A market model for inflation

N. BELGRADE∗, E. BENHAMOU†& E. KOEHLER ‡§

January 2004.

Abstract
The various macro econometrics model for inflation are helpless when it comes to the pricing of inflation
derivatives. The only article targeting inflation option pricing, the Jarrow Yildirim model [7], relies on non
observable data. This makes the estimation of the model parameters a non trivial problem. In addition,
their framework do not examine any relationship between the most liquid inflation derivatives instruments:
the year to year and zero coupon swap. To fill this gap, we see how to derive a model on inflation, based on
traded and liquid market instrument. Applying the same strategy as the one for a market model on interest
rates, we derive no-arbitrage relationship between zero coupon and year to year swaps. We explain how to
compute the convexity adjustment and what relationship the volatility surface should satisfy. Within this
framework, it becomes much easier to estimate model parameters and to price inflation derivatives in a
consistent way.
• Keywords: inflation index, forward, zero-coupon, year-on-year, volatility cube, convexity adjustment.
AMS Classification 60G35

∗ CDC IXIS-CM R&D and Maison des Sciences Economiques. Adress: CDC IXIS 47 Quai d’Austerlitz 75013 Paris, France.
Maison des Sciences Economiques, University of Paris 1 Panthéon-Sorbonne, 116-118 Boulevard de l’Hôpital 75013 Paris, France.
E-mail: [email protected].
† CDC IXIS-CM R&D, 47 Quai d’Austerlitz 75013 Paris, France. E-mail: [email protected]
‡ Risk Department CDC IXIS (26-28 Rue Neuve Tolbiac, 75658 Paris Cedex 13, France, ( +33 1 58 55 59 68) and Associated

Professor at Paris I, Sorbonne University (Maison des Sciences Economiques, University of Paris 1 Panthéon-Sorbonne, 116-118
Boulevard de l’Hôpital 75013 Paris). Email [email protected]
§ The ideas expressed herein are the authors’ ones and do not necessarily represent the ones of CDC IXIS CM.

1
1 Introduction
The standard approach for modelling inflation is based on econometrics models. Their aim is to forecast infla-
tion rate, provided a time series of data. Using sophisticated version of the so-called "Taylor" rule, economists
show how to relate inflation rate to various macro economic indexes such as short term interest rates and
monetary policy target.
When it comes to pricing inflation derivatives, this framework is helpless for many reasons:

• First, it does not provide any information concerning the hedging strategy making this approach of poor
usage for derivatives trading desk.
• Second, it does not offer any relationship between the various traded instruments. These relationships
are crucial to provide a model consistent with traded securities.
• Third, it uses discrete time modelling. This makes it not easy to tackle complex options where the set
up is based on continuous time modelling.

Surprisingly, there is not much in the literature on option pricing on inflation derivatives. The only paper
by Jarrow and Yildirim [7] uses the interest rate curve as a starting point. The authors model the inflation rate
as an exchange rate between the nominal and real zero-coupon bonds. Their key assumptions are deterministic
volatilities and non zero correlation between the different factors. Using no-arbitrage relationship, they derive
a model similar to a 3 factor HJM model. However, this model has the major drawback to use non observable
parameters. In order to infer the inflation rate, one needs to fit a model on the real interest rate curve, which
is even harder to estimate than the inflation rate itself. A second drawback is to provide no link between
zero-coupon and year-on-year products.
The two disadvantages of the Jarrow and Yildirim model are precisely the motivation of our model. Adapt-
ing the concept of market model, we explain how to use consistent information between the zero-coupon and
the year-on-year swap market. The first result concerns volatility information. The consistent relationship for
the volatility market is first examined in the general framework of a market model. We then compute explicitly
this relationship in the case of various model assumptions like Black-Scholes, homogeneous and Hull and White
volatilities. We then see how to compute convexity adjustment using the market model.

2 Primer on inflation modelling

2.1 Product overview
Over the last 3 years, the inflation swap market had been exploding with monthly transaction volume around
100 millions Euro in 2001, 500 millions in 2002 and 1500 in 2003 (source ICAP). In 2003, 56% of the transac-
tions concerned maturity below 7 years, with 28% between 1 and 4 years. The potential explanation of this
tremendous growth are numerous, ranging from
- interest for competitive inflation product (due to anticipated deflation)
- bigger liquidity provided by raising government inflation linked bond issue
- interest from corporate to issue inflation linked debt (DEXMA,RFF,...)
- ability to offer capital guaranteed structure guaranteed not in notional but real term.
The two liquid instruments are the zero-coupon swap and the year-on-year swap. More precisely:

• a zero-coupon swap (in its payer version) pays the inflation return

CP I(T )/CP I(0) − 1.

