PRE-PRINT

UD 2014-02

The Geometry of Interest Rate Risk

Michele Maio∗
and
Jorrit-Jaap de Jong†

uGly Duckling B.V.

Burgemeester le Fevre de Montignylaan 30,

3055LG Rotterdam, the Netherlands
June 2, 2015

Abstract
In this paper we consider the process of interest rate risk management. The
yield curve construction is revisited and emphasis is given to aspects such as input
instruments, bootstrap and interpolation. For various financial products we present
new formulas that are crucial to define sensitivities to changes in the instruments
and/or in the curve rates. Such sensitivities are exploited for hedging purposes. We
construct the risk space, which eventually turns out to be a curve property, and
show how to hedge any product or any portfolio of products in terms of the original
curve instruments.

Keywords: Yield curve, hedging, interest rate risk management.

∗ [email protected]
† [email protected]

1
Contents
1 Introduction 2

2 Set up 4

3 Zero rates from cash and swap instruments 8

3.1 Cash . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Swap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Interpolation 12
4.1 Linear Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Cubic Splines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2.1 Bessel-Hermite Cubic Splines . . . . . . . . . . . . . . . . . . . . . 14
4.2.2 Monotone-Preserving Cubic Splines . . . . . . . . . . . . . . . . . 14
4.3 Forward Monotone Convex Splines . . . . . . . . . . . . . . . . . . . . . . 15

5 Hedging 15
5.1 Portfolios of several products . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Risk space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

6 Main points 22

7 Conclusion 24

References 26

Appendix 27

A Hagan-West: Relevant Formulas 28

1 Introduction
In this paper we will describe a framework for managing interest rate risk of linear fixed-
income products. The aim is to find a way to hedge interest rate risk. In order to achieve
that, we reconsider well-known practices from the business side and put them in a wider
more mathematical framework. This is done by looking at the present value of one basis
point (PV01) or equivalently by the instrument value of one basis point (IV01). These
quantities allow us to define hedging strategies and work out the positions vector that
one should take to hedge the risk of small movements in the interest rates of market
instruments.
Many quantities of interest require the knowledge of the term structure of interest
rates at values that are different from the input data, namely the yield curve is needed.
Examples of such quantities are the discount factors for generic maturities or any prod-
uct’s PV01 which requires the derivatives of the yield curve with respect to its zero
nodes.
Two approaches are normally used to solve this issue: interpolation and parametric
fitting methods. When using interpolation one is guaranteed that the curve will pass

2
through all the input nodes by construction. This is not in general the case for paramet-
ric methods, where typically one is given a curve that depends on a number of parameters
whose values must be chosen appropriately. This is what happens for example in the
Nelson-Siegel-Svensson models [1, 2] where the parameters are fixed by minimizing the
sum of the squares of the distances between the observed input nodes and their corre-
sponding values expected from the curve. Other parametric models are those developed
by Wiseman [3, 4] by Bjork and Christensen [5], by James and Webber [6], and extensions
thereof. In this paper we will use interpolation only.
The yield curve will turn out to be crucial in constructing the hedging positions
for any product. In fact, the whole risk space will be determined by the properties of
the yield curve. On one side, those positions clearly must depend on the details of the
product that we want to hedge, and on the other side they are fixed by the yield curve.
Moreover, they are independent on the particular hedging strategy. A strategy based on
the curve PV01 will generally give the same positions as a strategy based on the curve
IV01. This will be easy to see once we realize that PV01 and IV01 are just two different
representations of the same concept but in two different basis, with a Jacobian matrix
relating them. Intuitively, this behavior is the counterpart of the bootstrap procedure
which relates instruments rates to zero rates.
To construct the framework three steps are needed.
First, we construct a yield curve. The various methods of curve construction all share a
bootstrap approach, but vary in interpolation technique. These techniques are not novel
and an overview of the most common approaches can be found in Bolder and Stréliski’s
report [7]. A recent contribution to the literature is the method described by Hagan and
West [8, 9]. These authors also give a list of good features that a yield curve should
have, that we can summarize with the following criteria: i) the curve should price back
the market (or at least deviations from market prices should be small), so that market
data are recovered; ii) forward rates should be continuous in order to avoid arbitrage,
but also positive and stable in order to price instruments correctly; iii) the curve should
be local, so that a small change in one of the input nodes modifies the curve values only
in the neighborhood of that node and not far away; iv) the hedges should be local, in
such a way that if one of the curve instruments is hedged then the hedge is assigned to
that instrument only.
Second, given the yield curve, we can compute the value of any linear fixed-income
product, namely a product that depends only on rates1 . Using the present value formula
as a sum of discounted cash flows, we can compute the sensitivity to changes in the
zero rates of the curve (PV01) or to the instrument rates (IV01). These quantities are
not completely independent of each other and will be crucial for hedging. We will see
that they are related by a change-of-basis transformation and that they are the starting
point to determine hedging positions. Consistently, the hedging positions that can be
determined for any product will be the same, independently from whether one uses the
product’s PV01 or IV01. Ultimately this happens because for any product the risk space
is determined by the yield curve, or more fundamentally by the instruments used to
bootstrap the yield curve. Our approach to compute PV01 and IV01 is fully analytic:
the relevant derivatives are computed exactly, however for the PV01 and IV01 we will
use a first-order approximation. This is a perturbative approach.
Thirdly, and finally, we can use these sensitivities to hedge (or replicate) any position.
1 This would is not the case for the so-called non linear products, which depend on rates, volatilities

and in some cases other quantities such as correlation.

3
For example, as referred to by Whittaker [10], the interest rate risk of a swap portfolio
can be managed in this way. To check whether the hedge is local, the preferred way
is to use exactly the set I as hedging instruments. Then, for any curve instrument we
can compute the change in the present value due to a small change in any of the curve
rates. We should find that this change in the instrument present value is zero for those
changes in the curve rates that do not involve the rate of that particular instrument.
Equivalently, changes in a given curve rate only triggers changes in the present value of
the instrument used to construct that curve rate.
As pointed out by Reitano [11], there are problems with this approach, mainly because
the replication is only a first-order approximation. Hence, all higher-order factors are
ignored. In practice this means that instruments with high-convexity will not be properly
hedged. This is indeed true, however when a portfolio is closely monitored and the hedge
is updated frequently, this first-order framework can be used.
Since the credit crunch crisis, it is common practice to use the multi-curve approach
for the valuation of fixed-income products. This approach is described by Bianchetti
[12] (see also [13, 14]) and is considered the market standard to accurately include basis
risk. In this paper the multi-curve framework is not considered, mainly for simplicity’s
reasons, however this framework can be extended to include multi-curve and we hope to
do that in the future.
The plan of this paper is as follows.
In Section 2 we define the set up and fix our notation. In particular, we describe the
role of bootstrap and interpolation for the yield curve construction, and we define the
PV01 and IV01. Section 3 explains how the zero rates, that are the node points of the
yield curve, are derived from various instruments’ rates. In particular we will compute
the zero rates related to cash and swaps. We will also compute the Jacobian matrix of
derivatives to change from zero to instrument rates.
Section 4 contains a summary of the various interpolation methods that we have used to
check our approach to measure risk in this paper, namely linear, cubic splines (Bessel-
Hermite and monotone-preserving), and forward monotone convex spline. The methods
that we use here are all partially treated in the Hagan-West papers [8, 9].
In Section 5 we describe how the PV01 and the IV01 are used for hedging purposes.
In section 6 we present a helicopter view of the quantities introduced in this paper,
how they are linked to each other and how they are used for risk management purposes.
Finally, in Section 7 we summarize our results, briefly discuss open issues, present possible
directions for future investigation, and throw our conclusions.

2 Set up
In this section we will explain our set up and fix our notation. Let us start by recalling
what the term structure of interest rates, or simply yield curve, is. A yield curve is
constructed out of a discrete set of instruments I which allows us to derive another
discrete set N of input nodes for the curve. The set I will contain several instruments
of various maturities, e.g.2

I = {cash2w , cash1m , cash3m , cash6m , swap1y , swap5y , swap10y , swap20y } . (2.1)

2 We will consider only cash and swap in this paper, but in principle any fixed-income instrument can

be used. The choice of the instruments used to construct the curve will be important when we move the
discussion to hedging.

