Building a Smooth Yield Curve

University of Chicago

Jeff Greco

email: [email protected]
 Preliminaries

As before, we will use continuously compound-

ing Act/365 rates for both the zero coupon rates
and forward rates. Time is measured in years from
now. Recall the following quantities:

Name Symbol Formulas

discount e−r(t)·t
P (t) Rt
factor e− 0 f (s)ds
forward e−f (t,T )(T −t)
discount P (t, T ) P (T )
factor P (t)
zero − 1t log P (t)
r (t) 1 R t f s ds
rate t 0 ( )
discrete r(T )·T −r(t)·t
forward f (t, T ) T −t
1
− T −t log P (t, T )
rate
instant- d log P t
− dt ()
aneous
f (t) limτ ↓0 f (t, t + τ )
forward d r t ·t
rate dt ( ( ) )

1
 Mechanics

Let
0 = t0 < t1 < · · · < tn
be an ordered set of times. A yield curve is rep-
resented by any of the following:

• Discount factors P (t1) , P (t2) , . . . , P (tn),

• Zero rates r (t1) , r (t2) , . . . , r (tn), or

¡ ¢
• Forward rates f (t0, t1) , f (t1, t2) , . . . , f tn−1, tn .

These are all equivalent, as can be seen by the

relations

P (ti) = e−r(ti)·ti ,
¡ ¢ ¡ ¢¡ ¢
r (ti) · ti = r ti−1 · ti−1 + f ti−1, ti ti − ti−1 .

2
 What is Smoothness?

Smoothness is a measure of the degree to which

a function wriggles. A smoothness penalty is of-
ten defined as
Z ³ ´2
00
f (t) dt
or
Z ¯ ¯
¯ 0 ¯
¯f (t)¯ dt.

The zero rates are averages of forward rates

Z
1 t
r (t) = f (s) ds
t 0
and are therefore smoother. This implies that if
we smooth the forward rates, the zero rates will
also be smooth. So what will be the criterion of
smoothness for our discrete sequence of forward
rates?
¡ ¢
f1 = f (t0, t1) , f2 = f (t1, t2) , . . . , fn = f tn−1, tn
3
We shall choose a discrete version of one of the
continuous time measures we have already seen.
One possibility is that a forward rate should be
“close” to its neighbors, e.g.
n
X
|fi − fi−1|
i=2
should be small. The absolute value is tricky to
deal with, so we will use the squares of the differ-
ences instead
n ¡
X ¢
S(f ) = fi − fi−1 2 .
i=2
Another possibility is
n−1
X ³ ´2
fi−1 − 2fi + fi+1 .
i=2
 Why is Smoothing Possible?

The prices of the traded input instruments (de-

posits, Eurodollar futures, and swaps) are func-
tions of the forward rates. The forward rates are
not uniquely determined by these prices. There
are always extra degrees of freedom. When we
bootstrapped the LIBOR curve in the previous lec-
ture, we chose a constant forward rate interpola-
tion method in order to specify the curve.

We used a total of 23 constraints (instruments) to

construct the curve. The number of unknowns
(forward rates) is equal to the number of knot times.
A knot time is a time whose corresponding dis-
count factor impacts the value of any of the in-
put instruments:

4
 • Deposit begin and end dates

• Futures IMM dates and end dates

• Swap fixed cashflow payment dates

We have at least 41 knot times and only 23 con-

straints. That leaves at least an extra 18 degrees
of freedom! We will use them for smoothing.

5
 The Dangers of Overfitting

Underfitting is when the number of constraints is

greater than the number of free variables. This is
relatively benign. Generally, not every constraint
can be satisfied. Fewer constraints than free vari-
ables results in overfitting. Ad-hoc procedures for
fixing extra variables usually result in unpleasant
consequences. Extra variables go haywire. Over-
fitting is extremely widespread but usually hidden!
Watch out for it in

• Curve building

• Model calibration

• Market (BGM) models

• Everywhere else

6
Overfitted models are unstable, have little explana-
tory and predictive power, yet provide a false sense
of security (because all inputs are matched).