T
versus receiving a pre-agreed zero-coupon rate (1 + Zc) − 1. By far, zero-coupon swaps are the most
liquid instruments.
• the year-on-year swap (also referred to as year-on-year, or annual swap) pays in its payer annual form
version the annualized CPI return. At time Ti + 1, the inflation leg pays CP I (Ti+1 ) /CP I (Ti ) − 1 versus
receiving a fix leg paying S.
• Last but not least, inflation bonds pays the compounded inflation return over time CP I(T )/CP I(0)

2
2.2 Modelling issue
To model inflation, one may think to use the numerous recent models in time series analysis targeting inflation
using discrete time modelling. One can find equivalents in continuous time but these models remain inefficient
for the evaluation of the inflation linked products. The major challenge comes from the difference of probability
measures between the historical and risk neutral ones. Econometric models are derived under historical
probability while option pricing requires to use the risk neutral probability 1 . This therefore prevents us from
using econometric models.
A first sight solution may be to use an adaptation of interest modeling. However, the inflation market offers
some additional challenging features:

• multi-curve environment: because of the inability to lock in a inflation zero-coupon with its notional
compounded by the inflation return, static replication of the year-on-year curve from the zero-coupon
one is impossible. Hence the year-on-year curve has to account for the additional convexity adjustment.
• correlation modelling: inflation should be rigorously connected to interest rate as the correlation structure
between forward CPI and interest rate has to be used for the convexity adjustment.
• multi-asset pricing dimension: because of the correlation between interest rates and inflation rates.

2.3 Modelling assumptions

Ignoring for now the multi-asset pricing dimension between interest rates and inflation, we examine how to
provide consistent information between the various inflation markets. Our modelling target is to provide a
model that can be

• simple enough to have only a few parameters.

• robust enough to replicate market prices.

At first sight, the CPI can be modelled as:

- either n sampling of one single process observed at different times
- or a single sampling of n different processes observed each at one time.
This fundamental difference can be related to the interest rate markets. We could think of the Libor or
swap rate as one single instrument observed at different time (assumption made when doing a pricing of
swaption and or cap in Black-Scholes) or we could see forward Libor rates as independent rates. The latter is
the approach taken in the Libor market models (also referred to as forward rate models or BGM or Jamshidian
models). Obviously the latter could also be seen as an extension of the first methodology but with a much richer
information on the correlation structure. From an econometric point of view, the first assumption (respectively
the second one) corresponds to a heteroscedastic process without (respectively with ) autocorrelation on errors.
Typically, inflation structures are based on CPI data fixing at various dates denoted by (Ti )1≤i≤n . Let us
denote by CP I (s, Ti , Ti + δ i ) the CPI rate observed at time s that fixes at time Ti and applying for a period
δ i . Typically, the tenors δ i are the same and are all equal to 1 month. In the following, we will drop the index i
and denote by δ. From a concrete point of view, δ of 1 month means monthly CP I data. Note that compared
to Libor modelling, the tenor δ is slightly different in the sense that it is not an interest period. However,
writing the CPI rate like this shows the similarities between the two instruments. In order to simplify even
further notations, we will write CP I (s, Ti ) dropping the third term in the parenthesis.

A simple but rich enough framework is to assume a market model for inflation where the forward CPI return
is modelled
³ as a diffusion
´ with a deterministic volatility structure. For this, we consider the filtered probability
space Ω, A, (Ft )t≥0 , P where P is the historical probability, and (Ft )t≥0 the natural filtration2 generated by
n¡ ¢ o
the standard multi n + 1 dimensional Brownian motion W i , B (t)1≤i≤n , t ≥ 0 , with correlation matrices
given by Ξ with the various terms of this matrix are given by
( ­ ®
d W.i , W.j (t) = ρInfi,j dt,
­ i ® (1)
d W. , B. (t) = ρB,I i dt.

1 The passage between the two is made by the determination the market risk premium which is a parameter which still

complicates the estimate

n¡ of the models. o
2 ∀t ≥ 0, F = σ
¢
t W i (s) 1≤i≤n , 0 ≤ s ≤ t .

3
 Where {B (t) , t ≥ 0} is the Brownian motion driving the zero-coupon bond

dB (t, T )
= r (t) dt + Γ (t, T ) dB (t) , t ≤ T.
B (t, T )

In this framework, the evolution of the forward CPI under the risk neutral probability measure Q corresponds
to a geometric Brownian motion:

dCP I (t, Ti )
= µ (t, Ti ) dt + σ (t, Ti ) dW i (t) , (2)
CP I (t, Ti )

where the volatility structure σ (t, Ti ) and the drift µ (t, Ti ) are deterministic. Specific forms of volatility are

• Black-Scholes [2] (independent of date of observation)

σ (t, Ti ) = σi ,

• Homogeneous case (σ (t, Ti ) = f (Ti − t)):

— Hull and White [5] type volatility (potentially time dependent)

σ (t, Ti ) = σ i (t) e−λi (t)(Ti −t) ,

— Integrated Hull and White volatility (potentially time dependent)

1 − e−λi (t)(Ti −t)

σ (t, Ti ) = σ i (t) .
λi (t)

• Many other type similar to the interest rate models like Mercurio Moraleda types and so on....

3 Consistent modelling of inflation volatility

CP I(.,Ti )
When examining CPI products, we can notice that they are based on a ratio of two CPIs CP I(.,Tj ) . Including
a strike, a vanilla inflation option (let) can be parameterized by the following three parameters:

• Date of fixing of the CPI used in the denominator of the ratio

• Difference between the fixing date of the CPI in the numerator
• strike of the option.
CP I(.,Ti )
An option on year-on-year inflation returns will therefore be a strip of call or put on CP I(.,Ti−1 ) .
CP I(.,Ti )
An option on zero-coupon will therefore be a call or put on CP I(.,T0 ) .
Although the volatility markets between the year-on-year and zero-coupon option may at first sight look
different, there are some connections. A year-on-year option may 1 day become a zero-coupon one as soon as
the CPI of the denominator has fixed. The modelling of the inflation volatility need to account for this.

