, there are three different barriers at each observation date tm : the autocall trigger level without

participation denoted by Dm , the autocall trigger level with participation denoted by U m , and the

coupon barrier without autocall denoted by C m . The relative positions of these barriers are as follows:

C m  Dm  U m , with Dm being defined as S1 t 0  in most traded contracts. The following table

(Table 1) recapitulates all possible payoffs at any observation date prior to expiry.

Table 1 : Possible payoffs at each observation date tm , 1  m  n prior to expiry, under a Black-
Scholes model

Event Outcome
S 1 tm   C m  the note proceeds to the next observation date

Dm  S 1 tm   C m  the note pays out a prespecified coupon N  z m

U m  S 1 tm   Dm
S 1 tm   U m and S1 tm   S 2 tm 
, foregoing missed coupons are recovered and the note
proceeds to the next observation date
 early exit with coupon : the investor’s initial
capital is fully redeemed, foregoing missed coupons
are recovered and the note pays out a final

prespecified coupon N  1  ym 
 early exit with participation in S1 : the investor’s
initial capital is fully redeemed, foregoing missed
coupons are recovered and the note pays out a final
coupon equal to a prespecified percentage 1 of the
   
ratio S 1 tm / S 1 t 0

S 1 tm   U m and S1 tm   S 2 tm   early exit with participation in S2 : the investor’s
initial capital is fully redeemed, foregoing missed

5
 coupons are recovered and the note pays out a final
coupon equal to a prespecified percentage 2 of the
   
ratio S 2 tm / S 2 t 0

Notes : This table shows the possible payoffs provided by the autocallable structure under consideration at an observation
date prior to expiry

Let us now turn to the profile of the note at expiry or maximum duration tn  T . The safety barrier,

denoted by B , is the European-type barrier under which the notional N will not be redeemed at par

and the reverse convertible mechanism will be activated.For the investor to receive a final coupon Yn ,

two conditions must be met : S1 tn  must lie in the range Dn ,U n  , Dn  U n , and the minimum
 
value reached by S1 at any time between tn1 and tn must not be smaller than a prespecified

American-type down-and-out barrier H . The way in which the investor may participate in the growth

of assets S1 and S2 also differs from previous observation dates, as potential participation in the

growth of S2 is not contingent on S1 tn  being above level Un . Rather, a European-type up-and-in

barrier W is specifically attached to S2 and may trigger participation in the growth of S2 even if

S1 tn  lies below Un . The best-of mechanism is thus less restrictive at expiry than at the previous

observation dates. Finally, a European-type barrier Cn determines whether the memory function or

snowball effect can be activated.

Thus, there are six different barriers at expiry or maximum duration tn  T :

- the safety barrier, denoted by B , under which the reverse convertible mechanism is activated

- an exit barrier denoted by Dn triggering a final coupon payment

- an American-type down-and-out barrier denoted by H and monitored during time interval tn 1, tn  ,
 
that determines the possibility of receiving a final coupon payment but not the possibility of receiving
a final payment through equity growth participation

- an exit barrier denoted by Un triggering a final payment contingent on the value of S1 tn 

- an exit barrier denoted by W triggering a final payment contingent on the value of S 2 tn 

- a coupon barrier Cn triggering the memory function of the note

The following order holds : B  C n  Dn  U n . The barrier W has to stand above B , but its

position can be freely chosen relative to barriers C n , Dn and Un , depending on how large the

6
influence of the best-of mechanism should be. The barrier H has to lie below Dn , but its position

can be freely chosen relative to barriers B and Cn .

The following table (Table 2) recapitulates all possible payoffs at expiry.

Table 2 : Possible payoffs at expiry or maximum duration tn  T , under a Black-Scholes model

Event Outcome
S 1 tn   B  the reverse convertible mechanism is triggered :
only a fraction S 1 tn  / S 1 t 0  of the investor’s
initial capital is redeemed, foregoing missed coupons
are definitely lost and no final coupon payment is
received
S1 tn   B,C  and S t   W
 n  2 n
 the investor’s initial capital is fully redeemed, but
foregoing missed coupons are definitely lost and no
final coupon is received
S1 tn   B,C  and S t   W
 n  2 n
 foregoing missed coupons are definitely lost, but
the investor’s initial capital is fully redeemed and the
note pays out a final coupon equal to a prespecified
percentage 2 of the ratio S 2 tn  / S 2 t0 
S 1 tn   C n , Dn  and S 2 tn   W  the investor’s initial capital is fully redeemed,
foregoing missed coupons are recovered but no final
coupon payment is received
S 1 tn   C n , Dn  and S 2 tn   W  the investor’s initial capital is fully redeemed,
foregoing missed coupons are recovered, and the note
pays out a final coupon equal to a prespecified
percentage 2 of the ratio S 2 tn  / S 2 t0 
S1 tn   Dn ,U n  and inf S1 t   H  the investor’s initial capital is fully redeemed,
tn 1 t tn foregoing missed coupons are recovered, and the note
pays out a prespecified final coupon N  yn
S1 tn   D ,U  and
 n n  inf S1 t   H  the investor’s initial capital is fully redeemed,
tn 1 t tn foregoing missed coupons are recovered, but no final
coupon is received as the American down-and-out
barrier has been breached
S 1 tn   U n and S1 tn   S 2 tn   the investor’s initial capital is fully redeemed,
foregoing missed coupons are recovered and the note
pays out a final coupon equal to a prespecified
percentage 1 of the ratio S 1 tn  / S 1 t 0 
S 1 tn   U n and S1 tn   S 2 tn   the investor’s initial capital is fully redeemed,
foregoing missed coupons are recovered and the note
pays out a final coupon equal to a prespecified
percentage 2 of the ratio S 2 tn  / S 2 t0 

Notes : This table shows the possible payoffs provided by the autocallable structure under consideration at expiry or
maximum duration

These multiple features are introduced in order to span a large variety of possible contracts. Despite its
rather complex payoff function, the fair value of the general autocallable structure under consideration
can be analytically obtained. Forthcoming Proposition 1 provides a formula for any number of

7
observation dates without snowball effect, while Proposition 2 provides a formula for four observation
dates. The reason why a formula is written down in the specific case of four observation dates is three-
fold : first, it is useful to provide a fully explicit example of how Proposition 1 is expanded, in case the
latter might look somewhat terse to the reader at first sight; second, an autocallable note with four
observation dates will serve as the basis for subsequent numerical examples; last but not least,
Proposition 2 includes the pricing of the snowball effect. The reason why the latter is not included in
Proposition 1 is because it seems impossible to come up with a compact formula for it in general, i.e.
without specifying the number of observation dates.

3. Risk management issues

Simpler notes can be designed on the basis of the above general autocallable structure.
Proposition 1 and Proposition 2 nest the valuation of all kinds of simpler structures by putting the
suitable inputs into the formulae :

- notes that do not allow exit through equity participation are valued by setting the parameter A1

equal to one and the parameters A2 and A3 equal to zero

- notes that allow exit through participation in asset S1 only are valued by setting the parameter A2

equal to one and the parameters A1 and A3 equal to zero

- the best-of provision is activated by setting the parameter A3 equal to one and the parameters A1

and the parameters A2 equal to zero

- the snowball effect is activated by setting the parameter A4 equal to one in Proposition 2

All other adjustments are fairly obvious. For example, if you do not want an American barrier to
condition exit with a coupon rate at expiry, just let H tend to zero; if you do not want coupon

payments prior to expiry, just set all the zi parameters equal to zero.

Investors need to be aware of the cost of benefitting from additional opportunities, compared to a
standard contract. Introducing intermediary coupon payments before expiry increases the note’s value
in a relatively straightforward manner, as long as they are fixed coupon payments. It is less easy to
anticipate precisely the effect of exit through equity participation, or the effect of a memory function,
or that of a best-of provision. In Table 3, the prices of four different types of contracts with a growing
number of options are compared. The first one is a plain single-asset note, with no autocall through

participation in the growth of S1 , and no memory function. The second one adds to the first one the

possibility of autocall through participation in the growth of S1 . The third one adds to the second one

a memory function. The fourth one adds to the third one a best-of provision.

8
In all three volatility settings of Table 3, the sharpest increase in value comes from adding the
possibility of autocall through participation in the growth of asset S1 . This effect is more and more

pronounced as volatility rises. The smallest increase in value results from adding a memory function,
but it is not negligible and it would be greater if the coupon barriers were higher. In the low volatility
setting, it is only slightly smaller than the increases in value resulting from adding other options. This
overall behavior is essentially a reflection of the probability of early autocall, as will be seen shortly.
The impact of the best-of provision is substantial, especially in the high volatility setting. Overall, the
introduction of non-standard features thus increases the value of the autocallable note to a large extent
: averaging across all three volatility settings, the price of contract 4 is 23.25% higher than that of
contract 1. There is a major difference between contracts 1, 2 and 3 on the one hand, and contract 4 on
the other hand, as regards the effect of volatility : the value of the former is a decreasing function of
volatility, whereas the value of the latter grows with volatility. We will come back in more detail to
the issue of price sensitivity later in this Section.

Table 3 : Prices of various types of 4-year maturity autocallable notes

Low volatility Medium volatility High volatility

1  12%, 2  15% 1  28%, 2  35% 1  45%, 2  60%

Value of contract 1 102.138322 90.7335053 80.5106173

Value of contract 2 104.916379 102.215849 101.203114
Value of contract 3 106.206245 103.849776 102.725194
Value of contract 4 107.733706 109.940314 115.210895

Notes : This table compares the values of four types of autocallable notes with maximum expiry 4 years and annual
observation dates. Contract 1 is a single-asset note linked to S 1 , with no early exit through participation, and no memory

function. Contract 2 includes the possibility of early exit through participation in the growth of S 1 . Contract 3 includes a
memory function. Contract 4 includes a best-of provision. All reported prices are obtained using forthcoming Proposition 2

with the following inputs :

S1 0  S 2 0  100, r  2.5%, 1  2  0,   35%, D1  D2  D3  D4  100, U 1  U 2  U 3  U 4  115

C 1  C 2  C 3  C 4  90, B  75, W  100, H  80, y1  y2  y3  y4  8%, z1  z2  z 3  z 4  5%, 1  2  1

Contract 1 is valued by setting 1  1 and  2   3   4  0 . Contract 2 is valued by setting   1 and

2

1   3   4  0 . Contract 3 is valued by setting  2   4  1,  1   3  0 . Contract 4 is valued by setting

1   2  0,     1 .
3 4

Whatever the number of options available in an autocallable contract, the issue of the expected
autocall date is fundamental. Indeed, potential automatic early redemption is the specific property of
these products, relative to other index-linked notes, and most investors who choose to put their money
in autocallable structures hope to get their capital back at an early stage. The invested notional can be

9
redeemed as soon as on the first observation date, t1 . In this respect, moderately bullish investors

would be well advised to set a high value for y1 , the coupon rate offered upon exit without

participation in either S1 or S2 at time t1 . They will thus be in a position to benefit from higher

returns than if they invest in standard fixed-income. On another hand, strongly bullish investors had

better trade off a high value of y1 for high levels of participation in the growth of assets S1 and S2 .