Ways to deal with overfitting:

• Recognize it

• Avoid it

• Formulate a goal that extra parameters are

designed to achieve

7
 A Simple Example

Consider two instruments

• A Eurodollar deposit

• A Eurodollar future

Assume that the deposit settles at time t0 and

pays at time t2. The Eurodollar covers the time
period from t1 to t3. The times are arranged in
order t0 < t1 < t2 < t3. Assume all day bases are
Act/365. Define

τ1 = t1 − t0,
τ2 = t2 − t1,
τ3 = t3 − t1.

8
The three forwards involved are

f1 = f (t0, t1) ,
f2 = f (t1, t2) ,
f3 = f (t2, t3) .
Two equations must be satisfied by the forward
rates,
exp (f1τ1 + f2τ2) − 1
D = ,
τ1 + τ2
exp (f2τ2 + f3τ3) − 1
F = ,
τ2 + τ3
where D is the deposit rate and F is the rate im-
plied by the futures price. Expressed in term of
forward rates, the constraints become

τ1f1 + τ2f2 = log (D (τ1 + τ2) + 1)

τ2f2 + τ3f3 = log (F (τ2 + τ3) + 1) .
We have two equations and three unknowns. We
want the smoothest curve possible. Recall our
definition of smoothness. The goal is to minimize

S (f 1 , f 2 , f 3 ) = ( f 2 − f 1 )2 + (f 3 − f 2 )2
using the extra degree of freedom. By direct sub-
stitution
log (D (τ1 + τ2) + 1) − f2τ2
f1 =
τ1
log (F (τ2 + τ3) + 1) − f2τ2
f3 = .
τ3
Plugging into S (f1, f2, f3) leaves a one dimen-
sional unconstrained optimization problem. Find
f2 where the minimum of
ÃÃ ! !2
τ log (D (τ1 + τ2) + 1)
S (f 2 ) = 1 + 2 f2 −
τ1 τ1
ÃÃ ! !2
τ2 log (F (τ2 + τ3) + 1)
+ 1+ f2 −
τ3 τ3
is attained.

Do this as an exercise!
 The Method in Detail

Step 1: Identify the instruments used in construct-

ing the curve

• Deposits / LIBOR rates

• Eurodollar futures

• Swaps

We will call them “curve instruments”.

9
Step 2: Identify knot times

The knot times are characterized as the set of

times whose corresponding discount factors im-
pact the value of the curve instruments. Knot
times include

• Today

• Deposit accrual begin dates and end dates

• Eurodollar futures underlying forward LIBOR be-

gin and end dates

• Swap settlement dates and fixed rate pay-

ment dates

Sort the knot times

0 = t0 < t1 < · · · < tn,
breaking the timeline into non-overlapping peri-
ods.
10
Step 3: Assign a forward rate (free variable) to
each time period

¡ ¢
f1 = f (t0, t1) , f2 = f (t1, t2) , . . . , fn = f tn−1, tn ,
establishing n independent variables.

Step 4: Express the curve instrument prices as func-

tions of the forward rates

• First express the curve instrument prices through

the discount factors at the knot times
P (t1) , P (t2) , . . . P (tn) .
This is always possible by definition of the knot
times.

• Second express the discount factors in terms

of the forward rates. This can be done recur-
sively
P (t0) = 1,
¡ ¢ ¡ ¡ ¢¢
P (ti) = P ti−1 exp −fi · ti − ti−1 .
11
Step 5: Define constraints

The smoothed curve should imply curve instru-

ment prices which are a close match for market
quotes. In other words

Vi (f1, f2, . . . , fn) = vi, i = 1, 2, . . . , m.