3.1 A general vol cube

A simple idea is to parameterize the vol structure in term of

• the fixing date of the CPI in the denominator date denoted by T ,

• either the tenor maturity denoted by δ representing the maturity between the numerator and the de-
nominator or the numerator date denoted by T + δ,
• the strikes denoted by K.

4
Definition 1 We call vol cube and we denote by ψ the 3 dimensional deterministic function of a fixing date
T , a tenor δ and a strike K defined as:

ψ : R+ × R+ ∗ ×R → R
+
³ ³ ´´
(T, δ, K) 7→ ψ (T, δ, K) = V ol max / min CPCP
I(T +δ,T +δ)
I(T,T ) − K, 0 .

Each of the three plans of the cube will be a real matrix obtained by fixing one variable among the three.

The market provides indication/information about only year-on-year options (the most liquid ones) and
zero coupon. Hence, looking at the vol cube, we can only get from the market two volatility surfaces (somehow
orthogonal) of the volatility cubes which are ψ (T, T + δ, K)T ≥0 and ψ (0, T, K)T >0 . For different values of the
strike K, a zero-coupon option of maturity one year being the first option year-on-year of tenor one year, the
two surfaces are dependant (∀K ∈ R, ψ (0, T, K)|T =δ = ψ (T, T + δ, K)|T =0 ). A first objective would be thus,
to bind these two surfaces in order to find the consistency relationships between them. The modelling issue is
to provide a way to interpolate/extrapolate volatility information for forward starting structure different from
year-on-year information.

3.2 Options on inflation description

As mentioned above, options on inflation available in the market are vanilla options, zero-coupon and year-
on-year . For some (α, β) ∈ R+ +
∗ × R , the general payoff of:

• an option on zero-coupon inflation is written as:

³ ´
Ti
max / min α CP I(Ti ,Ti )
CP I(0,0) − β, (1 + k0 ) . (3)

Here the nominator is fixed and known at t = 0. Let’s remark the particular expression of the strike
(1 + k0 )Ti , formulate as an actuarial rate. In the special case (α, β) = (1, 1), we simply compare the
inflation to an zero-coupon swap level.
• an option on year-on-year inflation as:
³ ´
max / min α CPCP I(Ti ,Ti )
I(Ti−1 ,Ti−1 ) − β, K . (4)

Here the nominator CP I (Ti−1 , Ti−1 ) can not be known before t = Ti−1 and the strike is crude.

Let’s note for a given strike K, σ BS (0, Ti ) (respectively σ BS (Ti−1 , Ti )) the Black-Scholes volatility of the
option zero-coupon inflation (respectively year-on-year inflation) with exercise date Ti . Keeping the same
notation as the last paragraph, we can write for a fixed strike K ∗ :
½
ψ (0, Ti , K ∗ ) = σ BS (0, Ti )
. (5)
ψ (Ti−1 , Ti − Ti−1 , K ∗ ) = σ BS (Ti−1 , Ti )

In the following section we’ll determine the relation between the Black-Scholes volatilities of the zero-coupon
options and the year-on-year ones for a fixed strike K ∗ , so we will use σBS instead ψ.

3.3 Model inter/extrapolation

Denoting by σ (t, Ti ) the volatility function:

σ : R+ × R+ ∗ → R+
(t, Ti ) 7→ σ (t, Ti )

Equations (2) and (3) gives3 :

Z Ti
V [ln (CP I (Ti , Ti ))] = Ti σ 2BS (0, Ti ) = σ 2 (s, Ti ) ds.
0
3V means variance. It does not depend on the measure.

5
So from (4), we obtain:
h ³ ´i
CP I(Ti )
V ln CP I(Ti−1 )

= Ti .σ 2BS (Ti−1 , Ti )
Ri
T Ti−1
R Ri
T ­ ®
= σ 2 (s, Ti ) ds + σ2 (s, Ti−1 ) ds − 2 σ (s, Ti−1 ) σ (s, Ti ) d W.i−1 , W.i (s)
0 0 0
TZ
i−1

= σ 2BS (0, Ti ) .Ti + σ 2BS (0, Ti−1 ) .Ti−1 − 2ρInf

i−1,i σ (s, Ti−1 ) σ (s, Ti ) ds.
0

where the instantaneous correlation ρInf

i−1,i is defined as in ??. The year-on-year volatility includes:

• the two corresponding zero-coupon volatilities

• the covariance between corresponding zero coupons.
Z Ti−1
Denoting by γ (Ti−1 , Ti ) ≡ σ (s, Ti−1 ) σ (s, Ti ) ds the full correlation integrated covariance4 between
0
the two CPIs, we find the explicit relation between the year-on-year volatilities and the zero-coupon ones:
s
Ti σ 2BS (0, Ti ) + Ti−1 σ 2BS (0, Ti−1 ) − 2ρInf
i−1,i γ (Ti−1 , Ti )
σ BS (Ti−1 , Ti ) = . (6)
Ti − Ti−1

This shows that the relationship between year-on-year and zero-coupon is model dependent through not
only the instantaneous correlation ρInf
i−1,i but also the full correlation integrated covariance γ (Ti−1 , Ti ) which
in terms depends on the volatility assumptions. In the next paragraph, we detail the result for various form
of volatilities.