They will thus be able to benefit from an amplification of upside market movements, relative to a long
position in the stocks themselves, while being protected from downside risk. Overall, in terms of
contract specifications, the chances to benefit from autocall as soon as the first or maybe the second
observation date depend mostly on the choice of the autocall trigger level. Investors who contemplate
lowering that level must be aware that this will cause the cost of the note to rise. For a given price, this
means that the coupon rates and the participation rates will be lower. The autocall trigger levels may
increase or decrease during the note’s life. The former case is often referred to as a step-up autocall in
the markets, while the latter is called a step-down autocall. When autocall levels vary, they usually do
so in a step-down pattern, so that the investor is more and more likely to kick out as time passes. When
dealing with notes that provide possible early exit with participation in the growth of risky assets, it is

obvious that lowering the U i ’s will increase the likelihood of early redemption with participation,

which may be called the probability effect. But investors should be aware that this positive probability
effect may be offset by the lower expected return received when the note autocalls, which may be
called the return effect, so that the overall effect on the value of the note is ambiguous. Roughly
speaking, the higher the volatility, the more the return effect tends to prevail over the probability

effect, as the likelihood of reaching relatively high U i ’s increases. Table 4 reports the maturity

breakdown on a four-year autocallable note including equity participation and a best-of provision, as a
function of volatility; for simplicity, volatility is supposed to be flat, i.e. the skew and the term
structure are not taken into account. Table 4 shows a couple of noticeable results. First, the most likely
outcome is to autocall on the first observation date, whether volatility is low or high. The probability
of autocall occurring as early as the first or the second observation date increases with volatility. This
increase is not linear, it accelerates when volatility is shifted from the “medium volatility” category to
the “high volatility” one. When volatility is low, the chances of autocall occurring at the fourth and
final observation date are quite substantial, but they go down as volatility gets higher. Again, this
decrease is not linear, it considerably accelerates as volatility moves to the “high volatility” regime, at
a faster pace than the observed increase in the probability of autocall at the first observation date.
Interestingly, autocall at the third observation date is always the least likely outcome, whatever the
volatility input, and its probability of occurrence changes more or less linearly, unlike the probability
of autocall at the first or at the last observation date.

10
 Table 4 : Maturity breakdown of a 4-year autocallable note as a function of volatility under risk
neutrality
Low volatility Medium volatility High volatility

1  12%, 2  15% 1  28%, 2  35% 1  45%, 2  60%

Maturity = 1 year 0.59292711 0.60704343 0.66483994

Maturity = 2 years 0.14079933 0.14795149 0.16333123
Maturity = 3 years 0.06899081 0.0738965 0.08193993
Maturity = 4 years 0.19728274 0.17110857 0.0898889

Notes : This table reports the risk-neutral probabilities that the actual maturity on a 4-year autocallable note providing the
payoff described in Table 1 and Table 2 will be either 1, 2, 3 or 4 years. For simplicity, the volatilities of S1 and S2 are
assumed to be flat. The reported probabilities are obtained using forthcoming Proposition 2 with the following inputs :
1  2  0,  3  1, S1 0  S 2 0  100, r  2.5%, 1  2  0,   35%,
D1  D2  D3  D4  100, U 1  U 2  U 3  U 4  115
Other inputs to Proposition 2 have no relevance here. The probability that the actual maturity is 1 year is obtained by adding
up      . The probability that the actual maturity is 2 years is obtained by adding up      . The
3 5 6 9 11 12

probability that the actual maturity is 3 years is obtained by adding up      . The probability that the actual
16 18 19
maturity is 4 years is equal to one minus the probability that the actual maturity is 1, 2 or 3 years.

For risk management purposes, the probabilities of autocall need to be computed under the physical

measure. In Table 5, the risky assets S1 and S2 include risk premia. Overall, the structure of the

maturity breakdown remains the same as under the risk-neutral measure. There is little difference
between the probabilities of early exit in year 2 and in year 3 in Table 4 and Table 5. However, under
the risk-averse, physical measure, the likelihood that the note will extend to its maximum maturity of
four years is even lower, while the probability that autocall will occur as early as in year 1 is even
higher. Under the high volatility regime, exit at maximum expiry becomes the least likely outcome in
Table 5. Interestingly, the probability of autocall in year 1 does not increase monotonically with
volatility, as it is lower in the medium volatility regime than in the low volatility one. This observation
points out the need to accurately compute the sensitivity of the probability of early exit at any
observation date with respect to volatility. This is easily done by mere differentiation thanks to the
formulae provided in this paper.

Table 5 : Maturity breakdown of a 4-year autocallable note as a function of volatility under risk
aversion
Low volatility Medium volatility High volatility

1  12%, 2  15% 1  28%, 2  35% 1  45%, 2  60%

Maturity = 1 year 0.663790051 0.641161764 0.687729344

11
 Maturity = 2 years 0.147111224 0.153577508 0.167175316
Maturity = 3 years 0.068736885 0.076020303 0.083416916
Maturity = 4 years 0.120361840 0.129240425 0.061678424

Notes : This table reports the probabilities that the actual maturity on a 4-year autocallable note providing the payoff
described in Table 1 and Table 2 will be either 1, 2, 3 or 4 years. The difference with Table 4 is that the returns on assets S
1

and S now include risk premia of 2% and 3% respectively.

2

Finally, two other features should be borne in mind by investors. First, it must be noticed that the best-

of provision plays a more important role at expiry tn than at previous observation dates, as it is not

contingent on S1 being above the U n barrier. Indeed, as long as S 2 tn  trades above the independent

barrier W , the best-of provision can be activated even if S1 tn  trades at levels below Dn , unless

the reverse convertible mechanism is activated (i.e., the reverse convertible provision dominates the
best-of one). One should also pay attention to the fact that the main role of the American down-and-
out barrier is to determine whether a final coupon payment is received in case the note does not
autocall before expiry, so that this barrier will not have a significant impact on the value of the note

unless a sufficiently high value for yn is agreed on in the contract, which makes sense as a

compensation to the investor when early redemption has not taken place.

Now, from the perspective of traders who sell these products, the risks associated with the various
components of the autocallable structure are very similar to those associated with digitals. These
components have positive Delta, and sellers of the option will have to buy Delta of the underlying,
meaning that they will be long dividends, short interest rates and long borrow costs of the underlying.
To tackle the discontinuity risk around the coupon barriers, a barrier shift will usually be applied so
that the resulting shifted payoff over-replicates the initial payoff by the least amount, while making the
Greeks of the new payoff manageable near the barrier. The position in volatility depends on the

coupon barrier levels and on the forward price of the underlying asset S1 . The Vega of the

autocallable’s components is positive if the underlying’s forward price is lower than the trigger level;
otherwise, Vega is negative. Hence, the Vega of the general autocallable structure under consideration
in this paper will depend on the respective weights of the downside and upside components within the

overall structure. Typically, with the Di ' s equal to the starting spot price S 1 0 and the C i ' s being

smaller than the Di ' s , the components of the autocallable structure providing fixed coupon payments

will have a negative Vega. The reverse holds for the components that provide participating payments.
In view of the potentially offsetting effects from the various components, one should always check

12
whether the Vega of the overall structure is positive or negative. Once endowed with the formulae
provided in this paper, this becomes an easy task using analytical or numerical differentiation.
One of the difficulties associated with hedging the sale of an autocallable note is that this product is
sensitive to the skew. Traders can hedge the components providing fixed payments by implementing a

simple roll-over strategy as follows : at each observation date ti , if there is no autocall and the note

proceeds to the next observation date ti 1 , they will take a long position in call spreads expiring at

time ti 1 . These call spreads will be more or less tight or wide depending on how conservative the

trader is, bearing in mind that the tighter the call spread the larger Gamma can become near the
coupon barrier. The catch is that the moneyness of the involved future call options cannot be known at
inception. As a result, the forward implied volatility inputs that might be used to price the structure at

time t0 will generally be wrong. Moreover, all components with expiry greater than t1 are conditional

on not having autocalled at a previous observation date, which introduces a path-dependent element in
the valuation problem. In this respect, the usefulness of the formulae provided in this article is two-

fold : first, they are a function of the skew at time t0 , which, by definition, is observable, i.e. they do

not involve pricing a combination of forward start digitals that would inevitably require the input of
the forward implied volatility surface; second, by giving the risk-neutral conditional probabilities of
autocalling at each observation date, these formulae provide traders with a simple way to accurately
weight the call spreads put in place for each maturity. Table 6 compares the prices of four-year
maturity autocallable notes with and without the possibility of exit through participation under three
stylized regimes of skew. For the sake of simplicity, the at-the-money volatility remains flat up until
maximum duration. In line with most observed data on equity, the slope of the upper part of the skew
curve (i.e. the part to the left of the at-the-money level) is greater than the slope of the lower part of the
skew curve (i.e. the part to the right of the at-the-money level); besides, the “small” skew decreases
more or less linearly with time, while the decrease in the “large” skew with respect to time displays
more curvature. As the downside components of the autocallable structure have negative Vega and the
upside components have positive Vega, the effect of the skew is to lower the value of the autocallable
structure, as the skew increases the volatility of downside components and decreases the volatility of
the upside components with respect to the at-the-money line. The effect of the skew on non-
participating autocallable notes is moderate but non-negligible, and would obviously be bigger under
more extreme skew curves than those chosen here. The effect of the skew on participating autocallable
notes is of great magnitude, so that pricing the autocallable note using flat at-the-money volatility will
largely overestimate the product’s value. It must be noticed that, for some components of the
autocallable structure, it is not obvious to know which point of the implied volatility surface to use,
due to the path-dependent nature of the product and to the many different coupon and early
redemption barriers. The rule followed in this paper is, for every component involved, to pick the

13
point in the implied volatility surface that matches the last barrier and the observation date in the

digital; for example, the digital providing a coupon equal to y 3 upon early exit at time t3 when

S 1 t 3   D ,U  , conditional on autocall not having occurred prior to t , is priced using the point
 3 3  3

with coordinates D3 , t 3  in the S1  volatility surface.

For implied volatility curves that “smile”, i.e. that display higher implied volatilities to the left but also
to the right of the at-the-money level, the previous analysis does not hold anymore. Indeed, the higher
volatility on the out-of-the-money components of the autocallable structure raises their prices since
they are Vega positive, so that the net effect of the smile is ambiguous and depends on the weight of
the upside digitals with respect to the downside digitals, “weight” meaning their respective
contribution to the total note’s value. Clearly, notes without the possibility to autocall through upward
market participation are skew negative, whether the implied volatility curve is skew- or smile-shaped.

Table 6 : Prices of two different kinds of autocallable notes under different regimes of skew

No skew Small skew Large skew

Contract 1 : no possibility 87.4125025 86.76425608 85.425556
of exit through
participation
Contract 2 : possibility of 111.102677 108.869329 105.180181
exit through participation,
including a best-of
provision

Notes : This table compares the prices of four-year maturity autocallable notes with and without the possibility of early exit
through participation under three stylized regimes of skew. The reported prices are computed using Proposition 2. The
“small skew” and “large skew” regimes are defined by the implied volatility surface displayed in Table 7. The “no skew”
regime assumes an identically flat volatility at 1  36% and 2  42% . The term-structure of the risk-free rate is given
by : 1-year at 2.5%, 2-year at 2.75%, 3-year at 3% and 4-year at 3.25%. The contract specifications are as follows:
 
S 1 0  S 2 0  100 ,   35% , t1  1 , t 2  2 , t 3  3 , t 4  4 , C 1  C 2  C 3  C 4  90

D1  D2  D3  D4  100 ,U 1  U 2  U 3  U 4  115 , B  75 ,W  100 , H  80 y1  y 2  y 3  y 4  8%

z 1  z 2  z 3  5% , A 1  A 4  1, A 2  A 3  0 for contract 1, A 1  A 2  0, A 3  A 4  1 for contract 2

Table 7 : Implied volatility surface used to compute the prices in Table 6

Small skew Small skew Large skew Large skew

S1  volatility S 2  volatility S1  volatility S 2  volatility
surface surface surface surface
Coordinates
C , t 
1 1
0.38 0.41

C , t 
2 2
0.375 0.40

C , t 
3 3
0.37 0.39

C , t 
4 4
0.365 0.38

14
 D , t 
1 1
0.36 0.36

D , t 
2 2
0.36 0.36

D , t 
3 3
0.36 0.36

D , t 
4 4
0.36 0.36

U , t 
1 1
0.35 0.39 0.32 0.38

U , t 
2 2
0.35 0.39 0.33 0.39

U , t 
3 3
0.355 0.395 0.34 0.40

U , t 
4 4
0.355 0.395 0.35 0.41

B, t 4
0.40 0.48

W , t 
4
0.42 0.42

Notes : This table reports the implied volatility inputs used to define the small skew and the large skew in Table 6. Only the
points that are needed to compute the autocallable note’s value are reported.