Here V1 (f ) , V2 (f ) , . . . , Vm (f ) are the prices as
a function of the forward rates, and v1, v2, . . . vm
are the market quotes we need to match. If m =
n we are done; there are no extra degrees of
freedom to use for smoothness. If m > n we have
too many constraints; we cannot match all the
prices. If, as is the usual case, m < n we can use
the extra degrees of freedom for smoothing.
Step 6: Formulate a smoothness criteria

Use
n ¡
X ¢2
S (f ) = fi − fi−1 ,
i=2
n−1
X ³ ´2
S (f ) = fi−1 − 2fi + fi+1 ,
i=2
or something else.

12
Step 7: Solve a non-linear constrained optimiza-
tion problem

Mathematically we must find f such that S (f ) is

a minimum among all possible solutions V (f ) =
v. This is a very hard problem. It is a multidi-
mensional (50+) minimization with non-linear con-
straints. This pushes the limit of numerical tech-
niques.

An alternative is to get rid of the hard constraints.

Use soft constraints instead. Replace V (f ) = v
with the mispricing objective function
n
X
M (f ) = (Vi (f ) − vi)2 .
i=1
The closer the fit to market prices, the smaller the
mispricing M (f ). The reformulated problem is to
find f such that

Ω (f ; w) = wS (f ) + M (f )
13
attains its minimum. We have included a control
variable w to specify the relative importance of
smoothness versus mispricing. This is a much sim-
pler unconstrained optimization problem. Even
Excel can handle it! The curve will no longer ex-
actly price the input instruments. This is not nec-
essarily bad – there is always a bid/ask spread.
We can fall within the bid/ask spread by choos-
ing the right value for w. Typical bid/ask spreads
are

1 bp
• Deposits: 2

• Eurodollars: 1 -
4 2
1 bp

• Swaps: 1-2bp
 Advanced: Connecting the Dots

We know the discrete forward rates for the knot

times. But, how do we value other instruments
that do not necessarily “hit” the knot times? In-
terpolate.

Assuming flat forward rates between the knot times

is the easy way out. We have worked hard to
make the curve “smooth”. Why not connect the
dots smoothly? There are several choices for in-
terpolation

• Polynomial

• Polynomial spline

• Exponential spline

• Other families of smooth functions

14
How do we choose? Be consistent.

For instance, if you chose

n−1
X ³ ´2
S (f ) = fi−1 − 2fi + fi+1
i=2
as your smoothness constraint, then why not try to
R ¡ 00 ¢2
extend this by minimizing f (t) dt? In other
words, find a twice differentiable function f (·)
such that
Z tn ³ ´2
00
f (s) ds (1)
0
is a minimum among all possible solutions
Z ti
1
f (s) ds = fi, i = 1, 2, . . . n. (2)
ti − ti−1 ti−1

15
In order to achieve this we will invoke the follow-
ing theorem∗.

Theorem. The term structure of instantaneous for-

ward rates f (t) , 0 ≤ t ≤ T , that minimizes the
smoothness criteria (1) subject to the constraints
(2) is a forth-order polynomial spline with knot times
{ti}ni=0 .

The coefficients of the spline are determined by

solving a linear system of equations.

∗ Donald R. van Deventer and Kenji Imai. Financial Risk An-

alytics: A Term Structure Model Approach for Banking, In-
surance and Investment Management, McGraw-Hill, 1996.

16
Final Words of Wisdom

• It is not always appropriate to smooth the yield

curve∗.

• If the market data used to construct the curve

is bad, no amount of smoothing will fix it.

• If different curve instruments have incompat-

ible prices (e.g. on-the-run and off-the-run
Treasuries), smoothing will make matters worse.

• When building a curve that has a scarce set

of curve instruments (e.g. Treasury curve), some
extra inputs should usually be used to “get it
rightӠ.
∗ Forinstance, when calculating price sensitivities to buck-
eted discrete forward rates, yield curve shocks will bleed
into other buckets.
† E.g.,
Treasury curve shape between 10 and 30 years
forced to match the shape of the LIBOR curve.
17