3.4 Specific form of the volatilities

3.4.1 Case of Black and Scholes
A Black and Scholes volatility is deterministic and homogeneous i.e. an one-dimensional positive function of
time:
σ (t, Ti ) = σ i , ∀0 ≤ t < Ti ,
leading to:
γ (Ti−1 , Ti ) = Ti−1 σ i−1 σi , ∀1 ≤ i ≤ n. (7)
And using the fact that
σ i = σ BS (0, Ti ) .
Hence, the the year-on-year volatilities become a function of the zero-coupon ones only:
s
Ti σ 2BS (0, Ti ) + Ti−1 σ2BS (0, Ti−1 ) − 2ρInf
i−1,i σ BS (0, Ti ) σ BS (0, Ti−1 )
σBS (Ti−1 , Ti ) = . (8)
Ti − Ti−1

3.4.2 Case of Hull and White

In the Hull and White (respectively the integrated) framework, we have an explicit exponential form of volatility
function increasing (respectively decreasing) by time. We can distinguish these special cases ∀0 ≤ t < Ti :
• Hull and White potentially time dependent5 σ (t, Ti ) = σi (t) e−λi (Ti −t) :
P
ni 2λi Tj
−e2λi Tj−1
Ti .σ 2B&S (0, Ti ) = e−2λi Ti σ 2 (j) e 2λi , (9)
j=1
nP
i−1
(λi−1 +λi )Tj −e(λi−1 +λi )Tj−1
γ (Ti−1 , Ti ) = e−(λi−1 Ti−1 +λi Ti ) σ 2 (j) e λi−1 +λi . (10)
j=1
4 We call this full correlation integrated covariance to mean that this would be the covariance if the instantaneous correlation

ρInf
i−1,i were equal to 1.
5 σ (t) is a step wise function: σ (t) =
Pn
i=1 σ (i) k{Ti−1 ≤i<Ti } , ∀0 ≤ t ≤ Tn

6
 −λi (Ti −t)
• Integrated Hull and White, potentially time dependent σ (t, Ti ) = σ i (t) 1−e λi :

P
ni 2λi (Tj −Tj−1 )−4e−λi Ti (eλi Tj −eλi Tj−1 )+e−2λi Ti (e2λi Tj −e2λi Tj−1 )
Ti .σ 2B&S (0, Ti ) = σ 2 (j) 2λ3i
, (11)
j=1
⎧ λi−1 Tj
⎫
⎪ −eλi−1 Tj−1 ⎪
⎪
⎨ Tj − Tj−1 − e−λi−1 Ti−1 e λi−1 ⎪
⎬
nP
i−1
σ 2 (j) (λi−1 +λi )Tj −e(λi−1 +λi )Tj−1
γ (Ti−1 , Ti ) = λi−1 λi +e−(λi−1 Ti−1 +λi Ti ) e . (12)
j=1 ⎪
⎪ λi−1 +λi ⎪
⎪
⎩ −e−λi Ti e
λi Tj
−eλi Tj−1 ⎭
λi

3.5 Numerical results

We present below a set of zero-coupon and year-on-year volatilities in each case of volatility function with
fixed values of parameters. Except in the Black-Schole case, the zero-coupon volatilities do not correspond to
the real market data. This inconsistency results from the choice of volatility function parameters.

Example 2 (Black-Scholes case with ρInf j,i = 0.98, ∀j < i) In this case, the year-on-year volatilities are
higher than the zero-coupon ones. We explain this in the subsection 5.1.

Zero-coupon volatilities from year-on-year

10.00% volatilities in Black Scholes case.
9.00%
8.00%
Vol ZC
Vol YoY
7.00%
6.00%
Value

5.00%
4.00%
3.00%

2.00%
1.00%
0.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Maturity of option

Example 3 (Homogeneous case with ρInf j,i = 1, ∀j < i) In this case, the year-on-year volatilities are lower
than the zero-coupon ones. But they can’t attain a minimum level. We explain this in the subsection 5.1.

Zero-coupon volatilities from year-on-

year volatilities in the homogeneous
6.00% case
Vol ZC
5.00% Vol YoY

4.00%
Value

3.00%

2.00%

1.00%

0.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Maturity of option

7
Example 4 (Hull & White case with ρInf
j,i = 0.98, ∀j < i)

Zero-coupon volatilities from year-on-year

volatilities in Hull & White case.
6.00%

5.00%

4.00%

Value
3.00%
Vol ZC
2.00% Vol YoY

1.00%

0.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Maturity of option

Example 5 (Integrated Hull & White case with ρInf

j,i = 0.98, ∀j < i)

Zero-coupon volatilities from year-on-year

volatilities in integrated Hull & White case.
7.00%

6.00%

5.00%

4.00%
Value

3.00%

2.00% Vol ZC
Vol YoY
1.00%

0.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Maturity of option

4 Convexity adjustment
As for the CMS (see [1] for a review on CMS pricing), the convexity adjustment of the inflation swaps results
from the difference of martingale measures between the numerator and the denominator.