4. Valuation formulae under a Black-Scholes model

This section includes the two formulae denoted by Proposition 1 and Proposition 2 in Section 1 and 2.

Proposition 1

Let S 1 t  and S 2 t  be two geometric Brownian motions driven by standard Brownian motions

B1 t  and B2 t  with correlation coefficient  . Under the risk-neutral measure denoted by   ,

RN

the dynamics of S 1 t  and S 2 t  are given by :

dS1  t    r  1  S1  t   1S1  t dB1  t  (1)

dS2  t    r  2  S2  t   2S2  t dB2 t  (2)

where r is the risk-free rate, 1 and 2 are two constant continuous dividend rates, and

1  0, 2  0 are two volatility inputs assumed to be extracted from an implied volatility surface

across a continuous range of strikes and maturities for both assets S1 and S2 .

Let the function n 1, 2 ,..., n ; 1, 2 ,..., n 1  , i  , i   1, 1 , i   , be defined as
 
follows : for n  1 and n  2 , n is the standard gaussian cumulative distribution function ; for
n  2 , n is given by the following special case of the multivariate standard gaussian cumulative
distribution function :

15
  2 n 1 2
 x
exp   
1
x i 1  i x i  

1 2 n  2
 i 1 2 1   i
2
 
 
n    ...  dx 1dx 2 ...dx n (3)
1   
n 1

2 
n /2
x1  x 2  x n   i
2
i 1

The numerical computation of the function n is explained in the Appendix.

Let the following notations hold :

     
- d  ln  Di  , u  ln  U i  , c  ln  C i  , i  1,..., n  (4)
i  S 0 i  S 0 i  S 0
 1   1   1 

- b  ln  B  , h  ln  H , w  ln  W  , x  ln  1  

      S 0 
(5)
 S 0  S 0  S 0 12  S 0
 1   2   2   2 
- 1  r  1  12 / 2 , 1  r  1  12 / 2 , 12  1  2   12  22  / 2  12 (6)
1  2
- 12  12  212  22 ,   (7)
12
- ˆ12  2  1   12  22  / 2  12 , ˆ1  r  1  12  12 / 2, ˆ2  r  2  22 / 2 (8)

- . is the indicator function, taking value 1 if the argument inside the braces is true and value zero
otherwise
- 1 is a parameter that takes value one if the autocall note does not allow exit through participation
in equity growth, and zero otherwise
- 2 is a parameter that takes value one if the autocall note allows exit through participation in asset
S1 and S1 only, and zero otherwise
- 3 is a parameter that takes value one if the autocall note allows exit through participation in asset
S1 or in asset S2 , i.e. if it includes a best-of provision, and zero otherwise
- the symbol ai ; i  j ...k denotes the sequence of real numbers ai , where i ranges from j to k

Then, the no-arbitrage value, VAUTOCALL , of an autocallable note with n observation dates until

maximum expiry tn , n   , including all the features defined in Table 1 and Table 2 except for the

snowball effect, that is, except for the possibility of recovering foregoing missed coupons at any
observation date, is given by the following formula :

exp r t   N  z  
n 1
VAUTOCALL   i i 1
(9)
i 1


 exp r  ti   N  1  yi   1  2  2  3   3 

16
 2  exp 1  ti   1  N  4   3  exp 1  ti   1  N  5

 3  exp 2  ti   2  N  6 
N  7  1  2   exp r  tn   N  8

3  exp r  tn   N  9  3  exp 2  tn   2  N  10


 exp r  tn   N  1  yn   1  11  2  3   12 

 exp r  tn   N  1  13  2  3   14 
 2  exp 1  tn   1  N  15   3  exp 1  tn   1  N  16

 3  exp 2  tn   2  N  17

where :

1 is the probability, under the money market numeraire, that the coupon N  z i is paid out to the

investor at time ti . It is given by : (10)

   
 d j  1t j c  1ti   d j  1t j d  1ti 
 ; j  1,..., i  1 i1, i ;  ; j  1,..., i  1 i1, i ;
  t 1 ti    t 1 ti 
1  i  1 j  
 i


1 j 

 tj ti1   tj ti1 
   
 t ; j  1,..., i  2 i2,  t i1   t ; j  1,..., i  2 i2,  t i1 
 j 1 i   j 1 i 

2 is the probability, under the money market numeraire, of autocall at time ti when the note does not
allow exit through equity participation. It is given by :
 
 d j  1t j d  1ti 
 ; j  1,..., i  1 i1, i ;
  t 1 ti 
2  i  1 j 

(11)
 tj ti1 
 
 t ; j  1,..., i  2 i2,  t i1 
 j 1 i 

3 is the probability, under the money market numeraire, of autocall at time ti without equity
participation when the note allows equity participation. It is given by : (12)

17
    
 d j  1t j d  1ti   d j  1t j u  1ti 
 ; j  1,..., i  1 i1, i ;  ; j  1,..., i  1 i1, i ;
  t 1 ti    t 1 ti 
3  i  1 j  
 i


1 j 

 tj ti 1   tj ti1 
   
 t ; j  1,..., i  2 i2,  t i1   t ; j  1,..., i  2 i2,  t i1 
 j 1 i   j 1 i 

4 is the probability, under the S1  numeraire, of autocall at time ti with participation in the growth

of S1 in the absence of a best-of provision. It is given by :

 
 d j  1t j ui  1ti 
 ; j  1,..., i  1 i1, ;
  t  t 
4  i  1 j 1 i 

(13)
 tj ti1 
 ; j  1,..., i  2  ,   
 t i2 t i1 
 j 1 i 

5 is the probability, under the S1  numeraire, of autocall at time ti with participation in the growth

of S1 when the contract includes a best-of provision. It is given by :

 
 d j  1t j u  1ti x 12  12ti 
 ; j  1,..., i  1 i1, i , ;
  t  t 12 ti 
5  i 1  1 j 1 i 

(14)
 tj ti 1 
 
 t ; j  1,..., i  2 i2,  t i1,  
 j 1 i 

6 is the probability, under the S2  numeraire, of autocall at time ti with participation in the growth

of S2 when the contract includes a best-of provision. It is given by :

 
 d j  ˆ1t j u  ˆ1ti x 12  ˆ12ti 
 ; j  1,..., i  1 i1, i , ;
  t  t 12 ti 
6 i 1  1 j 1 i 

(15)
 tj ti 1 
 
 t ; j  1,..., i  2 i2,  t i1,  
 j 1 i 

7 is the probability, under the S1  numeraire, of exit at maximum expiry tn under the reverse
convertible provision. It is given by :
 
 d  1ti b  1tn ti 
7  n  i ; i  1,..., n  1 , ; ; i  1,..., n  1  (16)
  t 1 tn ti 1 
 1 i 

18
8 is the probability, under the money market numeraire, of exit at maximum expiry tn without a
final coupon because the autocall barrier has not been reached, in the absence of a best-of provision. It
is given by :

 
 di  1ti d  1tn 1 b  1tn 
 ; i  1,..., n  2 n2, n 1 , ;
  t  t 1 tn 
8  n  1 i 1 n 1 

(17)
 ti t 
 ; i  1,..., n  2 n2,  n 1 
 ti 1 tn 
 
 
 di  1ti d  1tn 1 dn  1tn 
 ; i  1,..., n  2 n2, n 1 , ;
  t  t 1 tn 
n  1 i 1 n 1 

 ti t 
 ; i  1,..., n  2 n2,  n 1 
 ti 1 tn 
 

9 is the probability, under the money market numeraire, of exit at maximum expiry tn without a
final coupon because the autocall barrier has not been reached, when the contract includes a best-of
provision. It is given by :

 
 di  1ti d  1tn 1 b  1tn w  2tn 
 ; i  1,..., n  2 n2, n 1 , , ;
  t  t  t 2 tn 
9  n 1  1 i 1 n 1 1 n 

(18)
 ti t 
 ; i  1,..., n  2 n2,  n 1 ,  
 ti 1 tn 
 
 
 di  1ti d  1tn1 dn  1tn w  2tn 
 ; i  1,..., n  2 n2, n 1 , , ;
  t  t  t 2 tn 
n 1  1 i 1 n 1 1 n 

 ti t 
 ; i  1,..., n  2 n2,  n 1 ,  
 ti 1 tn 
 

10 is the probability, under the S 2  numeraire, of exit at maximum expiry tn with participation in

the growth of S2 , when the contract includes a best-of provision and when S1 is below the

participation barrier Un at time tn . It is given by :

19
  
 di  ˆ1ti d  ˆ1tn 1 b  ˆ1tn w  ˆ2tn 
 ; i  1,..., n  2 n2, n 1 , , ;
  t  t  t 2 tn 
10  n 1  1 i 1 n 1 1 n 

(19)
 ti t 
 ; i  1,..., n  2 n2,  n 1 ,  
 ti 1 tn 
 
 
 di  ˆ1ti d  ˆ1tn 1 dn  ˆ1tn w  ˆ2tn 
 ; i  1,..., n  2  n2, n 1 , , ;
  t  t  t 2 tn 
n 1  1 i 1 n 1 1 n 

 ti t 
 ; i  1,..., n  2 n2,  n 1 ,  
 ti 1 tn 
 

11 is the probability, under the money market numeraire, of exit at maximum expiry tn with a final

coupon N  yn when the note does not allow exit through equity participation. It is given by :


  

  di  1ti h  1tn 1 dn  1tn 

  ; i  1,..., n  2 n2, , ;
   t  t 1 tn 
11  
n 

1 i 1 n 1 


 ti t t
  
  ; i  1,..., n  3 n3,  n 2 n 2, n 1 

  ti 1 tn 1 tn  (20)

  
 
 di  1ti h  1tn 1 dn  2h  1tn 
2 1  ; i  1,..., n  2 n2, , ;
h   t  t  t 
e n  
12 1 i 1 n 1 1 n

 ti tn 2 tn 1 
 ; i  1,..., n  3 n3,  n 2,  
 ti 1 tn 1 tn 
 
 
 di  1ti d  1tn 1 dn  1tn 
 ; i  1,..., n  2 n2, n 1 , ;
  t  t  t 
n  1 i 1 n 1 1 n 

 ti tn 2 tn 1 
 ; i  1,..., n  3 n3,  n 2, 
 ti 1 tn 1 tn 
 
 
 di  1ti d  1tn 1 dn  2h  1tn 
2 1  ; i  1,..., n  2 n2, n 1 , ;
h   t  t 1 tn 
e 1 n  
2
 1 i 1 n 1

 ti t t 
 ; i  1,..., n  3 n3,  n 2 n 2,  n 1 
 ti 1 tn 1 tn 
 

12 is the probability, under the money market numeraire, of exit at maximum expiry tn with a final

coupon N  yn when the note allows equity participation. It is given by :

20
  
 di  1ti h  1tn 1 un  1tn 
 ; i  1,..., n  2 n2, , ;
  t  t  t 
12  11  n  1 i 1 n 1 1 n 