4
Instrument rates and maturities are in one-to-one correspondence with the zero nodes of
the yield curve. The exact relationship, that will be explained in section 3, allows us to
construct the set of zero nodes:

N = {(t1 , r1 ), (t2 , r2 ), . . . , (tn , rn )} , (2.2)

which are the curve inputs. In general, there exists a bootstrap algorithm B that allows us
to compute the zero rate of an instrument given its price and its principal value in a way
that is consistent with the zero rates of the other instruments with shorter maturities
(see e.g. section 4.5 of Hull’s book 9th ed. [15] for a review, or the Deaves-Parlar paper
[16] for a generalized approach to bootstrap):

B : I −→ N (2.3)

A yield curve is a function γ that maps the discrete set of zero rates into the real
numbers
int
γ ≡ γN : N −→ R . (2.4)
The suffix denotes that the curve construction is strongly dependent on interpolation. If
one uses parametric fitting methods (e.g. Nelson-Siegel-Svensson) to construct the curve,
then it will not necessarily pass through all the nodes in N . However, if interpolation is
used, then γ will go through all the node points by definition and consequently the market
is priced back. Equivalently, if the market instruments are priced using the yield curve
then the input prices are recovered. In any case, the curve can also be extrapolated
outside the interval [t1 , tn ] (e.g. linear extrapolation, or any other model-dependent
prescription), even if the interpretation of the values that one obtains might be subject
to discussions.
The yield curve is needed to compute the present value (P V ) of any product via
the discount factor. We use the continuously-compounded convention for interest rates
throughout this paper
D(t) = e−r(t)·t , (2.5)
but other choices are equally allowed. One such example is the the n-periodic compound-
ing
 −nt
1
D(t) = (2.6)
1 + nr
which converges to the continuous compounding in the n → ∞ limit.
Typically, a product promises intermediate cash flows in the future during its lifetime.
The present values of these cash flows are given by their discounted amounts. Hence, to
each product we can associate a set of discount factors

D = {D1 , D2 , . . . , Dm } , (2.7)

where Di = D(t̃i ) and t̃i , i = 1 . . . , m, correspond to the time of the m future cash flows.
Hence, to get the PV of a product we have to follow the steps as in the sequence (2.8):

I
B /N γ
/ Rγ D /D PV / RP V . (2.8)

Here, Rγ is the range of the curve γ which is normally R+ , RP V denotes the range of
the present value PV, typically R, and the discount function D represents the convention
chosen for discounting.

5
 One could also wonder what happens if one of the input zero nodes is changed by a
small amount, typically one basis point3 . In this case we can use the one-variable Taylor
series that approximates the present value with respect to changes in the interest rate:
∞
X 1 dn P V n dP V 1 d2 P V
P V (r + δr) = n
(δr) = P V (r) + δr + 2
(δr)2 + . . . (2.9)
n=0
n! dr dr 2 dr

In this paper we will focus on the changes in the present value due to chances in the
interest rate (changes in the term structure) for linear products, that is the value of a
product without optionality. As a result the first order approximation will suffice, for
which equation (2.9) reduces to:

dP V
δP V (r) = δr + O(δr2 ) (2.10)
dr
where δP V (r) = P V (r + δr) − P V (r). When δr = 1 basis point, then the l.h.s. gives
back the standard PV01. Clearly, in the limit when δr → 0, δP V vanishes as well, but
the derivative on the r.h.s. does not.
Actually the exact counterpart of formula (2.10) is
Z r+δr
d
δP V (r) ≡ P V (r0 ) dr0 . (2.11)
r dr0

From this we can see that if we know analytically the expression for the derivative of
the present value with respect to the rate, then we can re-construct the PV01 exactly by
integrating the derivative on the interval [r, r + δr] with δr = 1bp.
We will compute the derivative dPdrV , which depends on the definition of the financial
product, the formula used for the discount factors, the interpolation method and the
instruments used to construct the yield curve. In particular, we consider a yield curve
constructed out of n market instruments. Each instrument will result in fixing a node
point in the yield curve. It will be shown that the gradient vector ∇P V w.r.t. the node
points spans the complete risk space for linear fixed-income products (and it is the best
first-order approximation for more complex products).
Defining ri as the zero rate at node point i, from eq. (2.10) we get:

δP Vi = (∇P V )i · δri , (2.12)

∂P V
where ∇P V is the gradient vector with components (∇P V )i = ∂ri , and the total risk
for the product -given all the curve nodes- is:
X ∂P V
δP V (r) = ∇P V · ∆r = δri . (2.13)
i
∂ri

Here ∆r is the shift vector, ∆r = (δr1 , δr2 , . . . , δrn ). If δri is one basis point and all the
others are zero, i.e. δrj = 0.0001 · δi,j , then each component in the sum gives the present
value of one basis point, a.k.a. P V 01i , with respect to zero rate i:
∂P V
P V 01i = δri . (2.14)
∂ri
3 By definition of basis point, 1bp = 10−4 .

6
When the PV01 is implemented numerically or with a first-order Taylor approximation,
one should choose the bump δri = 0.0001. The exact counterpart of this is
Z ri +1bp
∂
P V 01i ≡ dri0 P V (ri0 ) . (2.15)
ri ∂ri0

PV01 measure is a widely-used metric for interest rate risk. In addition, if all the rate
changes in the vector ∆r are set to 0.0001 at the same time, then the result is a parallel
shift of the complete curve of one basis point. This is also often called DV01 (dollar value
of one basis point) or confusingly PV01 (which conflicts with our earlier definition).
Similarly to the zero rate ri , we can also consider the instrument rate xi at node i,
which is the rate specified by the instrument that is used to construct the corresponding
zero rate in the yield curve, and define the IV 01i (instrument value of one basis point)
at i as
IV 01i = P V (r(xi + 1bp)) − P V (r(xi )) . (2.16)
In order to compute this, since δx1 = 1bp is small enough to be in the linear regime, we
can use first-order Taylor approximation as
∂P V
IV 01i = δxi . (2.17)
∂xi
However, there is an exact counterpart in this case too. By extension of the argument
that led to the exact relations (2.11) and (2.15), mutatis mutanda, the same reasoning
holds for the calculation of the IV01, with the only difference that now the relevant
derivative is ∂P V
∂xi :
Z ri +1bp
∂
IV 01i ≡ dx0i 0
0 P V (xi ) . (2.18)
ri ∂x i

The present value of a product is by definition the discounted sum of its future cash
flows: X
PV = Acf D(tcf ) . (2.19)
cf

Here, starting from t = 0 as today’s date, the index cf runs over all the future cash
flows, Acf is the amount exchanged at time tcf and D(tcf ) = exp (−r(tcf )tcf ) is the cor-
responding discount factor, where r(tcf ) is the interest rate obtained from the knowledge
of the yield curve. In order to compute these derivatives, the best way is to use the chain
rule as follows:
∂P V X ∂P V ∂Dcf ∂r
= (2.20)
∂ri ∂Dcf ∂r ∂ri
cf

for the PV01i , and

∂P V X ∂P V ∂rj X ∂P V ∂Dcf X ∂r ∂rj
= = (2.21)
∂xi j
∂rj ∂xi ∂Dcf ∂r j ∂rj ∂xi
cf

for the IV01i . Here r(t) ≡ γ(t) is the yield curve. The key difference between PV01 and
∂r
IV01 is the presence of the Jacobian ∂xji . We will see later that the Jacobian is in general
not diagonal, but lower triangular matrix, as a consequence of the bootstrap procedure.
Formulas (2.20) and (2.21) are the mirror of the sequence (2.8). In fact:

7
 ∂rj
• ∂xi represents the bootstrap on the set I;
∂r
• ∂rj represents interpolation from the set N ;
∂Dcf
• ∂r depends on the choice of discounting convention;
∂P V
• ∂Dcf is related to the product only via its PV.

Many quantities in the formulas above require the knowledge of the term structure
of interest rates at values that are different from the input data, namely the yield curve
is needed. Examples of such quantities are the discount factors for the cash flows D(tcf )
∂r
or the derivatives ∂r i
. As discussed earlier, we will generate these values by using inter-
polation rather than parametric fitting methods.

3 Zero rates from cash and swap instruments

In this section we describe how to derive the zero rates that are then used to construct
the yield curve via bootstrap and interpolation from various instrument rates.
A zero rate is the interest rate matured by a zero-coupon bond. Zero rates are
constructed out of marked data, and in particular from instruments traded in the market.
Here we will focus on the relation between zero rates and cash rates as well as swap rates4 .
Typically cash rates are used to construct the short-term structure of interest rates,
while swap rates are responsible of the long-term yield curve.
Later we will move to hedging and we will need to compute the quantities PV01 and
the IV01. These two quantities differ by the presence of the Jacobian matrix5

∂(x1 , x2 , . . . , xn ) ∂xk
J ≡ Jx (r1 , r2 , . . . , rn ) = ⇐⇒ Jkl = (3.2)
∂(r1 , r2 , . . . , rn ) ∂rl

between zero rates and instrument rates in the latter. Note that J is a curve property,
since it depends only on the instruments used to construct the curve:

J ≡ J (I) . (3.3)

Actually, as already pointed out in (2.20) and (2.21), we will mostly need the inverse of
such a matrix, namely J −1 :

∂(r1 , r2 , . . . , rn ) ∂rl
J −1 = J −1


⇐⇒ lk
= (3.4)
∂(x1 , x2 , . . . , xn ) ∂xk

In order to compute (3.4) we will first compute the actual Jacobian (3.2) and then take
∂rl
its inverse, rather than directly computing the entries ∂x k
. In fact, it turns out that this
4 With the word swap we actually mean plain vanilla interest rate swap.
5 Recall that the notation for the Jacobian matrix is:
 ∂x ∂x1 ∂x1

1
∂r ∂r2
... ∂rn
 ∂x12 ∂x2 ∂x2
...