4.1 Intuition
The forward CP I (t, Ti ) fixing at time Ti is obviously a martingale under its payment probability measure
QTi . Similarly, for the forward CP I (t, Tj ) fixing at time Tj under the probability measure QTj , but not QTi .
Consequently, the expected value under the probability measure QTi of the ratio of the two CPIs (with time
Ti > Tj ) has to take into account various convexity adjustments6 :

• CP I (t, Tj ) is not a martingale under the QTi measure. Hence it has to be adjusted to account for the
change of measure between QTj and QTi . This adjustment should intuitively depend on the covariance
between the forward interest bond volatility (between Tj and Ti ) and the forward inflation rate in the
denominator CP I (t, Tj ). This change of measure is similar to the CMS adjustment.
• In addition, we pay CP I (t, Ti ) /CP I (t, Tj ). Because of the correlation between these two inflation
forward rates, we need to account for the their joint move. The adjustment should intuitively be depending
on the covariance between these two CPI rates. This is similar in a sense to a quanto adjustment.

The adjustment is therefore computed in two steps:

• change of measure between QTj and QTi .

6 The forward of the ratio of CPI is not equal to the ratio of the forward CPIs. One calls more or less improperly this adjustment

a convexity adjustment by extension from the one used in interest rates for various change of measures like CMS and in-arrears.

8
 • computation of the expected ratio.

Definition 6 We call the Inflation Convexity Adjustment h at time ¯t between

i Tj and Ti , and we denote
7 QTi CP I(Ti ,Ti ) ¯
λCvx (t, Tj , Ti ), the difference between the forward ratio E CP I(Tj ,Tj ) ¯ Ft and both zero-coupon swaps cor-
QT
E i [ CP I(Ti ,Ti )|Ft ] 8
responding to the two dates which frame it QT , where QTi is the Ti -forward neutral probability,
E j [ CP I(Tj ,Tj )|Ft ]
so: h ¯ i
CP I(Ti ,Ti ) ¯ CP I (t, Ti )
λCvx (t, Tj , Ti ) = EQTi CP I(Tj ,Tj ) ¯ Ft − .
CP I (t, Tj )

4.2 General framework

The forward CP I (t, Ti ) fixing at time Ti being a martingale under its payment probability measure QTi , it’s
dynamic is as follows:
dCP I (t, Ti )
= σ (t, Ti ) dWTi (t) , (13)
CP I (t, Ti )
where (WTi (t))1≤i≤n is a n dimensional Brownian motion under QTi . It is well known that the relationship
between the risk neutral measure and forward measure is given by

dWTi (t) ≡ dW i (t) − Γ (t, Ti ) dt, ∀t ≥ 0, ∀1 ≤ i ≤ n,

where Γ (t, Ti ) is the lognormal volatility of the zero-coupon bond B (t, Ti ) .

This change of measure forces us to specify an implied correlation between zero-coupon bonds and CPI
forwards9 . We suppose that zero-coupon bonds and CPI forwards are driven by different Brownian motions
W i (t) and B (t) with the correlation from (1) :
­ ®
dW i (t) , dB (t) = ρB,I
i dt. (14)

Solving the SDE (13) for Tj and Ti under the same probability QTi leads to:
³ ´
CP I(Ti ,Ti ) CP I(0,Tj )
ln CP /
I(Tj ,Tj ) CP I(0,Ti )

Ri
T T
Rj
= σ (s, Ti ) dWTi (s) − ρInf
j,i σ (s, Tj ) dWTi (s)
0 0
r ³ ´ 2 ¡ 2 ¢
− 1 − ρInf
j,i σ (s, Tj ) dWT⊥i (s) − 1
2 σ (s, Ti ) − σ 2 (s, Tj ) ds
T
Rj
+ρB,I
j σ (s, Tj ) {Γ (s, Ti ) − Γ (s, Tj )} ds.
0

The computation of the expectation of the forward CPI is then simply given by:
h i R Tj
CP I(Ti ,Ti ) CP I(0,Tj ) σ(s,Tj ){σ(s,Tj )−ρInf B,I
j,i σ(s,Ti )}+ρi σ(s,Tj ){Γ(s,Ti )−Γ(s,Tj )}ds
EQTi CP I(Tj ,Tj ) / CP I(0,Ti ) = e 0 .
h i
CP I(Ti ,Ti ) CP I(0,Tj )
The convexity adjustment at t = 0, defined by λCvx (0, Tj , Ti ) = EQTi CP I(Tj ,Tj ) − CP I(0,Ti ) is equal to:

CP I(0,T )
λCvx (0, Tj , Ti ) / CP I(0,Tji )
R Tj
σ(s,Tj )(σ(s,Tj )−ρInf B,I
j,i σ(s,Ti )+ρi {Γ(s,Ti )−Γ(s,Tj )})ds
= e 0 − 1.

This shows that this convexity adjustment depends on

CP I(T ,T )
• the covariance between the two CPI forwards CP I(Tjj ,Tji ) ,
∙ ¯ ¸
CP I(Ti ,Ti ) ¯¯
R Ti
7 Which corresponds to a zero coupon forward swap EQ e− t r(s)ds CP F with a discount factor B (0, Ti ) meadows.
I (Tj ,Tj ) ¯ t
8 This ratio is called naïve CPI forward.
9 This implied correlation should not be the same as the one between inflation and interest rate used in Economic’s models.

This latter is negative whereas the implied one could be positive.