(21)
 ti tn 2 tn 1 
 ; i  1,..., n  3 n3,  n2, 
 ti 1 tn 1 tn 
 
 
 di  1ti h  1tn 1 un  2h  1tn 
2 1  ; i  1,..., n  2 n2, , ;
h   t  t  t 
e n  
12 1 i 1 n 1 1 n

 ti tn 2 tn 1 
 ; i  1,..., n  3 n3,  n 2,  
 ti 1 tn 1 tn 
 
 
 di  1ti dn 1  1tn 1 un  1tn 
 ; i  1,..., n  2  n2
, , ;
  t  t  t 
n  1 i 1 n  1 1 n 

 t t t 
 i
; i  1,..., n  3  n 3
,  n 2
 n 2
, n 1

 ti 1 tn 1 tn 
 
 
 di  1ti d  1tn 1 un  2h  1tn 
21  ; i  1,..., n  2  n2, n 1 , ;
h   t  t 1 tn 
e n  
12 1 i 1 n 1

 ti t t 
 ; i  1,..., n  3 n3,  n 2 n 2,  n 1 
 ti 1 tn 1 tn 
 

13 is the probability, under the money market numeraire, of exit at maximum expiry tn without

receiving a final coupon N  yn because the down-and-out barrier has been crossed in the time

interval tn 1, tn  , when the note does not allow exit through equity participation. It is given by :
 
 
 di  1ti d  1tn 1 dn  1tn 
 ; i  1,..., n  2 n2, n 1 , ;
  t  t 1 tn 
13  n  1 i 1 n 1 
 11
(22)
 ti t 
 ; i  1,..., n  2 n2,  n 1 
 ti 1 tn 
 

14 is the probability, under the money market numeraire, of exit at maximum expiry tn without

receiving a final coupon N  yn because the down-and-out barrier has been crossed in the time

interval tn 1, tn  , when the note allows exit through equity participation. It is given by :
 

21
  
 di  1ti d  1tn 1 dn  1tn 
 ; i  1,..., n  2 n2, n 1 , ;
  t  t 1 tn 
14  n  1 i 1 n 1 

(23)
 ti t 
 ; i  1,..., n  2 n2,  n 1 
 ti 1 tn 
 
 
 di  1ti d  1tn 1 un  1tn 
 ; i  1,..., n  2 n2, n 1 , ;
  t  t 1 tn 
n  1 i 1 n 1 
 12
 ti t 
 ; i  1,..., n  2 n2,  n 1 
 ti 1 tn 
 

15 is the probability, under the S1  numeraire, of exit at maximum expiry tn with participation in

the growth of S1 in the absence of a best-of provision. It is given by :

 
 di  1ti d  1tn 1 un  1tn 
 ; i  1,..., n  2 n2, n 1 , ;
  t  t 1 tn 
15 n  1 i 1 n 1 

(24)
 ti t 
 ; i  1,..., n  2 n2,  n 1 
 ti 1 tn 
 

16 is the probability, under the S1  numeraire, of exit at maximum expiry tn with participation in

the growth of S1 , when the contract includes a best-of provision. It is given by :

 
 di  1ti d  1tn 1 un  1tn x12  12tn 
 ; i  1,..., n  2 n2, n 1 , , ;
  t  t  t 12 tn 
16 n 1  1 i 1 n 1 1 n 

(25)
 ti t 
 ; i  1,..., n  2 n2,  n 1 ,  
 ti 1 tn 
 

17 is the probability, under the S 2  numeraire, of exit at maximum expiry tn with participation in

the growth of S2 , when the contract includes a best-of provision and when S1 is above the

participation barrier Un at time tn . It is given by :

 
 di  ˆ1ti d  ˆ1tn 1 un  ˆ1tn x12  ˆ12tn 
 ; i  1,..., n  2 n2, n 1 , , ;
  t  t  t 12 tn 
17  n 1  1 i 1 n 1 1 n 

(26)
 ti t 
 ; i  1,..., n  2 n2,  n 1 ,  
 ti 1 tn 
 

22
 End of Proposition 1

Although Proposition 1 looks bulky, it may actually be regarded as quite compact, given that it
accomodates any number n of observation dates and that it nests a large variety of possible payoffs.
However, as mentioned earlier, Proposition 1 does not tackle the memory function embedded into
some autocallable notes. So let us now introduce Proposition 2, which expands Proposition 1 in the
special case n  4 , thus providing an easy way for readers to make sure they fully understand how to
expand Proposition 1 for any integer n ; furthermore, Proposition 2 prices the snowball effect.

Proposition 2

Let the assumptions of Proposition 1 hold. Then, the no-arbitrage value, VAUTOCALL , of an

autocallable note with 4 observation dates until expiry t4 , endowed with all the features described in

Table 1 and Table 2, including the snowball effect, is given by the following formula :

VAUTOCALL  exp r  t1   N  z 1  1 (27)


 exp r  t1   N  1  y1   1  2  2  3   3 
2  exp 1  t1   1  N 4
 3  exp 1  t1   1  N  5  3  exp 2  t1   2  N  6
 exp r  t2   N  z 2  7


 exp r  t2   N  1  y2   1  8  2  3   9 
2  exp 1  t2   1 N  10  3  exp 1  t2   1 N  11
 3  exp 2  t2   2  N  12  4  exp r  t2   N  z1  13
 exp r  t 3   N  z 3  14


 exp r  t3   N  1  y 3   1  15  2  3   16 
2  exp 1  t 3   1  N  17   3  exp 1  t 3   1  N  18
 3  exp 2  t 3   2  N  19
 4  exp r  t 3   N  z 1  z 2   20   4  exp r  t 3   N  z 2  21
N  22  1  2   exp r  t4   N  23  3  exp r  t4   N  24
3  exp 2  t4   2  N  25


 exp r  t4   N  1  y4   1  26  2  3   27 

 exp r  t4   N  1  28  2  3   29 
23
2  exp 1  t 4   1  N  30   3  exp 1  t 4   1  N  31
 3  exp 2  t 4   2  N  32
 4  exp r  t 4   N  z 1  z 2  z 3   33
 4  exp r  t 4   N  z 2  z 3   34   4  exp r  t 4   N  z 3  35

The parameter 4 takes value one if the autocall note includes a memory function and value zero

otherwise. The full expansion of all m terms, m  1, 2,..., 35 , is given in Appendix 2. There are

6 out of the 35 m terms in this formula that are not a direct expansion of Proposition 1 for n  4 ,

the role of which is to value the coupon-related memory function; they are 13 , 20 , 21 , 33 , 34

and 35 .

The meaning of these 6 terms can be intuitively defined as follows :

(i) exp  r  t2   N  z1  13 is the risk-neutral value of the digital allowing to recover at time t2

the coupon that was not received at time t1 ;

(ii) exp  r  t3   N   z1  z 2   20 is the risk-neutral value of the digital allowing to recover at

time t3 the coupon that was not received at time t1 and the coupon that was not received at time t2 ;

(iii) exp  r  t3   N  z 2  21 is the risk-neutral value of the digital allowing to recover at time

t3 the coupon that was not received at time t2 ;

(iv) exp  r  t4   N   z1  z 2  z 3   33 is the risk-neutral value of the digital allowing to

recover at time t4 the coupon that was not received at time t1 and the coupon that was not received at

time t2 and the coupon that was not received at time t3 ;

(v) exp  r  t4   N   z 2  z 3   34 is the risk-neutral value of the digital allowing to recover at

time t4 the coupon that was not received at time t2 and the coupon that was not received at time t3 ;

(vi) exp  r  t4   N  z 3  35 is the risk-neutral value of the digital allowing to recover at time

t4 the coupon that was not received at time t3

The full expansion of the 35 m terms in Proposition 2 is as follows :

24
      
 c1  1t1   d1  1t1   d1  1t1 
1  1    1   2  1  
 1 t1   1 t1   1 t1 
     
     
 d1  1t1   u1  1t1   u1  1t1 
3  1    1   4  1  
 1 t1   1 t1   1 t1 
     
   
 u1  1t1 x12  12t1   u1  ˆ1t1 x 12  ˆ12t1 
5  2  , ;  6 2  , ;  
 1 t1 12 t1   1 t1 12 t1 
   
   
d1  1t1 c2  1t2 t1   d1  1t1 d2  1t2 t1 
7  2  , ;   2  , ; 
 1 t1 1 t2 t2   1 t1 1 t2 t2 
   
 
d  1t1 d2  1t2 t 
8  2  1 , ; 1 
 1 t1 1 t2 t2 
 
   
d  1t1 d2  1t2 t   d  1t1 u2  1t2 t1 
9  2  1 , ;  1   2  1 , ; 
 1 t1 1 t2 t2   1 t1 1 t2 t2 
   
 
d  1t1 u2  1t2 t 
10  2  1 , ; 1 
 1 t1 1 t2 t2 
 
 
 d  1t1 u2  1t2 x 12  12t2 t 
11  3  1 , , ;  1 , 
 1 t1 1 t2 12 t2 t2 
 
 
d  ˆ1t1 u2  ˆ1t2 x 12  ˆ12t2 t 
12  3  1 , , ;  1 ,  
 1 t1 1 t2 12 t2 t2 
 
 
 c  1t1 c2  1t2 t 
13  2  1 , ; 1 
 1 t1 1 t2 t2 
 

  
  d1  1t1 d2  1t2 c3  1t 3 t1 t 
14    3  , , ; , 2 

   t 1 t2 1 t 3 t2 t3 
  1 1
 
  
 d  1t1 d2  1t2 d 3  1t3 t1 t  
3  1 , , ; ,  2  
 1 t1 1 t2 1 t 3 t2 t 3  
 
 
d  1t1 d2  1t2 d3  1t3 t1 t 
15  3  1 , , ; , 2 
 1 t1 1 t2 1 t3 t2 t3 
 
  d   t d   t d   t t t 


16  3  1 1 1
, 2 1 2
, 3 1 3
; 1 , 2 
   t 1 t2 1 t3 t2 t3 
  1 1 
 
 d  1t1 d2  1t2 u 3  1t 3 t1 t  
 3  1 , , ; ,  2  
 1 t1 1 t2 1 t 3 t2 t 3  
 

25
  
d1  1t1 d2  1t2 u3  1t3 t1 t2 
17  3  , , ; , 
 1 t1 1 t2 1 t3 t2 t3 
 
 
d1  1t1 d2  1t2 u3  1t3 x12  12t3 t1 t2 
18  4  , , , ; ,  , 
 1 t1 1 t2 1 t3 12 t3 t2 t3 
 
 
 d  ˆ1t1 d2  ˆ1t2 u3  ˆ1t3 x12  ˆ12t3 t1 t 
19  4  1 , , , ; ,  2 ,  
 1 t1 1 t2 1 t3 12 t3 t2 t3 
 
 
c  1t1 c2  1t2 c3  1t3 t1 t 
20  3  1 , , ; , 2 
 1 t1 1 t2 1 t3 t2 t3 
 
  
  c  1t1 c2  1t2 c3  1t3 t t 
21   3  1 , , ; 1 , 2 
   t 1 t2 1 t3 t2 t3 
  1 1 
  
 d  1t1 c2  1t2 c3  1t3 t t  
 3  1 , , ;  1 ,  2  
 1 t1 1 t2 1 t3 t2 t3  
 
 
d  1t1 d2  1t2 d3  1t3 b  1t4 t1 t2 t3 
22  4  1 , , , ; , , 
 1 t1 1 t2 1 t3 1 t4 t2 t 3 t 4 
 

  
  d1  1t1 d2  1t2 d 3  1t 3 b  1t4 t1 t2 t3 
23  
 4  , , , ; , ,  

  1 t1 1 t2 1 t 3 1 t 4 t2 t 3 t4 

  
 
 d1  1t1 d2  1t2 d3  1t3 d4  1t4 t1 t2 t3  
4  , , , ; , , 
 1 t1  t  t  t t2 t 3 t4  
 1 2 1 3 1 4 
 
 
  d1  1t1 d2  1t2 d 3  1t 3 b  1t 4 w  2t 4 t1 t2 t 
24  
5  , , , , ; , ,  3 ,  

   t 1 t2 1 t 3 1 t 4 2 t4 t t t 
  1 1
 2 3 4

 
d1  1t1 d2  1t2 d3  1t3 d4  1t4 w  2t4 t1 t2 t3 
5  , , , , ; , ,  ,  
 1 t1 1 t2 1 t3 1 t4 2 t4 t2 t 3 t4 
 