∂(x1 , x2 , . . . , xn )  ∂r1
 ∂r2 ∂rn

Jx (r1 , r2 , . . . , rn ) ≡ = . . (3.1)

∂(r1 , r2 , . . . , rn )  . .
.. ..
 . ··· .


∂xn ∂xn ∂xn
∂r ∂r
... ∂rn
1 2

8
approach is both easier and provides a natural way to deal with the first node. Moreover,
it is pretty straightforward to derive formulas for the instrument rates xl in terms of the
zero rates {rk } and in addition, even though the actual values depend on the specific
interpolation scheme, these formulas are interpolation-independent. We will also see
that generically both J and J −1 are lower triangular matrices and that this is the case
because the zero rate at position i is sensitive to changes in the previous instruments at
positions up to i, as consequence of the bootstrap procedure.
The first node of the curve is special because of the settlement delay. Once entering a
contract, there is always a settlement delay, that could in principle be zero, but in reality
is non-zero6 . As a consequence, almost any contract will be likely to start at a future
date7 . We will include this effect in our discussion and denote the real starting time by
s (s ≥ 0).
We will soon see that the Jacobian will be expressed in terms of derivatives of the
type ∂r(t)
∂ri . This quantity strongly depends on the interpolation method used to compute
the yield curve r(t). In this section we will focus on the Jacobian, while in section 4 we
will discuss interpolation.

3.1 Cash
Cash instruments are simply deposits that promise a pre-agreed interest over a pre-
determined time on an initial invested amount. Consider a set of cash instruments with
different maturities τ (e.g. τ = 2 weeks, τ = 1 month, τ = 3 months, τ = 6 months,
etc.). They all promise to pay back a simply-compounded interest at time t = τ + s,
where s is the settlement delay and t will be the location of the node used to construct
the yield curve.
For cash, the relation between the zero rate at time t and cash rate with maturity τ
is:
D(s) = (1 + x · τ )D(t) , (3.5)
where x is the cash rate of the cash instrument with maturity τ , τ = t − s, and D(s) and
D(t) are the discount factors at times s and t8 . Formula (3.5) can be inverted to give
the cash rate as a function of the node zero rates:
 
1 D(s)
x= −1 . (3.6)
τ D(t)

Observe that D(s) is typically function of the first node only and its value depends on
the interpolation method, while D(t) = D(τ +s) is independent of the interpolation since
t = τ + s defines a node point.
For more cash instruments, there will be more cash rates x1 , x2 , . . . , xn , with matu-
rities respectively 0 ≤ s ≤ τ1 ≤ τ2 ≤ · · · ≤ τn . Before going to the general case, let us
consider a simpler example. Let us take s = 2 days as settlement delay and n = 3 cash
6 E.g.in the case of the Eurozone the settlement delay is two days.
7 Exceptions to this case are e.g. all the overnight products, which start on the spot date.
8 The relation τ = t − s is actually valid only in our idealized mathematical world where there is only

one measure of times. In practice, conventions appear all over the places. The yield curve carries its own
convention to measure times, while instruments follow other conventions for day counting, business days,
and holidays, that are typically different from the curve one. In the real world, s and t are measured
with the curve convention, while τ with the instrument convention. The relation τ = t − s is only
approximately true.

9
instruments with tenors τ1 = 2 weeks, τ2 = 1 month, τ3 = 3 months. Formula (3.5) gives
now explicitly:
D(s) = (1 + x2w · τ1 ) D(t1 ) (3.7a)
D(s) = (1 + x1m · τ2 ) D(t2 ) (3.7b)
D(s) = (1 + x3m · τ3 ) D(t3 ) (3.7c)
where ti = τi + s. Without doing any calculation, we can immediately see that
∂x2w ∂x2w ∂x2w
6= 0 =0 =0 (3.8a)
∂r1 ∂r2 ∂r3
∂x1m ∂x1m ∂x1m
6= 0 6= 0 =0 (3.8b)
∂r1 ∂r2 ∂r3
∂x3m ∂x3m ∂x3m
6= 0 =0 6= 0 . (3.8c)
∂r1 ∂r2 ∂r3
This structure is typical for cash instruments only, where some of the lower-diagonal
entries are zero. The derivative with respect to the first node is non-zero due to the
discount factor D(s), which is usually fixed in terms of r1 9 , while the derivative with
respect to the node i is non-zero only for the cash instrument responsible for that node,
since D(ti ) is fixed by ri only. It is straightforward to compute these derivatives:
 
∂xj 1 D(s) ∂r(s) ∂r(tj )
= −s + tj . (3.9)
∂ri (tj − s) D(tj ) ∂ri ∂ri

One can rewrite this in various ways, using (3.5) and the definition of the discrete forward
rate:
∂xj 1 ∂
= (1 + x · τ ) · (−sr(s) + tj r(tj ))
∂ri (tj − s) ∂ri
 
∂ tj r(tj ) − sr(s)
= (1 + x · τ ) ·
∂ri tj − s
∂ d
= (1 + x · τ ) · f (s, tj ) ,
∂ri

where f d (s, tj ) is the discrete forward rate between times s and tj .

A few remarks are now in order. First. Notice that already for cash we see a lower
triangular matrix for the quantity J as defined in (3.2):

J is a lower triangular matrix . (3.10)

In addition, some entries in the lower triangle are zero. The non-zero entries are on the
diagonal and in the first column. This is true if only cash instruments are considered10 .
This happens because the zero rate at position i is sensitive only to its cash instrument
at i and to the initial cash instrument (via the discount factor D(s)). This structure
will appear again later when we will start adding swap instruments to the picture, but
those zeroes in the lower triangle will generically disappear. The zeroes in the triangle
are typical of cash instruments.
9 This is true in many cases, e.g. when flat extrapolation is used.
10 Otherwise the positions of the vanishing and the non-vanishing entries will generally be different.

10
 ∂ri
Second. Later we will need to compute the matrix of derivatives ∂x j
. This matrix will
simply be given by the inverse matrix of (3.10). If we denote the elements of the Jacobian
matrix (3.10) as
∂xk
Jkl = , (3.11)
∂rl
where for each cash k the entries are given by (3.9), then we have
∂ri
= J −1 ij .


(3.12)
∂xj

3.2 Swap
For plain vanilla interest rate swaps, or swaps in short, the reasoning is similar, but the
formulas are more complicated. The defining relation for the swap with total maturity
τ is:
Ncf
X
D(s) − D(t) = x αk D(tk ) . (3.13)
k=1
The notation here is as follows. The l.h.s. represents the floating side of the swap, where
s is the settlement delay and t is the end date11 using the conventions for the floating leg
of the swap. The r.h.s. is the fixed side of the swap, x is the swap rate, Ncf is the number
of fixed cash flows, tj is the time when the j th cash flow is paid, using the conventions
for the fixed leg of the swap, and αj is the tenor of each fixed cash flow12 . If we invert
this relation we obtain
D(s) − D(t)
x = PNcf (3.14)
k=1 αk D(tk )
for the swap rate and

  X Ncf
∂x 1 ∂r(s) ∂r(t)
= P 2  −sD(s) + tD(t) · αk D(tk )
∂ri Ncf ∂ri ∂ri
k=1 αk D(tk ) k=1

Ncf
  X ∂r(tk ) 
+ D(s) − D(t) · αk D(tk )tk  (3.15)
∂ri
k=1

for the derivative of the swap rate w.r.t. a given zero rate ri . Observe that the quantity
PNcf
k=1 αk D(tk ) appearing in the equation above is the PV01 (or better the IV01) of the
swap, obtained by deriving (3.13) by x.
In general, in the curve building we have more than just one swap instrument, and
hence there will be more swap rates x1 , x2 , . . . , xn , with tenors respectively 0 ≤ s ≤
τ1 ≤ τ2 ≤ · · · ≤ τn . As an example, for n = 3 we could have τ1 = 1 year, τ2 = 3 years,
τ3 = 5 years. As we have already done for cash, in this case we have
D(s) − D(t1 )
x1y = PNcf (3.16a)
k=1 αk D(tk )
11 In the mathematical world, t = τ + s and later t
Ncf = τ + s. However, the floating leg has generally
different conventions for measuring times than the fixed leg, hence it might happen that the end date
on the floating and fixed side are not equal.
12 It is computed as the time between the (j − 1)th and the j th fixed cash flow and it depends on the

index j because of the day count convention.