9
 B(t,Ti )
• the covariance between the zero-coupon forward bond B (t, Tj , Ti ) = B(t,T j)
observed at time Tj and the
forward CPI CP I (Tj , Tj )as nominal and inflation securities covariances.

If we note the correlation between CPI forward CP I (Tj , Tj ) and of the zero-coupon forward bond B (Tj , Tj , Ti )
as R Tj
ρB,I σ(s,Tj )(Γ(s,Ti )−Γ(s,Tj ))ds
ζ I,B
j,i ≡
j 0
Tj σ BS (0,Tj ) ΓBS (0,Tj ,Ti ) ,
where Z Tj
2
Tj Γ2BS (0, Tj , Ti ) ≡ (Γ (s, Ti ) − Γ (s, Tj )) ds
0

is the integrated volatility of B (Tj , Tj , Ti ) and replacing by the equation (6), we can write λCvx (0, Tj , Ti ) as
only a sum of volatilities:
I,B
λCvx (0,Tj ,Ti ) eTj σBS (0,Tj )ζ j,i ΓBS (0,Tj ,Ti )
CP I(0,Tj ) = 1 2 2 2 (15)
CP I(0,Ti ) .e− 2 {Ti σBS (0,Ti )−Tj σBS (0,Tj )+(Ti −Ti−1 )σBS (Tj ,Ti )} − 1.

4.3 Assumption about the forward bond volatility

Because of the uncertainty on the estimation of the instantaneous correlation between the forward bond and
the CPI, we take as an input the new integrated correlation ζ I,B j,i . In the case of constant CPI volatility σ (s, Tj )
and forward bond volatility (Γ (s, Ti ) − Γ (s, Tj )) , notice that the two correlations: the instantaneous ρB,I
i and
I,B
integrated one ζ j,i are the same.
Using the BGM model (see [4]), the volatility of the forward bond can be read directly from the volatility
structure of the caplets. This comes from the fact that
Y 1
B (t, Tj , Ti ) = ,
1 + δ k F (t, Tk , Tk + δ k )
k=j..i−1

where F (t, Tk , Tk + δ k ) is the forward Libor of period δ k fixing at time Tk and paid at time Tk + δ k and
observed at time t.
Applying Ito (and looking only at the stochastic part) leads immediately to
X B (t, Tj , Ti )
dB (t, Tj , Ti ) = δ k σ F (t, Tk ) F (t, Tk , Tk + δ k ) dBQk T (t) ,
1 + δ k F (t, Tk , Tk + δ k ) i
k=j..i−1

where σ F (t, Tk ) is the lognormal volatility of the forward Libor F (t, Tk , Tk + δ k ) and where the diffusion is
taken under the QTi probability measure. This means that the forward bond volatility is approximately given
by
X δ k F (0, Tk , Tk + δ k ) σ F (0, Tk )
ΓBS (0, Tj , Ti ) = ,
1 + δ k F (0, Tk , Tk + δ k )
k=j..i−1

where we have approximated the forward by its current value.

4.4 Specific form of volatilities

4.4.1 Case of Black and Scholes
In this case from (7), we get:

γ (Tj , Ti ) = Tj σ BS (0, Ti ) σ BS (0, Tj ) , ∀1 ≤ i ≤ n.

Hence, the inflation convexity adjustment become a function of the inflation zero-coupon volatilities and the
zero-coupon forward bond ones only:
I,B
ΓBS (0,Tj ,Ti )+σ BS (0,Tj )−ρInf
= eTj σBS (0,Tj )(ζ j,i j,i σ BS (0,Ti ))
λCvx (0,Tj ,Ti )
CP I(0,Tj ) − 1. (16)
CP I(0,Ti )

10
4.4.2 Case of Hull and White
We have only to replace the explicit expressions of covariances γ (Tj , Ti ) and the variances Tj σ 2BS (0, Tj ), from
(9) or (11) in (15):
B,I
Tj σ BS (0,Tj )ζ I,B
λCvx (0,Tj ,Ti ) eρi j,i ΓBS (0,Tj ,Ti )

CP I(0,Tj ) = 1 2 2 2 (17)
CP I(0,Ti ) .e− 2 {Ti σBS (0,Ti )−Tj σBS (0,Tj )+(Ti −Ti−1 )σBS (Tj ,Ti )} − 1.

4.5 Numerical Results

Keeping the data in the section (3.5) we present the following results about convexity adjustment of year-on-
year swaps:

Example 7 (Black-Scholes case with ζ I,B j,i = 0.3) In this example the year-on-year swap’s convexity ad-
justments is negative and decreasing with time. This means that the year-on-year swap rate is at least lower
than a year-on-year swap rate whose legs are priced with the raw CPI forwards.

Annual swap's convexity adjustment in

the Black-Scholes case
0.00
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

-0.01

-0.02 Cvx Ajst BS

Level

-0.03

-0.04

-0.05 Time

Example 8 (Homogenous case with ζ I,B j,i = 0.3, ∀j < i) In this case the year-on-year swap’s convexity ad-
justment is greater than in the Black-Scholes case, because for the same other data, the homogenous year-on-
year volatilities are lower then the Black-Scholes ones.