 
 
  d1  ˆ1t1 d2  ˆ1t2 d 3  ˆ1t 3 b  ˆ1t 4 w  ˆ2t 4 t1 t2 t 
25   5  , , , , ; , ,  3 , 

   t 1 t2 1 t 3 1 t4 2 t 4 t t t 
  1 1
 2 3 4

  
 d  ˆ1t1 d2  ˆ1t2 d 3  ˆ1t3 d 4  ˆ1t4 w  ˆ2t4 t1 t2 t 
5  1 , , , , ; , ,  3 ,   
 1 t1 1 t2 1 t3 1 t4 2 t4 t2 t 3 t 4  
 
  d   t d   t h   t d   t t t t 


26   4  1 1 1
, 2 1 2
, 1 3
, 4 1 4
; 1 , 2 , 3 
   t 1 t2 1 t3 1 t4 t2 t3 t4 
  1 1 
2 1  
d  1t1 d2  1t2 h  1t3 d4  2h  1t4 t1 t t 
h
12
e 4  1 , , , ; , 2 , 3 
 1 t1 1 t2 1 t3 1 t4 t2 t3 t4 
 

26
  
 d1  1t1 d2  1t2 d3  1t3 d4  1t4 t1 t2 t3 
4  , , , ; , , 
 1 t1 1 t2 1 t3 1 t4 t2 t3 t4 
 
 
t3  
2 1
d1  1t1 d2  1t2 d3  1t3 d 4  2h  1t4 t1 t2
h
12
e 4  , , , ; , , 
 1 t1 1 t2 1 t3 1 t4 t2 t3 t4  
 
  
  d1  1t1 d2  1t2 h  1t3 u 4  1t4 t1 t2 t 3 
27  26  4  , , , ; , , 
  1 t1 1 t2 1 t3 1 t4 t2 t3 t4 
  

2 1 
 d1  1t1 d2  1t2 h  1t3 u4  2h  1t4 t1 t2 t3 
h
12
 e 4  , , , ; , , 
 1 t1 1 t2 1 t3 1 t4 t2 t3 t4 
 
 
 d  1t1 d2  1t2 d3  1t3 u4  1t4 t1 t t 
4  1 , , , ; , 2 , 3 
 1 t1 1 t2 1 t3 1 t4 t2 t3 t4 
 
 
t  
2 1
 d  1t1 d2  1t2 d 3  1t3 u 4  2h  1t4 t1 t
h
,  2 ,  3  
12
e 4  1 , , , ;
 1 t1 1 t2 1 t3 1 t4 t2 t3 t4  
 
 
d  1t1 d2  1t2 d3  1t3 d4  1t4 t1 t2 t 
28  4  1 , , , ; , ,  3   26
 1 t1 1 t2 1 t3 1 t4 t2 t 3 t4 
 
  
  d1  1t1 d2  1t2 d 3  1t3 d 4  1t4 t1 t2 t3 
29   4  , , , ; , , 
   t 1 t2 1 t3 1 t4 t2 t 3 t4 
  1 1 
  
 d1  1t1 d2  1t2 d 3  1t 3 u 4  1t 4 t1 t2 t3  
 4  , , , ; , ,   27 
 1 t1 1 t2 1 t 3 1 t 4 t2 t 3 t4  
  

 
d1  1t1 d2  1t2 d3  1t3 u4  1t4 t1 t2 t3 
30  4  , , , ; , , 
 1 t1 1 t2 1 t3 1 t4 t2 t 3 t4 
 
 
 d  1t1 d2  1t2 d3  1t3 u4  1t4 x12  12t4 t1 t2 t 
31  5  1 , , , , ; , ,  3 , 
 1 t1 1 t2 1 t3 1 t4 12 t4 t2 t 3 t4 
 
 
 d  ˆ1t1 d2  ˆ1t2 d3  ˆ1t3 u4  ˆ1t4 x12  ˆ12t4 t1 t2 t 
32  5  1 , , , , ; , ,  3 ,  
 1 t1 1 t2 1 t3 1 t4 12 t4 t2 t 3 t4 
 
 
 c  1t1 c2  1t2 c3  1t3 c4  1t4 t1 t2 t 
33  4  1 , , , ; , , 3 
 1 t1 1 t2 1 t3 1 t4 t2 t 3 t4 
 
 
 
  c1  1t1 c2  1t2 c3  1t 3 c4  1t 4 t1 t2 t3 
34  4  , , , ; , , 

  1 t1 1 t2 1 t3 1 t4 t2 t3 t4 
 
 
 
 d  1t1 c2  1t2 c3  1t3 c4  1t4 t t2 t3  
 4  1 , , , ; 1 , , 
 1 t1 1 t2 1 t3 1 t4 t2 t3 t 4  
 

27
    
c  1t1 d1  1t1   c  1t2 d2  1t2  c3  1t 3 c4  1t 4 t1 t t 
35  4  1 , ,  2 , , , ; , 2 , 3 
  t 1 t1   1 t2 1 t2  1 t 3 1 t 4 t2 t3 t4 
 1 1 

  
where the function 4  11, 12 , 21, 22 , 3 , 4 ; 1, 2 , 3  is the same as in eq. (3) except for the
 
bounds of the x 1  integral which are changed from x 1   (lower bound) and x 1  1 (upper
bound) to x 1  11 and x 1  12 respectively, and for the bounds of the x 2  integral which are
changed from x 2   (lower bound) and x 2  2 (upper bound) to x 2  21 and x 2  22
respectively.

5. Analytical formula for the value of a general autocallable note under a jump-diffusion equity
model correlated with a two-factor stochastic interest rate process

The autocallable payoff structure considered in this section is described in the following Table 8 and
Table 9.

Table 8 : Possible payoffs at each observation date tm , 1  m  n prior to expiry under the
jump-diffusion stochastic interest rate model

Event Consequence
S  tm   C m  the note proceeds to the next observation date
Dm  S  tm   C m  the note pays out a prespecified coupon
M  z m and the note proceeds to the next
observation date
U m  S  tm   Dm  early exit with coupon : the investor’s initial
capital N is fully redeemed and the note pays
out a final prespecified coupon M  ym
S  tm   U m  early exit with participation in S : the
investor’s initial capital M is fully redeemed
and the note pays out a final coupon equal to a
prespecified percentage 1   of the ratio
S  tm  / S  t0 

28
 Table 9 : Possible payoffs at expiry or maximum duration tn  T under the jump-diffusion
stochastic interest rate model

Event Consequence
S  tn   H  the reverse convertible mechanism is
triggered : a fraction S  tn  / S  t0  of the
investor’s initial capital M is redeemed
H  S  tn   Dn  the investor’s initial capital M is fully
redeemed
U n  S  tn   Dn  the investor’s initial capital M is fully
redeemed and the note pays out a prespecified
coupon M  yn
S  tn   U n  the investor’s initial capital M is fully
redeemed and the note pays out a final coupon
equal to a prespecified percentage 1   of the
ratio S  tn  / S  t0 

The new modeling framework is now introduced.

Let W1  t  , t  0  , W2  t  , t  0  and W3  t  , t  0  be three correlated Brownian motions,

whose constant pairwise correlation coefficients are denoted by 1.2 , 1.3 and 2.3 .

The default-free interest rate process  r  t  , t  0  is driven by :

dr  t     t  dt  1dW1  t   2dW2  t  (28)

where 1 and 2 are two positive constants and   t  is a non-random, piecewise continuous function

satisfying a linear growth condition. Eq. (28) is a two-factor Ho-Lee model (1986). The reason that it
was chosen is two-fold : (i) it allows calibration to the observed data by fitting a suitable   t  ; (ii) the

finite-dimensional distributions of  r  t  , t  0  are multivariate normal, which enables to preserve

the analytical tractability of the full model, as will be seen later in the proof of forthcoming
Proposition 3. The latter point should be emphasized as there are other classical interest rate models
that are univariate normal and that can be appropriately calibrated but whose finite-dimensional
distributions are not multivariate normal, such as the two-factor Hull and White model (1994) and its
generalization known as the G2++ model. The reader wishing details about interest rate models and
how they can be fitted to market data may refer to Brigo and Mercurio (2006).
The underlying equity asset process  S  t  , t  0  is driven by :

dS  t 
  r  t     dt    t  dW3  t   I  t dN t 
S  t 
(29)

29
where :
t
(i)   t  is a positive, piecewise continuous, non-random function such that  2  s ds   ,
0

t  0

(ii)  N  t  , t  0  is a Poisson process of constant intensity   0 .

(iii) I  t   U n    n 1 , n


t  , where n  inf t  0, N  t   n    and U n is a sequence of
n

independent, identically distributed random variables taking values in  1,  

Let J n  ln  1  U n  be normally distributed with mean  and variance 2 ; then, the parameter 

is defined by :   exp    2 / 2   1 .

It is assumed that all random processes are defined on a suitable probability space  , , t ,   , in

which  t , t  0  is the smallest   algebra generated by the random variables Wi  s  , N  s  and

U n   n  N  t   , s : 0  s  t , n   , i   1, 2, 3  . It is recalled that a Brownian motion

and a Poisson process relative to the same filtration must be independent (Shreve, 2004).
Thus, the stochastic differential equation for the equity asset price is a jump-diffusion process
extending the seminal Merton model (Merton, 1976) by integrating a stochastic interest rate process
driven by eq. (28). The reason for that extension is that it is essential to factor in the correlation
between sources of randomness in the equity world and in the fixed income world, as explained in the
introduction of this article.
The valuation formula can now be stated under the heading “Proposition 3”.