11
 D(s) − D(t2 )
x3y = PNcf (3.16b)
k=1 αk D(tk )
D(s) − D(t3 )
x5y = PNcf (3.16c)
k=1 αk D(tk )
where ti = τi + s. From these expressions it easy to see that
∂x1y ∂x1y ∂x1y
6= 0 =0 =0 (3.17a)
∂r1 ∂r2 ∂r3
∂x3y ∂x3y ∂x3y
6= 0 6= 0 =0 (3.17b)
∂r1 ∂r2 ∂r3
∂x5y ∂x5y ∂x5y
6= 0 6= 0 6= 0 . (3.17c)
∂r1 ∂r2 ∂r3
Hence we recover the lower triangular structure as in (3.10), but now all the entries are
generically non-zero13 because of the intermediate cash flows coming from the fixed leg.
∂x
For instance, this generically happens for the entry ∂r5y
2
, where the r2 -dependence is in
the denominator of x5y .
Extending this n = 3-result to the general situation is straightforward. We will have
a Jacobian matrix of the form
∂xk
Jkl = , (3.18)
∂rl
where for each swap k the entries are given by (3.15) and whose inverse is
∂ri
= J −1 ij .


(3.19)
∂xj
Moreover, both matrices are lower triangular because the zero rate at position i is sen-
sitive to all the swap instruments up to i (through the settlement discount D(s) and all
the fixed-leg cash flows).

4 Interpolation
We have seen in section 3 that the Jacobian J is computed in terms of the matrix
∂r(t)
, (4.1)
∂ri
which is interpolation-dependent. Therefore it is important to understand how it is
computed in various interpolation schemes. For all the methods considered here, except
Hagan and West’s interpolation, we use flat extrapolation outside the data points (ti , ri ),
with i = 1, . . . , n:

r(t) = r1 t ≤ t1 and r(t) = rn t ≥ tn . (4.2)

This immediately implies that for those methods one has:

∂r(t) ∂r(t)
= δi,1 t ≤ t1 and = δi,n t ≥ tn , (4.3)
∂ri ∂ri
13 Some of them could still be zero, but it would be by chance and case dependent.

12
where δi,j denotes the Kronecker delta. This is not the case for Hagan and West’s method.
For all the methods however the following is true: for any time t within the input range,
we can always determine the interval [tî , tî+1 ] for a specific index î that encloses t, i.e.
tî ≤ t ≤ tî+1 14 .
In the following subsections we will give the exact analytic expressions for the deriva-
tive (4.1) for some specific interpolation schemes. Observe that, within our framework,
actually all the higher-order derivatives are always known exactly. In practice, the high-
est order is limited by the specific interpolation method used to construct the curve.
Typically, spline-like methods result in curve which are C 1 , sometimes C 2 , with vanishing
higher derivatives, so that in a Taylor-like approach the sum will stop quite soon.

4.1 Linear Interpolation

In linear interpolation, the curve r(t) is specified by
rî+1 − rî
 
r(t) = (t − tî ) + rî , t ∈ [tî , tî+1 ] . (4.4)
tî+1 − tî

Hence our derivative is

  
t−tî

 1 − t −t if i = î

 î+1 î


∂r(t)  
t−tî

= if i = î + 1 . (4.5)
∂ri 
 tî+1 −tî




0 otherwise


4.2 Cubic Splines

In interpolations based on cubic spline the curve r(t) is specified by

r(t) = ax + bî (t − tî ) + cî (t − tî )2 + dî (t − tî )3 , t ∈ [tî , tî+1 ] . (4.6)

Hence our derivative is generically:

∂r(t) ∂aî ∂b ∂c ∂d
= + î (t − tî ) + î (t − tî )2 + î (t − tî )3 . (4.7)
∂ri ∂ri ∂ri ∂ri ∂ri
The choice of the coefficients a, b, c and d defines different methods. For the spline-like
methods that we will consider here, it is enough to specify the b’s, i.e. the derivatives of
the curve at the node points. In fact, the a’s are fixed by the nodes, i.e. aî = rî , and for
c’s and d’s we have (see [8]):
3mî − bî+1 − 2bî bî+1 + bî − 2mî
cî = and dî = , (4.8)
hî h2î

where we have defined

aî+1 − aî
hî = tî+1 − tî and mî = . (4.9)
hî
14 For Hagan and West’s interpolation in section 4.3 the interval will be denoted by [tî−1 , tî ].

13
Consequently, we have
∂aî
= δi,î , (4.10)
∂ri
and the derivatives
∂cî ∂dî
and (4.11)
∂ri ∂ri
are fixed in terms of the derivatives of the b’s. The only method-dependent quantity is
then
∂bî
. (4.12)
∂ri
We will consider two kinds of cubic splines, namely the Bessel-Hermite and the monotone-
preserving splines, and show how the b’s and its derivative are computed there.

4.2.1 Bessel-Hermite Cubic Splines

In Bessel-Hermite splines the b-coefficients are fixed by considering three consecutive
points, drawing the only parabola passing through all of them and computing the deriva-
tive at the node. This procedure defines formulas for these coefficients, that we recall
here [8]:
 
1 (t3 + t2 − 2t1 )(r2 − r1 ) (t2 − t1 )(r3 − r2 )
b1 = − (4.13)
t3 − t1 t2 − t1 t3 − t2
 
1 (ti+1 − ti )(ri − ri−1 ) (ti − ti−1 )(ri+1 − ri )
bi = + (1 < i < n) (4.14)
ti+1 − ti−1 ti − ti−1 ti+1−ti
 
1 (tn − tn−1 )(rn−1 − rn−2 ) (2tn − tn−1 − tn−2 )(rn − rn−1 )
bn = − − .
tn − tn−2 tn−1 − tn−2 tn − tn−1
(4.15)
∂b
The calculation of ∂rîi is now straightforward. First, for any given value of t to interpolate,
one has to determine the lower index î that identifies the corresponding bî , and then the
derivative can be carried out as usual.

4.2.2 Monotone-Preserving Cubic Splines

As the name suggests, these splines preserve the monotonicity of the curve by following
the trend of the input points. This is achieved in the following way [8]. First, the
derivatives at the boundary are set to zero, b1 = bn = 0. Then, if the curve has a turning
point at position î, its derivative is also set to zero, bî = 0, otherwise its value is

bî = βî , (4.16)

where βî is equal to

3mî−1 mî
βî = . (4.17)
max(mî−1 , mî ) + 2 min(mî−1 , mî )

In the original paper [8] the following additional adjustment was included:


min max(0, bî ), 3 min(mî−1 , mî )
if mî > 0
bî = . (4.18)
max min(0, bî ), 3 max(mî−1 , mî ) if mî < 0

14
However, it can be shown [18, 19] that the latter adjustment is not necessary, since mono-
tonicity had already been achieved at the previous step. At this point the calculation of
∂b ∂b
the derivative ∂rîi is straightforward: if the curve has a turning point at î, then ∂rîi = 0,
otherwise the curve is locally monotonic at î and one has:
∂bî ∂βî
= (4.19)
∂ri ∂ri
(see [18, 19] for the relevant expressions).

4.3 Forward Monotone Convex Splines

This method was introduced by Hagan and West [8]. In [9] they also give the algorithm
to implement it.
The method starts from the realistic assumption that forward rates are constant in
the intervals [ti−1 , ti ], with 1 ≤ i ≤ n, t0 = 0. Eventually, the yield curve is recovered
from the forward function as
d(r(t)t) 1
+ fîd (t − tî−1 ) + It ,

f (t) = =⇒ r(t) = r t t ∈ [tî−1 , tî ] , (4.20)
dt t î−1 î−1
where Z t
It = g(τ )dτ (4.21)
tî−1

and t ∈ [tî−1 , tî ] for a specific index î ∈ [1, n]. This integral is also defined in four regions.
Finally, since we are interested in the Jacobian J , we need to compute the derivative
of the yield curve w.r.t. the input zero rates:
   
∂r(t) 1 ∂ d ∂
= δ t + f (t − tî−1 ) + It . (4.22)
∂ri t i,î−1 î−1 ∂ri î ∂ri
The derivative of the integral is done region by region, while the derivative of the forward
rates gives a linear-like contribution:
   
∂ d t − ti−1 t − ti−1
fî = δi,î−1 1 − + δi,î . (4.23)
∂ri ti − ti−1 ti − ti−1
∂It
The relevant formulas for g(τ ), It and ∂rj can be found in appendix A.