Annual swap's convexity adjustment in

the homogeneous case
0.00
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
-0.01

Cvx Ajst Hmg

-0.02
Level

-0.03

-0.04

-0.05
Time

5 Coherence tests
From the preceding sections, one can notice that the market’s unkonwns are threesome: the Black-Scholes
volatilities of zero-coupon options, the year-on-year ones and the CPIs’ implicit correlations.
However, these three data are not observable at the same time. In fact, one can find some consistency rela-
tionship between these three data according to the assumed volatility function. We present conditions of level
for volatilities and confidence intervals for each component of this trio.

5.1 Level volatility condition

The aim of this subsection, is to provide a minimal level for year-on-year volatilities from a zero-coupon
volatility curve.

11
 In the log-normal deterministic-volatility model, we have the inequality of the Black-Scholes year-on-year
volatility of maturity T and tenor δ:
Z T
δσ 2BS (T − δ, T ) ≥ σ 2 (s, T ) ds. (18)
T −δ

If we consider the two chief forms of volatility function of the last sections, we find respectively:

• Black-Scholes type σ (s, T ) = σ (T ) :

Z T
σ 2 (s, T ) ds = δσ 2 (T ) = δσ 2BS (0, T ) ,
T −δ

so
σ BS (T − δ, T ) ≥ σ BS (0, T ) . (19)
This implies that the year-on-year volatility curve must to be above the zero-coupon one. This does not
correspond to the reality of the market.
• Homogeneous case σ (s, T ) = f (T − s) :
Z T Z T Z δ
v=T −s
σ 2 (s, T ) ds = f 2 (T − s) ds = f 2 (v) dv
T −δ T −δ 0
Z δ
u=δ−v
= f 2 (δ − u) du
0
Z δ
= σ 2 (u, δ) du = δσ2BS (0, δ) .
0

Then
σ BS (T − δ, T ) ≥ σ BS (0, δ) . (20)
This means that the value of the volatility of a year-on-year option of a fixed maturity and a fixed tenor,
is at least greater than the volatility of a zero-coupon option of a maturity which is worth the tenor (the
first zero-coupon volatility and then the year-on-year one when δ = 1 year).

5.2 Triangle bounds

We allow ourself to name the trio Black-Scholes zero-coupon, year-on-year volatilities and CPIs’ correlation
implicit the "market data triangle". Independently from any model or an hypothesis on volatility function,
we want to have an order of size of one of the heads of the market data triangle, by fixing the two others. To
have a intuitive size of each component, we use the main relation (6) rewritten as follows:

δσ 2BS (T − δ, T ) = T σ 2BS (0, T ) + (T − δ) σ 2BS (0, T − δ) − 2ρT −δ,T γ (T − δ, T ) . (21)

The different intervals are obtained according to the choice of the volatility function. As the Black-Scholes
case doesn’t match to the market, we will restrict to the homogeneous case, in which we can distinguish
two cases ∂σ(s,T
∂T
)
≥ 010 and ∂σ(s,T
∂T
)
≤ 011 . These inequalities involve the relation between zero-coupon,
year-on-year volatilities and covariance below:
½
σ (s, T − δ) ≥≤ σ (s, T ) ⇒ T σ 2 (0, T ) − δσ 2 (0, δ) ≥ γ (T − δ, T ) ≥ (T − δ) σ 2 (0, T − δ) .
BS BS ≤ ≤ BS (22)
0 ≤ σ (s, T ) ≤ 1

It leads to these interval bounds of each component of the market data triangle, resumed on the tables
below:
−λT (t)(T −t)
1 0 For exemple, the integrated Hull and White volatility σ (t, T ) = σ T (T − t) 1−e λ (t) .
T
1 1 For exemple, the Hull and White type volatility σ (t, Ti ) = σT (T − t) e−λT (t)(T −t) .

12
 © ª
(∗) For κ ∈ ρT −δ,T , σ BS (T − δ, T ) ,

κ ∈ [min (κ1 , κ2 ) , max (κ1 , κ2 )] , (23)

where
κ2 ³ κ2 ´
2 2 2
ρT −δ,T 1 T σ BS (0,T )+(T −δ)σ BS (0,T −δ)−δσ BS (T −δ,T ) 1 T σ 2BS (0,T )−δσ 2BS (T −δ,T )
2 2 1+ 2
q (T σ2 (0,T )−δσ2 (0,δ))
BS BS q
(T −δ)σ BS (0,T −δ)
σ BS (T −δ,T ) (1−2ρT −δ,T )T σ2BS (0,T )+(T −δ)σ2 (0,T −δ)+2ρT −δ,T δσ2 (0,δ) T σ 2BS (0,T )+(1−2ρT −δ,T )(T −δ)σ 2 (0,T −δ)
δ δ

(∗∗) For σ BS (0, T ), we have

σ BS (0, T ) ∈ [min (σ 1 , σ 2 ) , max (σ 1 , σ 2 )] k{0≤ρ≤ 1 }

2
(24)
+ [0, min (σ 1 , σ 2 )] ∪ [max (σ 1 , σ 2 ) , +∞[ k{ 1 ≤ρ≤1} .
2

Where ⎧ q 2
⎨ σ = δσ (T −δ,T )−(1−2ρδ )(T −δ)σ 2 (0,T −δ)
1
q 2 T
.
⎩ σ = δσ (T −δ,T )−(T −δ)σ 2 (0,T −δ)−2ρδ δσ 2 (0,δ)
2 T (1−2ρ ) δ