Proposition 3
Under the model assumptions of Section 5, the fair value, V , at time t0  0 , of the autocallable

contingent claim whose payoff is defined by Table 8 and Table 9, with four observation dates
t1  t2  t3  t4 , is given by :

n1 n2 n3 n4 t1n1  t2  t1   t 3  t2   t 4  t 3 

    n2 n3 n4

V  M  exp  t 4    n1 ! n2 ! n 3 ! n 4 ! (30)

n1  0 n2  0 n3  0 n 4  0

   t1    1  y1  1  2   z1  4  5      t1  1    3 
 
   t    1  y       z          t  1     
 2 2 6 7 2 9 10 2 8 
   t3    1  y 3  11  12   z 3  14  15      t 3  1    13 
 
   t4    1  y 4  16  17   19  20     t 4    1    18  21  


30
  ln U 1 / S 0     t1     t1, t1    t1    t1  
1    
   t1  
 ln  D1 / S 0     t1     t1, t1    t1    t1  
2    
   t1  
 ln  S 0 / U 1     t1     t1, t1    t1    t1  
3    
   t1  
 ln  D1 / S 0     t1     t1, t1    t1    t1  
4    
   t1  
 ln C 1 / S 0     t1     t1, t1    t1    t1  
5    
   t1  
 ln  D1 / S 0     t1     t1, t2    t1    t2  ln U 2 / S 0    t2    t2, t2   t2   t2   t1  
6  2  , ; 
   t1    t2   t2  
 ln  D1 / S 0     t1     t1, t2    t1    t2  ln  D2 / S 0     t2     t2 , t2    t2    t2   t1  
7  2  , ; 
   t1    t2    t2  
 ln  D1 / S 0     t1     t1, t2    t1    t2  ln  S 0 / U 2     t2     t2 , t2    t2    t2    t1  
8  2  , ; 
   t1    t2    t2  
 ln  D1 / S 0     t1     t1, t2    t1    t2  ln  D2 / S 0    t2    t2, t2   t2   t2   t1  
9  2  , ; 
   t1    t2   t2  
 ln  D1 / S 0     t1     t1, t2    t1    t2  ln C 2 / S 0     t2     t2 , t2    t2    t2   t1  
10  2  , ; 
   t1    t2    t2  
 ln  D1 / S 0     t1     t1, t3    t1    t3  ln  D2 / S 0     t2    t2 , t3   t2   t3  
 , , 
   t1    t2  
11  3   
 ln U 3 / S 0     t3     t3 , t3    t3    t3    t1    t2  
 ; , 
   t3    t2    t 3  

 ln  D1 / S 0     t1     t1, t 3    t1    t3  ln  D2 / S 0     t2     t2 , t 3    t2   t 3  
 , , 

12  3 
  t1    t 2  

 ln  D3 / S 0     t3     t3 , t 3    t 3    t 3    t1    t2  
 ; , 
   t3    t2    t 3  


 ln  D1 / S 0     t1     t1, t 3    t1    t 3  ln  D2 / S 0     t2     t2 , t3    t2    t3  
 , , 

13  3 
  t1    t 2  

 ln  S 0 / U 3     t3     t3 , t3    t3    t3    t1    t2  
 ; , 
   t3    t2    t 3  


31
  ln  D1 / S 0     t1     t1, t3    t1    t3  ln  D2 / S 0    t2    t2 , t3    t2   t3  
 , , 

14  3 
  t1    t2  

 ln  D3 / S 0     t3     t3 , t3    t3    t3    t1    t2  
 ; , 
   t3    t2    t3  


 ln  D1 / S 0     t1     t1, t 3    t1    t3  ln  D2 / S 0     t2     t2 , t 3    t2   t 3  
 , , 

15  3 
  t1    t 2  

 ln C 3 / S 0     t3     t3 , t 3    t 3    t 3    t1    t2  
 ; , 
   t3    t2    t 3  

 ln  D1 / S 0     t1     t1, t4    t1    t4  ln  D2 / S 0     t2     t2 , t4    t2   t4  
 , , 
   t    t  
 1 2 
 ln  D3 / S 0     t3     t3 , t4    t3    t4  ln U 4 / S 0     t4     t4 , t4    t4   t4  
16  4  , ; 
   t3    t4  
 
  1   2   3 
 t  t  t 
 , , 
   t2    t3    t4  
 ln  D1 / S 0     t1     t1, t4    t1    t4  ln  D2 / S 0     t2     t2 , t4    t2    t4  
 , , 
   t1    t2  
 
 ln  D3 / S 0     t3     t3 , t4    t3    t4  ln  D4 / S 0     t4     t4 , t4    t4    t4  
17  4  , ; 
   t3    t4  
 
   t1    t2    t3  
 , , 
   t2    t3    t4  
 ln  D1 / S 0     t1     t1, t4    t1    t4  ln  D2 / S 0     t2     t2 , t4    t2    t4  
 , , 
   t    t  
 1 2 
 ln  D3 / S 0     t3     t3 , t4    t3    t4  ln  S 0 / U 4     t4     t4 , t4    t4   t4  
18  4  , ; 
   t3    t4  
 
   t1    t2    t3  
 , , 
   t    t    t 
2 3 4

 ln  D1 / S 0     t1     t1, t4    t1    t4  ln  D2 / S 0     t2     t2 , t4    t2   t4  
 , , 
   t    t  
 1 2 
 ln  D3 / S 0     t3     t3 , t4    t3    t4  ln  D4 / S 0     t4     t4 , t4    t4   t4  
19  4  , ; 
   t3    t4  
 
   t1    t2    t3  
 , , 
   t    t    t 
2 3 4

 ln  D1 / S 0     t1     t1, t4    t1    t4  ln  D2 / S 0     t2     t2 , t4    t2    t4  
 , , 
   t    t  
 1 2 
 ln  D3 / S 0     t3     t3 , t4    t3    t4  ln  H / S 0     t4     t4 , t4    t4    t4  
20  4  , ; 
   t3    t4  
 
   t1    t2    t3  
 , , 
   t    t    t 
2 3 4 

32
  ln  D1 / S 0     t1     t1, t4    t1    t4  ln  D2 / S 0     t2     t2 , t4    t2    t4  
 , , 
   t    t  
 1 2 
 ln  D3 / S 0     t3     t3 , t4    t3    t4  ln  H / S 0     t4     t4 , t4    t4    t4  
21  4  , ; 
   t3    t4  
 
   t1    t2    t3  
 , , 
   t    t    t 
2 3 4


where the following definitions hold :




ti
  ti   exp  r  0  ti  
u ti3   1
2 
 2 1.2   2222 1 


  s dsdu  6


(31)
 0 0 

 ti i  ti i 

 1 1 
  ti   exp  ti     s  ds    nk     2  s ds  2  nk 
 2 

 
(32)
2 0 2  0 

 k 1 k 1 
ti u ti i
1
  ti   r  0  ti     s  dsdu  ti    2  s  ds    nk (33)
0 0
2 0 k 1

  ti  

1/2
  3/2 2  t 3/2   2 
 i  ti  1  2 1.2  ti
 i ti
  
 2
 2  s  ds    2  s  ds  
2 21
 n k  
 3
 1.3    3
 2.3 1   
 k 1  0   0  
  (34)
 2 ti 2 
  s  ds 
 3 1.2 
  
0 

 3
 ti
  ti   
  1
2

1/2
 2 1.2   2222 1 

 (35)
 3 
 

 ti i 1/2

  ti      2  s  ds  2  nk 

(36)

0 k 1 

  3/2   

  ti  1  2 1.2    ti3/2  1  2 1.2  ti
    
   1.3    s  ds  
2
  3   3  

  ti , t j     /    ti    t j  
0

  ti3/2 22 1 ti   t 3/2   

    
    
   
i 2 2 1
 2.3 1    
2
s ds  
  3   3  
  0 
(37)

33
  ti , t j 
  3/2  
 ti
 ti  1  2 1.2  ti
 
 1.3
   s ds 
2
3
 1.3    s  ds  
2




 
  /    t    t  
 
0 0
 t 3/2     i j
 ti
 ti

ti i 
 
 2 s ds  3  2.3 1  2 s ds   32 1.2  2 s ds  2 k
i 2 2 1
  nk 
 2.3 1 
 0  0  0 1 
(38)
2.3  1.2 1.3
2 1  1  1.2
2
, 2.3 1  , 3 1.2  1  1.3
2
 2.3
2
(39)
2 1 1

The functions   b1  and 2  b1, b2 ;   are equal to the univariate and bivariate standard normal

cumulative distribution functions, respectively. For n  2, the function

n  b1,..., bn 1, bn ; 1,..., n 2 , n 1  is defined by eq. (3).

End of Proposition 3.

We now proceed to the proof of Proposition 3.

The proof relies on three lemmas called Lemma 1, Lemma 2 and Lemma 3. Unless stated otherwise,
the notations used here have already been defined in (28) – (39).

Lemma 1
Let r t , t  0  and  S  t  , t  0  be driven by eq. (28) and eq. (29), respectively. Let b be a

positive number. Then, given N  t  , we have :

 ln b / S 0     t  
  S  t   b     
 t 
(40)
 

Proof of Lemma 1 :
Given N  t  , let P    S  t   b     ln  S  t  / S 0   ln b / S 0  

For t  0 , the solution of eq. (29) yields, conditional on N  t  :

 S  t   t t t N t 
1
ln   N  t    r  s  ds  t    2  s ds    s dW3 s    J n
 S  0  
(41)
0
2 0 0 n 1

where each J n is a normal random variable with expectation  and variance 2 , which will be

denoted as follows :

J n    ; 2 

Integrating eq. (28) yields :

34
 t
r t   r  0     s ds  1W1 t   2W2 t  (42)
0

Let us denote by L2 the Hilbert space of random variables with finite variance on  , F ,   . Then,

given W1  t  , it is well known that W2  t  admits the following orthogonal decomposition in L2 (see,

e.g., Shreve, 2004) :

W2  t   1.2W1  t   2 1B2  t  (43)

where B2 is a standard Brownian motion defined on the same probability space as W1 and W2 and

independent of W1 . Furthermore, for any given Brownian motion W  s , s  0  defined on a

suitable probability space and any fixed t  0 , we have:

t t
 t3 
 W s ds   t  s dW s     0;  (44)
 3 
0 0

which implies that, for any given x   and fixed t  0 , we have :

t   t 

   W  s  ds  x     W  t   x 

(45)
 0   3 

t
Thus, as far as the computation of P is concerned, the integral  r s ds can be expanded as
0

follows :

t  1  2 1.2  t 22 1
t u
r  0 t     s  dsdu  W1  t   B2  t  (46)
0 0 3 3

Next, given W1  t  and W2  t  , the random variable W3  t  admits an orthogonal decomposition in

L2 as follows :

W3  t   1.3W1  t   a2B2  t   a 3B3  t  (47)

where a2 and a 3 are real-valued scalars and B3 is a standard Brownian motion defined on the same

probability space as the processes W1 , W2 and W3 and independent of the processes W1 , W2 and B2 .

Note that a 3 must be positive since, by definition of the multivariate normal random vector

W1  t  ,W2  t  ,W3  t   , a 3 t is the standard deviation of W3  t  conditional on W1  t  and W2  t 

 
From the definition of linear correlation and the bilinearity of covariance, we obtain :
cov W2  t  ,W3  t  
2.3 
t

1
t
cov  1.2W1  t  , 1.3W1  t    cov  2 1B2  t  , a2B2  t  
   (48)

35
 2.3  1.2 1.3
 a2   2.3 1
2 1

The real number 2.3 1 is the partial correlation between W2  t  and W3  t  conditional on W1  t  .

Next, from the standard deviation of W3  t  and the stability under addition of a set of uncorrelated

normal random variables, we obtain :


t 1.3
2
 2.3
2
1
 a 32   t  a3  1  1.3
2
 2.3
2
1
 3 1.2 (49)

Now, as the function   t  is non-random, the isometry property of the Ito integral implies that the

t t
random variable    s dW3 s  is normally distributed with mean zero and variance  2 s ds ,
0 0

t  0 .
t
Hence, with regard to the computation of P , the integral   s dW3 s  can be expanded as :
0

t
1

t 0

 2  s  ds W1  t  1.3  B2  t  2.3 1  B3  t  3 1.2  (50)

Since the random variables J n are pairwise independent and are also independent of W1  t  , W2  t 

and W3  t  , the following equality thus holds in distribution for any given t  0 :

 S  t  
ln   N  t 
 S  0  

 3/2 
t u
1
t
 t  1  2 1.2  t

 r  0 t     s  dsdu  t     s ds  N  t    Z1 
2
 1.3   s ds 
2
2 0  3 
0 0  0 

 t 3/2   t  t
 
Z 2    s  ds   Z 33 1.2  2 s ds   N t  Z 4
2 21

 3
 2.3 1  2

(51)
 0  0

where Z1, Z 2 , Z 3 and Z 4 are four uncorrelated standard normal random variables

 S  t  
Consequently, given N  t  , the random variable ln   is normally distributed with mean given
 S  0  

by   t  and standard deviation given by   t  .