5 Hedging
In this section we will describe how hedging works. There are various approaches to
compute hedging-related quantities. One such approach is the so-called wave or scenario
method, which allows one to separate the risks of the yield curve from the instruments.
This is sometimes desirable since in principle the curve instruments do not need to be
exactly the same as the hedging instruments. However, we will not pursue this approach
here, but leave it for the future.
We will assume that the fundamental vector is a vertical column and hence its trans-
pose is a horizontal row. Moreover, the relevant vectors will be denoted by lower-case
Greek letters, while matrices by upper-case Greek letter (except for the Jacobian, which
is still denoted by J ). Due to the amount of notation used in this section, let us list the
main quantities first:

15
 • J is the Jacobian introduced earlier
• Ψ will denote the matrix whose columns are the IV01 of the curve instruments I
• Ψi , with i = 1, . . . , n will denote the columns of Ψ
• ψ will denote the IV01 of an arbitrary product that we want to hedge
• Π will denote the matrix whose columns are the PV01 of the curve instruments I
• Πi , with i = 1, . . . , n will denote the columns of Π
• ξ will denote the PV01 of an arbitrary product that we want to hedge
• ω will denote the vector with the hedging position
We have already defined the PV01 and the IV01 in (2.14) and (2.17). Let us recall
them here for convenience:
∂P V ∂P V
P V 01i = δri , IV 01i = δxi . (5.1)
∂ri ∂xi
These are vector components (not summed over), corresponding to the nodes in the yield
curve. The PV01 measures the sensitivity of any product’s present value to changes in
the zero rates by one basis point, while the IV01 measures a similar sensitivity when
the underlying instrument rate changes by the same amount. We have already seen that
they are connected by the Jacobian J
∂xk ∂rk
J −1


Jkl = ⇐⇒ kl
= (5.2)
∂rl ∂xl
via the relation
X X
P V 01j J −1


P V 01i = IV 01j Jji ⇐⇒ IV 01i = ji
(5.3)
j j

where the sum in j runs over the node points of the yield curve, or in matrix notation
P V 01 = J T · IV 01 ⇐⇒ IV 01 = J −1T · P V 01 , (5.4)
where T denotes transposition.
In hedging, we want to replicate a product (or a portfolio of products) using the
instruments that are available in the market, in such a way that fluctuations on the
present value of our original product due to fluctuations in the underlying rates are
matched by the same fluctuation in the portfolio. In this way one can make a portfolio
immune to small changes in the yield curve.
In this framework, the hedging construction is achieved in the following way, by
looking both at the curve and at the product to be hedged.
Given a yield curve constructed out of n instruments, we can define an n-by-n matrix
Ψ whose columns are the IV01’s of each instrument. Since the zero rate at a generic
position i is sensitive to all the previous instruments up to and including i, due to the
bootstrap procedure, then the matrix Ψ will be upper triangular:
  ψ ψ12 . . . ψ1n

.. .. .. ..

11
 . . . .   0 ψ22 . . . ψ2n 
Ψ ≡ Ψ(I) =  Ψ Ψ . . . Ψ n =
  . ..  , (5.5)

 1 2
.. .. ..
.. .. .. ..  . . . 
. . . . 0 0 0 ψnn

16
where Ψi are the column vectors defined as the IV01 of the ith instrument, with ψij =
(Ψj )i . Also note that Ψ is a curve property since it depends only on the set of instruments
N . Moreover, the diagonal entries of Ψ are strictly non-zero, hence this matrix is always
invertible.
Now consider a product to be hedged. We can work out its IV01 as explained in
section 3. It is a vector whose entries depend on its complexity. Let’s call it ψ:
 
ψ1
 ψ2 
ψ= . , ψi ∈ R . (5.6)
 
 .. 
ψn

The hedging strategy is now to write the vector ψ as a linear combination of the col-
umn vectors of the matrix Ψ with some suitable vector of coefficients ω = (ω1 , ω2 , . . . , ωn )T ,
where T denotes transposition:

ψ = ω1 · Ψ1 + ω2 · Ψ2 + · · · + ωn · Ψn . (5.7)

More formally15 :
n
X
ψi = Ψij ωj , ωj ∈ R ⇐⇒ ψ = Ψ·ω, (5.8)
j=1

in components and matrix notation respectively. Since Ψ is invertible, for any given
product that needs to be hedged it is always possible to find a solution for ω. The
solution for the hedging vector is unique and given by

ω = Ψ−1 · ψ . (5.9)

These values for the ωi ’s tell us how to replicate the initial product. In particular, if we
choose the product ψ to be exactly one of the curve instruments in I, then we expect
all the components of the ω vector to vanish, except for the component corresponding to
the original product which should be equal to one.
In detail, in order to insure ourselves against moves in the present value of a product
due to changes in the underlying curve rates, we can build a portfolio with the same
instruments in I, each in the right abundance, such that the replicated portfolio will
move in the opposite direction:
n
M
δζ ∼ ωj ιj , ιj ∈ I . (5.10)
j=1

We will refer to this fact by saying that for each product the curve instruments spans the
complete risk space. We will be more formal in the next subsection.
Ψ and J are both curve properties. Once we know them, we can define another
matrix Π as the product
Π = JT · Ψ, (5.11)
15 We define the positions ω with a plus sign. That will allow us to replicate the original product or

portfolio of products. For hedging purposes one must use opposite positions, obtained by changing the
sign of ω.

17
which is called the PV01 hedging matrix in the IV01 hedging strategy for reasons that
will be clear in a moment. This matrix is also a curve property:

Π ≡ Π(I) . (5.12)

Since the Jacobian J is lower triangular and Ψ is upper triangular, Π will also be upper
triangular. We can now compute the PV01 of our original product as follows, using (5.4)
and (5.8). If we denote it by ξ
 
ξ1
 ξ2 
ξ= . , ξi ∈ R , (5.13)
 
 .. 
ξn

then we have:
ξ = Π·ω, (5.14)
in terms of the replicated portfolio. This justifies its name. The figure below summarizes
the process.
Ψ−1 /ω
ψ
JT
 Π=J T ·Ψ
ξ
Similarly, we could repeat the same construction and start from the PV01 instead.
We can define the matrix Π̂ whose columns are the PV01’s of each instrument:
  ξ ξ12 . . . ξ1n

.. .. .. ..

11
 . . . .   0 ξ22 . . . ξ2n 
Π̂ ≡ Π̂(I) =  Π̂ Π̂ . . . Π̂ n  =  ..
  ..  , (5.15)

 1 2 .. ..
.. .. .. ..  . . . . 
. . . . 0 0 0 ξnn

th
where
  Π̂i is the column vector containing the PV01 of the i instrument, with ξij =
Π̂j . Considerations similar to the previous case hold here as well. In particular, Π̂
i
is upper triangular and we can write the product’s PV01 as a linear combination of the
columns of Π̂
Xn
ξ= Π̂ij ω̂j ⇐⇒ ξ = Π̂ · ω̂ (5.16)
j=1

with coefficients ω̂j . Hence

ω̂ = Π̂−1 · ξ . (5.17)
The process diagram is now the following:

Π̂−1 / ω̂
ξO
JT −1
Ψ̂=(J T ) ·Π̂
ψ

18

−1
where Ψ̂ = J T · Π̂ is the IV01 hedging matrix in the PV01 hedging strategy, since
the product’s IV01 is recoved as

−1
−1
ψ = JT ·ξ = JT · Π̂ · ω̂ ≡ Ψ̂ · ω̂ . (5.18)

By gluing the two diagrams together one gets:

Ψ−1 /ω
@ψ (5.19)
Ψ̂
JT
 Π

ω̂ o ξ
Π̂−1

This diagram summarizes all the relations between the matrices and vectors that we have
introduced so far. From the diagram it follows e.g. that

ξ = Πω = J T Ψ̂ω̂ (5.20)

which we will solve in a moment.

Recall that each column vector of Π̂ is equal to the matrix product of J T with the
corresponding column of Ψ. This tells us that these two matrices are related by

Π̂ = J T · Ψ , (5.21)

which is exactly Π. Using this observation, together with the diagram (5.19), it is now
easy to show that the position vectors are also equal:

−1
−1
−1
ω̂ = Π̂−1 · ξ = Ψ−1 J T · ξ = Ψ−1 J T · J T · ψ = Ψ−1 J T ·JT ·Ψ·ω,

i.e.
ω̂ = ω . (5.22)
Similarly, the sets of matrices in the two cases are the same:

Π̂ = Π and Ψ̂ = Ψ . (5.23)

This implies that, independently of the hedging strategy (either using PV01 or IV01),
all the relevant quantities are invariant (for the position vector we have ω = ω̂, for the
curve PV01 matrix we have Π = Π̂, for the curve IV01 matrix we have Ψ = Ψ̂).
We can summarize all the relations involving the yield curve with the following dia-
gram:
curve space I
B /N (5.24)
δ δ δ
  JT

risk space Ψ /Π

hedge
 hedge
 
hedge

hedge space ω
which shows all the main features, namely

19
 • the translation from the original instruments and nodes to the risk space is obtained
by going to the derivative space and using first-order Taylor approximation16 ;

• the bootstrap B between the instrument set I and the curve node set N is translated
into the Jacobian matrix J T in the risk space;
• the hedging positions are strategy-invariant.
Based on this result, from now on we will use only one notation for the curve IV01
and PV01, as well as for the position vector.