5.3 Numerical Results

5.3.1 Triangle heads
Example 9 (Impicit correlation from zero-coupon and year-on-year volatilities.) We present below
a set of market data of zero-coupon and year-on-year volatilities and the implicit CPIs’ correlations deduced
from a homogeneous volatility calibration function schema. Let’s remark that the size are unrealistic at the
medium and long term. This means that the market can be incoherent and that there is at least a freedom
degree in excess.
Zero-coupon and year-on-year
volatilities mareket data
6.00%
Vol ZC
5.00% Vol YoY

4.00%
Value

3.00%

2.00%

1.00%

0.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Maturity of option

Implicit CPIs' correlations in the

101.00% homogeneous case
Rho

100.00%
Level

99.00%

98.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Time

13
5.3.2 Triangle bounds
Example 10 (CPIs’ implicit correlations bounds from volatilities) We can notice that the implicit
CPIs correlation implied by the market is always above 1. The confidence interval, the lower bound at least,
aims at providing the trader a credible and intuitive level of the market correlations.

Exemple of CPIs' implicit correlations.

150.00%

140.00% Rho
130.00% Lower bound
Upper bound
120.00%

110.00%
Level

100.00%

90.00%

80.00%

70.00%

60.00%

50.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Fixing date

Example 11 (Year-on-year volatilties’ bounds from zero-coupon ones and correlations) As the sec-
tion (5.1) shows, the homogeneous year-on-year volatility curve is decreasing but doesn’t go under the first point.

Exem ple of year-on-year volatilties' bounds

7.00%

6.00% Vol YoY

Borne_inf
5.00% Borne_sup

4.00%
Value

3.00%

2.00%

1.00%

0.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Maturity of option

Example 12 (Zero-coupon volatilties’ bounds from year-on-year ones and correlations) We can ob-
serve
¡ that in the1 beginning
¢ of the curve σ BS (0, T ) ∈ [min (σ 1 , σ¡2 ) , max (σ 1 , σ 2¢)]
⇒ 0 ≤ ρ ≤ 2 , and after σ BS (0, T ) ∈ [max (σ 1 , σ 2 ) , +∞[ ⇒ 12 ≤ ρ ≤ 1 . This means that zero-coupon
volatilities provide already an interval of the size of the CPIs’ implicit correlations.

Exemple of zero-coupon volatilties' bounds

7.00%
Vol ZC
6.00% Borne_inf
Borne_sup
5.00%

4.00%
Value

3.00%

2.00%

1.00%

0.00%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Maturity of option

6 Conclusion
In this paper, we derive a market model for the inflation derivatives. Under weak assumptions, we can set up a
model driven only by the term structure of volatilities, describing CPIs forwards. This allows us in particular

14
to relate zero-coupon swaps ( swap market inputs) and volatilities of year-on-year options ( inputs of the option
market). This term structure of volatility as well as assumptions on the implicit correlations (between CPIs
and CPI-zero-coupon nominal Bond) allows us to:

- to infer zero-coupon volatilities from the vol cube information,

- to price year-on-year swaps with consistent convexity adjustments.

Compared to previous models, the offered market models gives consistent assumptions between the zero-
coupon and year-on-year inflation swap market.
Although it is not possible to estimate accurately implicit correlation, we show that these correlations
should satisify certain boundaries conditions. We notice that these boundaries conditions imply unrealistic
level of correlation under certain model hypotheses. We also provide confidence interval for the three unknowns
of the inflation market, leading to what we called the "market data triangle" inequalities. These relationship
relates two of the unkowns to the remaing one.

References
[1] BENHAMOU Eric "Pricing Convexity Adjustment with Wiener Chaos", London School of Economics
Working paper, 2000.
[2] BLACK F. and SCHOLES M., "The Pricing of Options and Corporate Liabilities", Journal of Political
Economy 81, 637—659 1973.
[3] BRACE A., GATAREK D. and MUSIELA M. "The Market Model of Interst Rate Dynamics" Mathe-
matical Finance, 7,2 , pp. 127-55 1997.
[4] BRIGO Damiano, MERCURIO Fabio "Interest Rate Models" Springer, 2001.
[5] HULL John "Option, Futures & Other Derivatives", 4th édition, Prentice-Hull International, Inc 2000.
[6] HULL John "Pricing Interest Rate Derivatives Securities", Review of financial Studies 3, 573—592. Third
Edition 1990.
[7] JARROW Robert & YILDIRIM Yildiray "Pricing Treasury Inflation Protected Securities and Related
Derivatives using an HJM Model " 2000.
[8] LAMBERTON Damien & LAPEYRE Bertrand "Introduction au Calcul Stochastique en Finance", édition
Ellipse, 1997.
[9] PROTOPAPADAKIS Benninga, Simon and Aris, "Real and Nominal Interest Rates under Uncertainty:
The Fisher Problem and the Term Structure", Journal of Political Economy, vol. 91, pp. 856—67, October
1983.
[10] REBONATO Riccardo "Interest-Rate Option Models", 2ème édition, John WILEY & Sons edition 1998.
[11] SARTE Pierre-Daniel G. "Fisher’s Equation and the Inflation Risk Premium in a Simple Endowment
Economy", Federal Reserve Bank of Richmond Economic Quarterly Volume 84/4 Fall 1998.

15