36
Lemma 2
Let  r  t  , t  0  and  S  t  , t  0  be driven by eq. (28) and eq. (29), respectively. Let t1 , t2 , t3 ,

t4 be four non-random times such that : 0  t1  t2  t3  t4 . Let b1 , b2 , b3 and b4 be four

positive numbers. Then,

  S  t1   b1, S  t2   b2 , S  t3   b3 , S  t4   b4 

n1 n2 n3 n4 t1n1  t2  t1   t3  t2  t4  t3 

    n2 n3 n4

    exp  t4 
n1 ! n2 ! n 3 ! n 4 !
n1  0 n2  0 n 3  0 n 4  0

 ln b1 / S 0     t1  ln b2 / S 0     t2  ln b3 / S 0     t3  ln b4 / S 0    t4  

 , , , ; 

4 
  t1    t 2    t 3    t 4  

   t1    t2    t3  
(52)
 , , 
   t2    t3    t4  


Proof of Lemma 2 :
Given N  t4  , let P    S  t1   b1, S  t2   b2 , S  t3   b3 , S t4   b4  . The first step is to shift

to log coordinates as in the proof of Lemma 1. Each random variable ln  S  ti  / S 0  , i  1, 2, 3, 4  ,

is continuous. Hence, one can express P as the following quadruple integral :

ln b1 /S 0  ln b2 /S 0  ln b3 /S 0  ln b4 /S 0 

P      dx 4dx 3dx 2dx1

    (53)
  S  t1    S  t2    S  t3    S  t4   
  ln    dx1, ln    dx 2 , ln    dx 3 , ln    dx 4 
  S 0  
 S 0  
 S 0  
 S 0  

Then, by conditioning and using the weak Markov property of the process  S  t  , t  0  , we have :

ln b1 /S 0  lnb2 /S 0  ln b3 /S 0  ln b4 /S 0 

P      dx 4dx 3dx 2dx1

   
  S  t1      S  t2    S  t1   
  ln    dx1     ln    dx 2 ln    dx1 
  S 0     S   S  
(54)
  0   0 
  S  t3    S  t2      S  t4    S  t3   
   ln    dx 3 ln    dx 2     ln    dx 4 ln    dx 3 
  S 0  
 S 0    
  S 0  
 S 0  

Using (51), it is straightforward to show that the correlation between ln  S  ti  / S 0  and

ln  S  t j  / S 0  , i  j , is equal to the standard deviation of ln  S  ti  / S 0  divided by the standard

deviation of ln  S  t j  / S 0  . Thus, from the definition of the bivariate normal distribution, we have :

37
   S  t    S  ti   
  ln 
j  
  dx j ln    dx i 

(55)
  S 0   S 0  

  x    t    t  x    t  2 
1  1  j j i i i  
 exp      
 t j  2  1    ti  /   t j    2  1    ti  /   t j   
   j
t   j
t   t   
2 2 2 2
i 

Hence, (53) becomes :

ln b1 /S 0  t1  ln b2 /S 0  t2  ln b3 /S 0   t 3  ln b4 /S 0 t 4 
  t1    t2   t 3   t 4 

P      dx 4dx 3dx 2dx1

   
 x 2 
(56)
1  x 2 1 3  i 1     ti  /   ti 1   xi  
exp   1     
 2 2 i 1 1  2  ti  / 2  ti 1 
3
  
4 2  1  2  ti  / 2  ti 1   
i 1

Summing over all possible values of the process N  t  in each time interval  ti , ti 1  and multiplying

by their respective probabilities of occurrence yields Lemma 2.

The smaller-dimensional case of   S  t1   b1, S  t2   b2 , S  t3   b3  is nested by Lemma 2 in an

obvious manner.


Lemma 3

Let Y be a normal random variable with mean Y and variance Y2 . Let  X1, X 2 , X 3 , X 4  be a

quadrivariate normal random vector such that :

  X 1  x 1, X 2  x 2 , X 3  x 3 , X 4  x 4 

 x1  X x 2  X x 3  X x 4  X 

 4  1
, 2
, 3
, 4
; X .X , X .X , X .X  (57)
 X X X X 1 2 2 3 3 4

1 2 3 4

where X and X are the mean and the standard deviation of Xi , respectively, and X .X is the
i i i j

correlation coefficient between Xi and X j ,   i, j   2 . Then, if the random vector

Y , X1, X 2 , X 3 , X 4  is multivariate normal, we have :

 

E  exp Y   X x ,X x ,X x ,X x  
 1 1 2 2 3 3 4 4 

38
  x1  X  X .Y X Y x 2  X  X .Y X Y 


1 1 1
, 2 2 2
, 
  X1 X2 
 
 Y2   x 3  X 3  X 3 .Y X 3 Y x 4  X 4  X 4 .Y X 4 Y 

 exp  Y     , ;  (58)
 2  4  X X 
 3 4 
 X .X , X .X , X .X 
 1 2 2 3 3 4 
 


where X .Y is the correlation coefficient between Xi and Y

i

Proof of Lemma 3 :
Let fY ,X ,X
1 2 ,X 3 ,X 4
 y, x 1 , x 2 , x 3 , x 4  denote the joint density of the standardized random vector

Y  Y Xi  X
Y , X1, X 2 , X 3 , X 4  , where Y  and Xi  i
  Y X
.
i

Then,

E  exp Y   X x ,X x ,X x ,X x   (59)
 1 1 2 2 3 3 4 4 

x1  X x 2 X x 3 X x 4 X

1 2 3 4

 X X X X
1 2 3 4

      exp  y  fY ,X ,X  y, z1, z 2, z 3, z 4 dydz 4dz 3dz 2dz1

1 2 ,X 3 ,X 4
y  z1  z 2  z 3  z 4 

Following the same approach as in the proof of Lemma 1, a little algebra shows that the normal
random variables Xi , i   1, 2, 3, 4  , admit the following orthogonal decompositions :

X1  X .YY  X Y Z1 (60)
1 1

X 2  X .YY  X .X Y Z1  X Y .X1 Z 2 (61)

2 1 2 2

X 3  X .YY  X .X Y Z1  X Y .X1 Z 2  X Y .X1 .X2 Z 3 (62)

3 1 3 2 .X 3 3

X 4  X .YY  X .X Y Z1  X Y .X1 Z 2  X Y .X1 .X2 Z 3  X Y .X1 .X2 .X3 Z 4 (63)

4 1 4 2 .X 4 3 .X 4 4

where the Z i ' s are pairwise independent standard normal random variables that are all independent of

Y and where the following definitions hold :

X  X X
a .Xb a .Y b .Y
X Y  1  X2 , X  , X  1  X2  X2 (64)
a .Xb Y Y .Xa
a a .Y X Y
b b .Y a .Xb Y
a

X  X X  X Y Xa .Xc Y
b .Xc b .Y c .Y a .Xb
X  (65)
b .Xc Y .Xa
X
b Y .Xa

X Y .Xa .Xb  1  X2  X2  X2 (66)

c b .Y a .Xb Y b .Xc Y .Xa

39
X
c .Xd Y .Xa .Xb

X
1
 X .X
c d
 X
c .Y
X
d .Y
 X
a .Xd Y Xa .Xc Y  X
b .Xd Y .Xa Xb .Xc Y .Xa  (67)
c Y .Xa .Xb

X Y .Xa .Xb .Xc  1  X2  X2  X2  X2 (68)

d b .Y a .Xb Y b .Xc Y .Xa c .Xd Y .Xa .Xb

Thus, we have :

X1 Y   X .YY ; X  1 1Y
 (69)

 X1  X .YY 

X 2 Y , X1    X .YY  X .X Y 1
; X Y .X  (70)
 2 X Y 1 
 
2 1 2
1

X 3 Y , X1 , X 2 (71)

 X1  X .YY X .X Y .X  X1  X .YY  
   ; 
  Y   1
 3 2 1
 X 2  X2 .YY  X 2 .X1 Y
1
 
   X 3 .Y X 3 .X1 Y
X Y X Y .X    
 1 2 1 X1 Y

 X Y .X .X 
 3 1 2

X 4 Y , X 1, X 2 , X 3 (72)

 X1  X .YY X .X Y .X  X1  X .YY 
 
   X .YY  X .X Y 1
 4 2 1 
X
 2   Y   1

 4   
X2 .Y X2 .X1 Y

 X2 Y .X1  
4 1
X1 Y X1 Y
X .X Y .X .X   X1  X .YY X .X Y .X   X1  X .YY 

 4 3 2  1  
 X 3  X 3 .YY  X 3 .X1 Y  3 2  X 2  X2 .YY  X2 .X1 Y  ;
1 1 1

X Y .X .X  X Y X Y .X  X Y 



3 1 2 1 2 1 1

X Y .X .X .X
4 1 2 3

From (69) - (72), one can derive fY ,X ,X

1 2 ,X 3 ,X 4
 y, z1, z2 , z 3 , z 4  as the following product of conditional
densities :
fY ,X ,X
1 2 ,X 3 ,X 4
 y, z 1 , z 2 , z 3 , z 4  (73)

 fY  y  fX Y  z1 y  fX Y , X1  z2 y, z1  fX Y , X1 , X 2 z3 y, z1 , z2  fX Y , X1 , X 2 , X 3  z4 y, z1 , z 2 , z 3 
1 2 3 4

Substituting (73) into (59) and performing the necessary calculations then yields Lemma 3.


The proof of Proposition 3 can now ensue. According to the no-arbitrage theory of valuation in a
complete market (Harrison and Pliska [1981] ), the price of the contingent claim under consideration is
equal to the expectation of its discounted payoff under the equivalent martingale measure. But the
market here is incomplete, due to the introduction of jumps. This raises the question of the choice of a
relevant equivalent martingale measure. A necessary condition for the discounted price process
  t  
 exp   r  s  ds  S  t , t  0  to be a martingale is to set the drift coefficient in the dynamics of
   
 
  0 

40
S t , t  0  equal to r  t    . This can be shown in exactly the same way as when the riskless

rate process r t , t  0  is constant (see, e.g., Lamberton and Lapeyre, 1997), so the details are

omitted. According to the classical argument by Merton (1976) that jump risk is diversifiable and
therefore not rewardable with excess return, this condition should be both necessary and sufficient. If
we consider this argument to be true, then the dynamics of  S  t  , t  0  can only be given by eq.