5.1 Portfolios of several products

This reasoning can be easily extended to a portfolio of many products by using linearity.
Formally, a portfolio ζ contains m individual products ζk , with k = 1, . . . , m:

ζ = {ζ1 , . . . , ζk , . . . , ζm } , (5.25)

Then, by linearity, the total IV01 and PV01 will be just the corresponding sums:
m
X m
X
ψ= ψk , ξ= ξk . (5.26)
k=1 k=1

Using linearity and the fact that the matrices Ψ and Π are curve properties (i.e. they
are product-independent), it is easy to see that:

ψ =Ψ·ω and ξ =Π·ω (5.27)

where
m
X
ω = Ψ−1 · ψ = ωk with ωk = Ψ−1 · ψk . (5.28)
k=1

is the hedging vector for the whole portfolio expressed as a sum of the hedging vectors
for the single products.
Linearity and the fact that ωk is fully determined by the individual product ζk allow
us to compute the complete hedging vector ω for the entire portfolio and hence its full
IV01 and PV01. Thus the risk management for portfolios ζ with many products is
solved in terms of the risk management for the individual products ζk . As a last remark,
observe that this generalization is straightforward because of linearity of the portfolio,
whose value is linearly determined by the values of the underlying products inside the
portfolio itself.

5.2 Risk space

We have seen earlier that the Ψ and Π matrices are triangular. Let us focus on Ψ for the
moment. We can define the risk space for Ψ to be the range of Ψ:

RΨ ≡ Range(Ψ) . (5.29)
16 Or the exact integral formalism.

20
The range of Ψ is the span generated by the columns of Ψ. Since Ψ is triangular, it is
easy to see that its columns are linearly independent, hence they form a basis for the
range and therefore for the risk space (5.29).
However, we could have also started from the matrix Π. In this case we can define
the risk space for Π to be the range of Π:

RΠ ≡ Range(Π) . (5.30)

Then, the columns of Π form a basis. Inspection of equation (5.4) shows that
 
RΠ = J T RΨ = span{J T Ψ1 , J T Ψ2 , . . . , J T Ψn } . (5.31)

In fact, for any product,

X X X
J T · Ψi ω i = J T · Ψi ω i = J T · ψ .


ξ =Π·ω = Πi ω i = (5.32)
i i i

Triangular matrices are not diagonalizable, but one can still find an orthogonal (and
orthonormal) basis {ei }ni=1 out of the columns of Ψ. Given the shape of Ψ, we can guess
that one orthonormal basis is the canonical basis
     
1 0 0
 0   1   0 
e1 =  .  , e2 =  .  , . . . , en =  .  . (5.33)
     
 ..   ..   .. 
0 0 1
We can easily check that this is indeed the case. Alternatively, one can derive the same
result by applying the standard Gram–Schmidt algorithm [20] to triangular matrices.
So, if we denote by Ψ1 , Ψ2 , . . . , Ψn the columns of Ψ, then an orthonormal basis {ei }ni=1
is constructed recursively as
i−1
X uj · Ψi ui
ui = Ψi − uj and ei = 1/2
, (5.34)
u · uj
j=1 j (ui · ui )

for i = 1, . . . , n, and indeed we get back (5.33).

Looking back at (5.31), we can see that the more stringent relation

RΠ = RΨ (5.35)

holds. One way to see this is to note that these two spaces have a common set of
orthonormal basis, namely the canonical basis (5.33), since both Π and Ψ are triangular.
Another way is to recall that the two sets of basis for RΠ and RΨ respectively are related
by a linear transformation and by definition they span the same space. Moreover, these
spaces are equivalent to the whole Rn , since they all have the canonical vectors (5.33) as
basis.
One can be even more explicit and write down the component of the vectors ψ and
ξ with respect to the canonical basis. Using some manipulations for the indices we find
the linear combinations
 
X Xn Xi Xn Xn Xn
ψ= ωi Ψi = ωi ej ψji = ei  ψij ωj ≡
 ei cψ
i (5.36)
i i=1 j=1 i=1 j=i i=1

21
 Pn
with coefficients cψ
i = j=i ψij ωj for ψ, and
 
X n
X i
X n
X n
X n
X
ξ= ω i Πi = ωi ej ξji = ei  ξij ωj  ≡ ei cξi (5.37)
i i=1 j=1 i=1 j=i i=1

Pn
with coefficients cξi = j=i ξij ωj for ξ, where ξij = k J T ik ψkj .
P

As a final remark, observe that for any given product all the other hedge quantities are
fixed by the curve via its instruments. This means that the choice of the instrument set
I is crucial for hedging. In fact, in the diagram (5.24) everything starts from I and the
resulting hedge or replication of a product (or a portfolio of products) strongly depends
on which instruments are part of I. Only those instruments relevant for measuring value
and risk of a product should be used. So I should be chosen with care, because it dictates
the value and risk one will perceive17 .

6 Main points
In this section we would like to summarize, generalize and collect together the main
points of the discussion so far.
We start from an obvious result.

Theorem 1 (Bootstrapping prices back market). Present Values of any instrument using
the bootstrapped yield curve is equal to market quoted value.
Proof. This is true by construction of the yield curve. The set of discrete input data are
the market data and interpolation is used to determine a value when the data points are
not available. Thus, the yield curve passes through all the market data. Consequently,
in the present-value calculation for any instrument where the yield curve is used the
computed value will match the market value.
Observe that the statement of this theorem will not be true if other methods were used
to generate the yield curve. In particular, using best-fit-like approaches (such as Nelson-
Siegel-Svensson) would give us a smooth curve which will not however pass through the
data (market) points.
The next theorem shows the relation between the PV01 and the IV01, which are two
sides of the same coin, since they are connected by a change of coordinates.
Theorem 2 (PV01 and IV01). For a given yield curve γ, the corresponding Jacobian
(3.2), which relates zero rates to the underlying instrument rates, also relates PV01 and
IV01 for any product ζ. If ξ and ψ are the product’s PV01 and IV01 vectors respectively,
then
ξ = JT ·ψ ⇐⇒ ψ = J −1T · ξ . (6.1)
Proof. The theorem was shown to be true in equation (5.4).
The next theorem is about risk management and replication for any product or port-
folio of products.
17 In principle, many instruments are available for hedging (e.g.in terms of maturities for IRS contracts),

however hedging with instruments unrelated to the curve will typically result in bad positions, since in
general the yield curve will not go through the desired value, unless it is a node point.

22
Theorem 3 (Ψ spans complete risk space). Let ζ be any product or a portfolio of prod-
ucts, and I = {ι1 , ι2 , . . . , ιn } be the set of instruments used to construct the yield curve
γ. Then the column vectors of the matrix Ψ as defined in (5.5) are a basis for the risk
of any product and portfolios of products:

RΨ = Range(Ψ) . (6.2)

Explicitly, the product’s IV01 vector ψ can be written as a linear combination of the form

ψ = Ψ·ω, (6.3)

for a given ω. The change in the portfolio’s value ζ −→ ζ + δζ due to changes in the
instruments rates is then replicated by the instruments in I together with the vector ω of
linear coefficients as
Mn
δζ ∼ ωj ιj . (6.4)
j=1

Similarly, the product’s PV01 vector ξ can be written as a linear combination of the form

ξ = Π·ω, (6.5)

where Π is given by
Π = JT · Ψ. (6.6)
Moreover, the matrices Ψ and Π are curve-dependent only and the vector ω is independent
of the hedging strategy. Finally,
RΠ = RΨ , (6.7)
i.e. the risk space based of Π coincides with the one based on Ψ.
Proof. This is straightforward from section 5. For one single product ζ, the result follows
from equations (5.6), (5.9), (5.10) and (5.13). The statement about the independence
of the hedging strategy is a consequence of equation (5.22). The statement about the
basis and the one about equality between the two risk spaces derive from the arguments
in subsection 5.2. If ζ is a portfolio of products, then one needs to use the analogous
equations from subsection 5.1.

Observe that once the product ζ and the curve γ have been specified the hedging
vector ω, and hence the replication strategy, is fully determined. Also note that what
is being replicated is the change in the value of the portfolio, not the portfolio itself. In
fact, the portfolio’s present values is generically different from the weighted sum of the
instruments’ present values with weights given by the ω’s. Similar statements, mutatis
mutanda, hold true for the risk space RiskΠ .
We finish with two remarks, one about which instruments should be used to construct
the yield curve and for hedging, and one about the use of numerical calculations in this
framework.
Remark 1 (Choice of curve instruments). The choice of the instrument set I dictates
the resulting hedge or replication of a product or a portfolio of products. Only those
instruments relevant and available for hedging should be used.

23
Proof. The proof is trivial, since the hedge strongly depends on the choice for the curve
instruments, as we mentioned at the end of subsection 5.2.

Remark 2 (No numerical computations). There are no numerical computations in this

framework.
Proof. Only analytic formulas were used in this paper.
Observe that the absence of numerical computation is important for many reasons.
First of all, numerical computations are intrinsically associated to approximation errors
which, despite very small, could lead to significant losses for financial institutions who
work with large amounts on a daily basis. Banks’ portfolios typically vary from hundreds
of billions to a few trillions of Euros (≈ 1011 − 1012 ), hence a small numerical error in
the hedging strategy, e.g. O(10−9 ), can produce sizable losses.
Secondly, typically a numerical problem could be solved by using various techniques,
which perform very well when dealing with a specific aspect, but quite badly with other
aspects. As an example, when computing derivatives one can choose to implement it as
a right derivative, or as a left derivative, or as a combination of left and right derivative.
Clearly this choice is crucial for piece-wise function since it can affect the value of the
derivative at the boundary points.
Lastly, when dealing with extremely large or extremely small numbers (e.g. close to the
upper or lower limits of the numerical range for a given machine), numerical computations
may have stability issues and produce machine-dependent results.
For all these reasons, when possible analytic formulas should be used.

7 Conclusion
In this paper we have considered the problem of evaluating and managing risk for interest
rate derivatives, in particular for cash and swap instruments. The crucial ingredient is
the construction of the yield curve. It depends on a set I of input instruments which
are uniquely related to the set N of input zero rates via the bootstrap procedure. The
detailed way of how to extract the zero rates from the instruments has been done here for
cash and swap instruments only, but it can be generalized further to any other instrument
that is part of the curve construction. The starting point is always the formula for the
present value of the product, which gives the desired relationship between zero rates and
instrument rate.
A crucial ingredient in this framework is the yield curve. The yield curve depends on
the specific interpolation method used to generate the values away from the node points.
A poor interpolation choice will result in a poor yield curve, where quality criteria such as
smoothness, locality, stability and positivity are problematic to have. We have considered
only a few interpolation techniques here. The linear method is easy to use and implement,
but it is often not accurate enough for our purposes (e.g. it gives serious problems when
the derivatives of the curve are needed, since they will be then discontinuous at the node
points). Other methods typically suffer of an unnatural wiggly/zig-zag behaviour which
is not intrinsic of the initial data, but is a spurious effect of the method. This problem
can be corrected by monotonicity constraints, and they have been implemented in some
of the splines. Many formulas for basic interpolation uses where already available in the
literature.

24
 Interest rate risk management is strictly related to hedging. Both concepts are based
on the knowledge of how the value of a particular product changes as a consequence
of a change in the yield curve due to small jumps, typically of one basis-point size, in
the input rates. This knowledge is mathematically encoded in quantities such as the
PV01 and the IV01 of a product, which are defined as the change in the product’s price
when the input rates change by one basis point. In order to compute them, one can
use a numerical approach. However, we have shown that it is possible to compute them
analytically, either exactly or at first-order approximation, by focusing on the derivatives
of the product’s present value with respect to the yield curve zero rates and instrument
rates. These derivatives in turn are expressed in terms of the derivatives of the curve
with respect to its inputs. The latter derivatives are interpolation-dependent. While
many useful interpolation formulas were already known explicitly in the literature, this
was not the case for these derivatives. We have given their analytic expression in the
main body of the paper as well as in the appendix.
PV01 and IV01 are not independent but related by a change of coordinates, which
is mathematically expressed by the Jacobian matrix between instrument rates and zero
rates. We have seen that the Jacobian is a triangular matrix. We have also defined
the curve PV01 and IV01 matrices and found that they are triangular too. This is a
consequence of the bootstrap procedure. We have showed that the whole risk space for
any arbitrary product is ultimately determined by the yield curve, since the change of
the present value for any product due to changes in the yield curve can be reproduced by
a portfolio of instrument consisting exactly of those instruments used to bootstrap the
yield curve. In this sense, the yield curve -through its instruments- spans the complete
risk space. Consequently, the choice for the instruments used to bootstrap the curve is
crucial and only those instruments should be used that are relevant and available for
hedging.
Clearly further generalizations of what we have described here could be considered in
the future. One obvious extension of this work would be to include many other instru-
ments in the treatment besides cash and swaps. Moreover, one could ask what happens
when a multi-curve framework is used instead of the single-curve one. This would be
particularly relevant, since the multi-curve is becoming more and more popular after the
credit crunch crisis in 2008. Furthermore, many financial institutions have moved to the
wave or scenario method for handling risks. We have used a more standard approach,
but we believe that our results could be easily generalizable to this case as well. Finally,
in this paper we have considered only linear products, namely those products where con-
vexity effects can be neglected, however it would be interesting to include convexity and
higher-order effects18 into the picture too by working with the exact integral expression
and to estimate the error of the approximation in concrete cases. We leave all these point
for the future.
18 Observe that, within our framework, at each order in perturbation theory the derivatives are always

known exactly. In practice, the highest order that can be reached in the perturbation expansion is limited
by the specific interpolation method used to construct the curve. Typically, spline-like methods result
in curve which are C 1 , sometimes C 2 , with vanishing higher derivatives, so the Taylor sum stops quite
soon.

25
Acknowledgements
We would like to thank Patrick Hagan, Julian Wiseman, Stratos Nikolakakis and Sander
Scheerders for carefully reading the manuscript and providing us with useful feedback,
as well as Laurent Lefort for pointing out some typos.

References
[1] C. Nelson, A. Siegel (1987), Parsimonious modeling of the Yield Curve, Jour-
nal of Business 60, no 4, pp. 473-489, http://www.math.ku.dk/~rolf/teaching/
NelsonSiegel.pdf.

[2] L. Svensson (1995), Estimating Forward Interest Rates with the Extended Nelson and
Siegel Method, Sveriges Riskbank Quarterly Review 3, http://www.diva-portal.
org/smash/get/diva2:342900/FULLTEXT01.pdf.
[3] J. Wiseman (1998) The Magpie Yield Curve Model, Technical report, J.P. Morgan
Fixed Income Research.
[4] J. Wiseman (2012) The Magpie Yield Curve Model at SG, Technical report, Societe
Generale Cross Asset Research - Fixed Income.
[5] T. Bjork, B. Christensen (1997), Interest Rates Dynamics and Consistent For-
ward Rates Curve, University of Aarhus, http://www.blackwell-synergy.com/
doi/abs/10.1111/1467-9965.00072.
[6] J. James, N. Webber (2000), Interest Rate Modelling, Wiley, ISBN: 978-0-471-97523-
6.
[7] Bolder, David, and David Stréliski (1999), Yield Curve Modelling at the Bank of
Canada, Available at SSRN 1082845 http://epe.lac-bac.gc.ca/100/200/301/
bankofcanada/technical_reports-ef/no084/tr84.pdf.
[8] P. S. Hagan, G. West (2006), Interpolation Methods for Curve Construction,
Applied Mathematical Finance, 13: 2, 89 - 129 http://dx.doi.org/10.1080/
13504860500396032, http://kalx.net/dsS2011/HagWes2006.pdf.

[9] P. S. Hagan and G. West (2008), Methods for Constructing a Yield Curve, Wilmott
Magazine, 70 - 81 http://www.math.ku.dk/~rolf/HaganWest.pdf.
[10] Whittaker, J Gregg (1987) Interest rate swaps: Risk and regulation Economic review,
volume= 72 - 3, pages 3-13, Federal Reserve Bank of Kansas City https://www.
kc.frb.org/Publicat/Econrev/EconRevArchive/1987/1q87whit.pdf.
Reitano, Robert R. ”Non-parallel yield curve shifts and immunization.” The Journal
of Portfolio Management 18.3 (1992): 36-43.
[11] Reitano, Robert R (1992) Non-parallel yield curve shifts and immuniza-
tion, The Journal of Portfolio Management volume 18-3, pages 36-43
http://www.robertrreitano.com/files/Non-parallel%20Yield%20Curve%
20Shifts%20and%20Immunization.pdf.

26
[12] Bianchetti, M. (2008). Two Curves, One Price Pricing & Hedging Inter-
est Rate Derivatives Decoupling Forwarding and Discounting Yield Curves.
arXiv:0905.2770v2 [q-fin.PR] http://arxiv.org/abs/0905.2770.

[13] M. Bianchetti, M. Carlicchi (2008). Interest Rates After The Credit Crunch:
Multiple-Curve Vanilla Derivatives and SABR, arXiv:1103.2567v2 [q-fin.PR] http:
//arxiv.org/abs/1103.2567.
[14] C. P. Fries (2013). Curves and Term Structure Models. Definition, Cali-
bration and Application of Rate Curves and Term Structure Market Mod-
els. preprint, 2013 http://www.fm.mathematik.uni-muenchen.de/download/
publications/fries_curves_tsm.pdf.
[15] J. C. Hull (2014), Options, Futures and Other Derivatives, 9th edition, Pearson,
ISBN-13: 978-0133456318.