(29). However, one must bear in mind that this argument is debatable, as empirical evidence suggests
that there are industry wide shocks and even country wide shocks that are not easily diversifiable, so
that the possibility of perfect hedging remains theoretical. The literature on mean-variance hedging
(Schweizer, 1992) and quantile hedging (Föllmer and Leukert, 1999) can be consulted for alternative
approaches.
Denoting by  the probability measure under which eq. (29) holds, and assuming that this is the
correct pricing measure, then the autocallable payoff defined in Table 8 and Table 9 implies that, at
each fixing date t j , j  1, 2, 3  , prior to maximum expiry t4 , the following expectations must be

computed :
  tj  
   
   exp 
  r  s  ds   M   1  y j   S  t D ,...,S t D ,D S t U  N t j   (74)
  0 
 1 1 j 1 j 1 j j j 
  
  tj  
   S t j  
   exp 
  r  s  ds   M   1      S  t D ,...,S t D ,S t U  N  t j  
S 0
(75)
  0 
 1 1 j 1 j 1 j j 
  
  tj  
   
   exp 
  r  s  ds   M  z j  S  t D ,...,S t D ,C S t D  N  t j   (76)
  0  

1 1 j 1 j 1 j j j

  
At maximum expiry t4 , the following expectation is also required :

  t4  S  t4  
  
   exp   r  s  ds   M   S  t D ,S  t D ,S t D ,S t H  N  t4  
  
(77)
  0 
 S 0 1 1 2 2 3 3 4

 
It is clear that the valuation problem comes down to computing two kinds of expectations, denoted by
E1 and E2 ,   i, j   1, 2, 3, 4  , i  j :

  tj  
   
E1     exp 
    
r s ds     S  ti     ci   N  j  
t (78)
  0 
 
  tj  
   S  t j  
E2     exp 
  r   
s ds      S  ti     ci   N  t j 
 S  0 
(79)
  0     
  

41
where the symbol  .  inside the indicator functions is a vector notation and where the components of

 ci  , 0  i  j , are positive constants

 
Thus, the value of the contingent claim under consideration can be obtained by applying Lemma 3,
provided the right parameters are entered into the formula.

 S  ti  
tj

Let Y   r  s  ds and let X  ti   ln   N  ti  . From eq. (46), the expectation of Y is

 S  0  
0

given by :
tj u

r  0  t j     s dsdu (80)
0 0

while the standard deviation of Y is given by   t j  .

From eq. (51), the expectation and standard deviation of X  ti  are given by   ti  and by   ti  ,

respectively. Hence, the covariance between Y and X  ti  is given by the numerator in   ti , t j  .

tj
 S  t  
Now, let Y   r  s  ds  ln   N  t j  . Then, the expectation and the variance of Y are
j 

0
 S  0  

given, respectively, by :
tj j
1
E Y   t j    2  s  ds    nk (81)
2 0 k 1

tj j
var Y     2  s  ds  2  nk (82)
0 k 1

Hence, the covariance between Y and X  ti  is given by the numerator in   ti , t j  .

Adding all required E1  type and E2  type expectations together and then summing over all

possible values of the process N  t  in each time interval  ti , ti 1  weighted by their respective

probabilities of occurrence, one can obtain Proposition 3.

It must be emphasized that the validity of the proof depends on the multivariate normality of the
random vector Y , X1, X 2 , X 3 , X 4  in Lemma 3. It is well known that the univariate normality of each

marginal does not imply the multivariate normality of the joint density (see, e.g., Tong [1989]).

42
References

Alm, T., Harrach, B., Harrach, D., Keller, M., A Monte Carlo pricing algorithm for autocallables that
allows for stable differentiation, to appear in Journal of Computational Finance

Black, F., Scholes, M., The Pricing of Options and Corporate Liabilities, The Journal of Political
Economy, Vol. 81, No. 3 (May - Jun., 1973), pp. 637-654

Bouzoubaa, M., Osseiran, A. (2010) Exotic Options and Hybrids, Wiley Finance

Broadie, M., Yamamoto, Y. (2005), A double-exponential fast Gauss transform algorithm for pricing
discrete path-dependent options, Operations Research, vol. 53 (5), 764-779

Brigo, D., Mercurio, F. (2006), Interest Rate Models – Theory and Practice, Springer Finance

Davis, P.J., Rabinowitz, P. (2007), Methods of Numerical Integration, Dover Publications

De Weert, F. (2008), Exotic Options Trading, Wiley Finance

Deng G., Mallett, J., McCann, C. (2011), Modeling autocallable structured products, Journal of
Derivatives and Hedge Funds, vol.17, 326-340

Föllmer, H., Leukert, P. (1999), Quantile hedging, Finance and Stochastics, 3, 251-273

Fries C.P., Joshi, M.S. (2011), Perturbation stable conditional analytic Monte-Carlo pricing scheme
for autocallable products, International Journal of Theoretical and Applied Finance, vol. 14 (02), 197-
219

Gatheral, J. (2006), The Volatility Surface, Wiley Finance

Harrison, J.M., Pliska, S. (1981) Martingales and stochastic integrals in the theory of continuous
trading, Stochastic Processes and their Applications, 11, 215-260

Hull, J., White, A. (1994), Numerical procedures for implementing term structure models II : two-
factor models, The Journal of Derivatives, 2, 37-47

Ho S.Y., Lee S.B. (1986), Term structure movements and pricing interest rate contingent claims,
Journal of Finance, vol. 41 (5), 1011-1029

Kamtchueng, C. (2011), Autocallable more American than European ?, Working Paper, Barclays
Capital

Lamberton, D., Lapeyre, B. (1997), Introduction to Stochastic Calculus Applied to Finance, Chapman
& Hall, 1997

Merton, R.C. (1976), Option pricing when underlying stock returns are discontinuous, Journal of
Financial Economics, 3, 125-144

Schweizer, M. (1992), Mean variance hedging for general claims, Annals of Applied Probability, 2 (1),
171-179

Shreve, S.E. (2004), Stochastic Calculus For Finance II Continuous-time Models, Springer Finance

Tong, Y.L. (1990), The multivariate normal distribution, Springer

43
Appendix

From a numerical point of view, the valuation formulae presented in this paper raise the question of

the computation of the function n . As the number n of possible exit dates increases, so does the

dimension of numerical integration. The latter can be reduced by using the following identities :

(i) when there are three observation dates, the actual numerical dimension of the function 3 can be

brought down from 3 to 1 by using :

2

exp x 22 / 2     
 1  1x 2   3  2x 2 
 3 1, 2 , 3 ; 1 , 2    2
1   
 1  2  1  1  2  2
 dx (83)
x 2   1   2 

(ii) when there are four observation dates, the actual numerical dimension of the function 4 can be

reduced by a factor of two by using :

3 2x2
2 122

 4 1 , 2 , 3 , 4 ; 1, 2 , 3     (84)

x 2  x 3 

 x 2  x 2 
 3 
exp  
2

 2     
  1  1x 2   4  3 1  2 x 3  3 2x 2 
2

1     dx 2dx 3
2  1  2  1  1   2 
 1   3 
(iii) more generally, when there are over four observation dates, the actual numerical dimension of the

function n can always be reduced by a factor of 2 by using :

n 1 , 2 ,..., n 1 , n ; 1 ,..., n 2 , n 1  (85)

2 3 n 1

  x ... 
x2  3  x n 1 

 x2 
2
 2 1 1
exp      
2
 x   x ...  x   x
 2 2 1   2  3 2 2
2 1  n 2 
2 n 1 n 2 n 2 
 2 

1    2
n 2 n 2

 2
i
2
i 2

   
 1  1x 2   n  n 1x n 1 
1     dx dx ...dx n 1
 1  2  1  1  2  2 3
 1   n 1 

44
The functions to be integrated in (83)-(84) are so smooth that a mere 16-point Gauss-Legendre
quadrature with a lower bound cut off at 5 is enough to achieve a high level of accuracy for pricing
purposes. When 2 is close to 1, the upper bound of the inner integral in (84) becomes very

« large ». To avoid numerical errors, it is safe to prespecify a maximum value for all possible multiple
integral endpoints. Given the standard normal nature of the integrands, bounding at an absolute value

of 5 will entail negligible loss of accuracy. The fact that the arguments in the 1 x  functions in (83)-
 
(85) may rise sharply when the i coefficients, i  1,..., n  1 , are close to  1 , is not a cause for

concern since these functions are, by definition, bounded at 1 and have already reached their

maximum to 109 accuracy at x  5 .

The quality of numerical integration using (83) – (85) was tested up to six observation dates by
applying a mere 16-point Gauss-Legendre quadrature rule. A first set of tests were performed using
analytical benchmarks which consisted of products of univariate normal cumulative distribution
functions, by assuming that all correlation coefficients were equal to zero. In every dimension from 3
to 6, some 500 integral convolutions were computed by means of Proposition 1 and Proposition 2 with

randomly drawn parameters. The results always matched the analytical benchmarks to at least 108
accuracy. Then, option prices with non-zero, arbitrary values of the correlation coefficients were
computed and compared with numerical values obtained by Monte Carlo simulation using antithetic
variates and the Mersenne Twister random number generator. In terms of efficiency, it always took
less than one second to compute prices of 4-year autocallable notes with yearly observation dates by
means of Proposition 2 on a mainstream commercial PC, which is very fast in absolute terms and
dramatically more efficient than simulation techniques, which required from four to ten minutes to
yield accurate approximations in dimensions 5 and 6. Moreover, the efficiency gains are even greater
when it comes to the computation of the option sensitivities or greeks. A clear pattern of convergence
of the Monte Carlo estimates to the analytical values was noticed, as more and more simulations were
performed. For autocallable notes with 4 observation dates, 400,000 simulations along with a time

discretization factor of 8 quotations per business day in the time interval tn 1, tn  were necessary to
 
achieve 104 convergence, and it took 1,000,000 simulations to achieve 105 convergence.
Thus, in moderate dimension, the analytical formulae provided by Proposition 1 and Proposition 2
give very efficient and accurate numerical results, while the computational time required for a Monte
Carlo simulation to return reasonably precise approximations is clearly not satisfactory for practical
purposes. It must be stressed that the valuation of the majority of autocallable notes does not involve
more than moderate dimension, as most traded products include between four and eight possible exit
dates, being typically annual observation dates on 4 to 8 – year maturity notes. In higher dimension, as
more and more exit dates might be added, the quality of a plain Gauss-Legendre implementation of

45
(85) will deteriorate. One solution, then, would be to implement an adaptive Gauss-Legendre
quadrature, in order to control the error resulting from numerical integration. Combining this with a
Kronrod rule will reduce the number of required iterations (Davis and Rabinowitz, 2007).
Alternatively, another solution is to notice that the function  that needs to be computed in increasing
dimension has the attractive feature that it matches the special structure of Gaussian convolutions
handled by the efficient Broadie-Yamamoto algorithm (Broadie and Yamamoto, 2005).
Regarding Proposition 3, the quadruple infinite series can be truncated in a simple manner by setting a
convergence threshold such that no further terms are added once the difference between two
successive finite sums of the quadruple series becomes smaller than that prespecified level. The speed
of convergence is inversely related to the intensity of the Poisson process. Numerical experiments
were carried out that showed that, when  was equal to 0.1 , only three or four terms had to be
computed in each summation operator for the percentage variation in two successive approximations
of the quadruple infinite series to decrease below 0.0001%,; when  was equal to 1 , seven terms were
required to achieve the same level of convergence. It should be pointed out that the rate of decay of the
denominator in (30) is “fast”, so that numerical errors may arise for “large” values of n1 , n2 , n 3 and

n 4 . Fortunately, finance applications do not require the level of accuracy that might raise such

concerns, i.e. a small number of terms in the infinite series will suffice to achieve adequate
convergence for all practical purposes. The total computational time required by Proposition 3, on a
mainstream commercial PC, ranges between 0.8 and 18 seconds, depending on the number of terms to
be computed in each summation operator (from 4 to 8) . Such an efficiency is in sharp contrast with
the slowness of a Monte Carlo simulation approximation. For only a moderate level of accuracy to be

achieved ( 103 divergence from the analytical formula) , the latter requires between one and two
hours of computational time, depending on the level of  (from 0.1 to 1), and assuming that the

maximum expiry of the autocallable structured product under consideration is t4  four years. It must

be emphasized that, compared with a standard geometric Brownian motion model with constant
parameters, the introduction of additional stochastic factors such as jumps and a multi-factor term
structure of interest rates heavily deteriorates the quality of Monte Carlo simulation approximations, as
the payoff under consideration is path-dependent and requires time discretization. Hence the
usefulness of accurate, quickly computed analytical formulae.

46