[16] R. Deaves, M. Parlar (2000), A Generalized Bootstrap Method to Determine the Yield
Curve, Applied Mathematical Finance 7, 257- 270 http://www.tandfonline.com/
doi/abs/10.1080/.VC049CmSzPY#.VC1SzymSwuE.
[17] J. M. Hyman (1983), Accurate Monotonicity Preserving Cubic Interpo-
lation, SIAM J. Sci. Stat. Comput. Vol. 4, No. 4, 645 - 654 http:
//www.researchgate.net/publication/230676072_Accurate_monotonicity_
preserving_cubic_interpolation/file/d912f50c0a8c9585cd.pdf.
[18] M. Maio, J. de Jong (2014), On the Non-differentiability of Hyman Monotonic-
ity Constraint, pre-print nr UD2014-01 http://uglyduckling.nl/library_files/
PRE-PRINT-UD-2014-01.pdf.

[19] M. Maio, J. de Jong (2014), Curiosities on Monotone Preserving Cu-

bic Spline, pre-print nr UD2014-03 http://uglyduckling.nl/library_files/
PRE-PRINT-UD-2014-03.pdf.
[20] G. B. Arfken, H. J. Webber, F. E. Harris (2012), Mathematical Methods for Physi-
cists, 7th ed., Elsevier Academic Press, ISBN: 978-0-12-384654-9.

[21] M. Maio (2014), Interpolation methods and the Hagan-West paper, blog
post, http://www.uglyduckling.nl/financial-engineering/yield-curve/
interpolation-methods-and-the-hagan-west-paper/.
[22] P. F. du Preez, E. Mare (2013), Interpolating Yield Curve Data in a Manner that
Ensures Positive and Continuous Forward Curves, SAJEMS NS 16 (2013) No 4:395-
406 http://sajems.org/index.php/sajems/article/view/388.

Appendix

27
A Hagan-West: Relevant Formulas
In this appendix we give the collection of relevant formulas for each of the four regions.
The method starts from the realistic assumption that forward rates are constant in the
intervals [ti−1 , ti ], with 1 ≤ i ≤ n. Their values are specified as usual by:
ri ti − ri−1 ti−1
fid = (1 ≤ i ≤ n , r0 = t0 = 0) . (A.1)
ti − ti−1

These discrete rates fid belong to the entire interval [ti−1 , ti ]. Then one defines the
instantaneous forward rates at ti as:
ti − ti−1 d ti+1 − ti d
fi = fi+1 + f (i = 1, 2, . . . , n − 1) (A.2a)
ti+1 − fi−1 ti+1 − fi−1 i
1
f0 = f1d − (f1 − f1d ) (A.2b)
2
1
fn = fnd − (fn−1 − fnd ) . (A.2c)
2
where the values f0 = f (t0 ) and fn = f (tn ) are such that f 0 (0) = f 0 (tn ) = 0, and
interpolates these instantaneous forward rates with an interpolator f (t) whose average
on each interval is the discrete rate. In addition f (t) is constructed to be positive and
continuous19 . Continuity is automatic, while positivity is enforced by the substitution:

f0 −→ collar(0, f0 , 2f1d ) (A.3a)

fi −→ collar(0, fi , 2 min(fid , fi+1

d
)) (i = 1, 2, . . . , n − 1) (A.3b)
fn −→ collar(0, fn , 2fnd ) (A.3c)
where collar(a, b, c) = max(a, min(b, c)).
In addition, one can enforce monotonicity and convexity. This is achieved by noting
that the interpolant f (t) is defined modulo a smooth function g(t) whose average is zero
on each interval. Then f (t) will be given by

f (t) = fid + g(t) , (A.4)

with Z ti Z 1
g(t)dt = g(x)dx = 0 . (A.5)
ti−1 0

Define: g0 = g(ti−1 ), g1 = g(ti ) and x = tt−t i−1

i −ti−1
. Inspection of the derivative of g allows
us to construct this function in four different regions. We will give explicit expressions
below.
Define:
t − ti−1
x= , t ∈ [ti−1 , ti ] . (A.6)
ti − ti−1
For the four regions we have:
19 Note that the forward function f (t) in not constructed to be differentiable. In fact, its derivative

can have discontinuities, as already observed in e.g. [21] and [22].

28
i) if g0 > 0, − 12 g0 ≥ g1 ≥ −2g0 and g0 < 0, − 21 g0 ≤ g1 ≤ −2g0 , then:

g(t) = g0 (1 − 4x + 3x2 ) + g1 (−2x + 3x2 ) (A.7)

It = (ti − ti−1 ) g0 (x − 2x2 + x3 ) + g1 (−x2 + x3 )

 
(A.8)
∂It
= (ti − ti−1 ) g00 (x − 2x2 + x3 ) + g10 (−x2 + x3 )
 
(A.9)
∂rj

ii) if g0 < 0, g1 > −2g0 and g0 > 0, g1 < −2g0 , then:

g0 g1 + 2g0
η =1+3 = (A.10)
g1 − g0 g1 − g0

for 0 ≤ x ≤ η
(
g0
g(x) = 
x−η
2 (A.11)
g0 + (g1 − g0 ) 1−η for η < x ≤ 1
(
h(ti − ti−1 )g0 x for 0 ≤ x ≤ η
It = (A.12)
i
(x−η)3
(ti − ti−1 ) g0 x + 31 (g1 − g0 ) (1−η)2) for η < x ≤ 1

(ti − ti−1 )g00 x



 for 0 ≤ x ≤ η


∂It  h 3
= (ti − ti−1 ) g00 x + 31 (g10 − g00 ) (x−η)
∂rj 
 (1−η)2 + i
2
− g0 )η 0 (2x + η − 3) (x−η)
 1
3 (g1 for η < x ≤ 1


(1−η)3
(A.13)
iii) if g0 > 0, 0 > g1 > − 12 g0 and g0 < 0, 0 < g1 < − 21 g0 , then:
g1
η=3 (A.14)
g1 − g0
(  2
η−x
g(x) = g1 + (g0 − g1 ) η for 0 < x < η (A.15)
g1 for η ≤ x < 1
( h  3 3
i
(ti − ti−1 ) g1 x + 31 (g0 − g1 ) η −(η−x)η 2 for 0 < x < η
It = (A.16)
1

(ti − ti−1 ) g1 x + 3 (g0 − g1 )η for η ≤ x < 1
 h  3 3



 (ti − ti−1 ) g10 x + 13 (g00 − g10 ) η −(η−x)η 2 +
  2
i
∂It  1
(g − g )η 0
1 − (η−x)
(η + 2x) for 0 < x < η
= 3 0 1 η 3
∂rj 


(ti − ti−1 ) g10 x + 31 (g00 − g10 )η + 13 (g0 − g1 )η 0
  
for η ≤ x < 1

(A.17)
iv) if g0 ≥ 0, g1 ≥ 0 and g0 ≤ 0, g1 ≤ 0, then:
g1 g0 g1
η= A=− (A.18)
g0 + g1 g0 + g1

29
   2
 A + (g0 − A) η−x

for 0 < x < η
η
g(x) =  2 (A.19)
 A + (g1 − A) 1−ηx−η
 for η < x < 1
  
1 (η3 −(η−x)3 )
(ti − ti−1 ) Ax + 3 (g0 − A) for 0 < x < η


η2
It = h 3
i
(ti − ti−1 ) Ax + 31 (g0 − A)η + 13 (g1 − A) (x−η) for η < x < 1


(1−η)2

 (A.20)
(η3 −(η−x)3 )



 (t i − t i−1 ) A0 x + 31 (g00 − A0 ) η2 +

  i
 2
1
(g0 − A)η 0 1 − (η−x) (η + 2x) for 0 < x < η



 3 η 3
∂It 
=
∂rj (ti − ti−1 ) A0 x + 31 (g00 − A0 )η + 13 (g0 − A)η 0 +
 

 3
0 (x−η)

1 0
3 (g1 − A ) (1−η)2 +




 2
i
1
(g − A)η 0 (x−η) (2x + η − 3)


3 1 (1−η)3 for η < x < 1
(A.21)
dη
Here a prime typically denotes a derivative w.r.t. rj , e.g. η 0 ≡ dr j
. Above we have
used the fact that g0 ≡ g(ti−1 ) = f (ti−1 ) − fid and g1 ≡ g(ti ) = f (ti ) − fid , to compute
the derivatives g00 ≡ dg 0 dg1
drj and g1 ≡ drj in terms of the derivative of f and f . This
0 d

calculation is in principle straightforward but cumbersome because of all the constraints

(i.e. positivity, monotonicity and convexity) that have been enforced on the forward
curve. We will not give the formulas for g00 and g10 here.

30