stochastic modeling of yield curve shifts using functional...

44
Stochastic modeling of yield curve shifts using functional data analysis RASMUS R EHN Master of Science Thesis Stockholm, Sweden 2014

Upload: others

Post on 19-Mar-2020

10 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

Stochastic modeling of yield curve shifts using functional data analysis

R A S M U S R E H N

Master of Science Thesis Stockholm, Sweden 2014

Page 2: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts
Page 3: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

Stochastic modeling of yield curve shifts using functional data analysis

R A S M U S R E H N

Master’s Thesis in Mathematical Statistics (30 ECTS credits) Master Programme in Mathematics (120 credits) Royal Institute of Technology year 2014 Supervisor at KTH was Boualem Djehiche Examiner was Boualem Djehiche

TRITA-MAT-E 2014:38 ISRN-KTH/MAT/E--14/38--SE Royal Institute of Technology School of Engineering Sciences KTH SCI SE-100 44 Stockholm, Sweden URL: www.kth.se/sci

Page 4: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts
Page 5: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

Stochastic modeling of yield curve shifts using

functional data analysis

Rasmus Rehn

[email protected]

Royal Institute of Technology

June 16, 2014

Abstract

This thesis approaches the problem of modeling the multivariate distribution

of interest rates by implementing a novel tool of statistics known as functional

data analysis (FDA). This is done by viewing yield curve shifts as distinct but

continuous stochastic objects defined over a continuum of maturities. Based on

these techniques, we provide two stochastic models with different assumptions

regarding the temporal dependence of yield curve shifts and compare their per-

formance with empirical data. The study finds that both models replicate the

distributions of yield changes with medium- and long-term maturities, whereas

none of the models perform satisfactory at the short segment of the yield curve.

Both models, however, appear to accurately capture the cross-sectional depen-

dence.

Page 6: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts
Page 7: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

Acknowledgements

I would like to express my great appreciation to my tutor Boualem Djehiche

at Royal Institute of Technology for introducing me to functional data analysis

as a modeling tool for finance and risk management. Also, I would like to

thank Jonas Nilsson, Fredrik Bohlin, Tomas Hirsch, Ulrika Trolle and Cecilia

Pettersson for giving me the opportunity to write this thesis at Handelsbanken

Capital Markets and for providing me with the data required for the study.

1

Page 8: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts
Page 9: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

Contents

1 Introduction 3

2 Previous research 4

3 Data 5

3.1 Yield curve data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.2 Construction of yield curve shifts . . . . . . . . . . . . . . . . . . . . 6

3.3 Construction of continuous yield curve shifts . . . . . . . . . . . . . 10

4 Methodology 11

4.1 Functional data analysis . . . . . . . . . . . . . . . . . . . . . . . . . 11

4.2 Functional principal component analysis . . . . . . . . . . . . . . . . 14

4.3 Functional autoregression estimation . . . . . . . . . . . . . . . . . . 16

4.4 Model formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.5 Methods for evaluating model performance . . . . . . . . . . . . . . 20

5 Results 21

5.1 Empirical functional principal components . . . . . . . . . . . . . . . 21

5.2 FAR(1) predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

5.3 Simulation model performance . . . . . . . . . . . . . . . . . . . . . 27

6 Conclusion 33

References 34

2

Page 10: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts
Page 11: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

1 Introduction

The value of a bond portfolio depends on yield levels of its included maturities. An

investor holding such a portfolio is therefore exposed to interest rate risk which is

simply the uncertainty regarding the future portfolio value due to stochastic changes

in yields. To fully understand this risk exposure, one needs to evaluate the joint prob-

ability distribution of yield changes over given time intervals. However, determining

this distribution function might be difficult. Doing so requires an accurate under-

standing of how multiple stochastic variables are distributed and how they depend

on each other.

As a result, risk analysts often use simplifications and approximations when analyz-

ing interest rate risk of portfolios. Usually, this is done by considering how much

portfolio values change if the yield curve shifts in parallel by a certain amount. Al-

though such methods might give useful indications of interest rate sensitivity, they

are still unrepresentative of true value fluctuations. The parallel shift assumption

might be sufficient for adjacent maturities or over small time intervals, but does not

realistically describe a general yield curve transformation. Most importantly, the

parallel shift assumption does not give any information about the probability dis-

tribution of the future portfolio value which is what a rational investor cares about

ultimately.

We approach this problem by modeling yield curve shifts as distinct but continuous

stochastic objects defined over a continuum of maturities rather than multivariate

stochastic variables defined for a finite number of maturities. The framework un-

derlying this technique is known as functional data analysis (FDA) which is a novel

field of statistics dating back to the early 1990’s [9]. It however relies on findings

– mostly the theory of Hilbert spaces and functional analysis – which are far older

branches of mathematics.

Based on this framework, we provide two stochastic models which attempt to de-

scribe realistic yield curve shifts. The first of these models assumes that yield curve

3

Page 12: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

shifts are independent across time whereas the second model incorporates temporal

dependence through the introduction of the functional autoregressive (FAR) opera-

tor. Both models are found to replicate the marginal distributions of yield changes

for medium-term and long-term maturities, whereas distributions for the shorter

segment of the yield curve appear less well-captured. Still, both models appear

to resemble much of the cross-sectional dependence inherent in the yield curve, as

measured by sample correlation estimates.

The remainder of this thesis is organized as follows. Section 2 provides an overview

of existing yield curve modeling literature. Thereafter, Section 3 describes the yield

curve data used in the study and the main steps of the construction of continuous

yield curve shifts. Furthermore, Section 4 outlines the mathematical framework

underlying FDA, functional principal component analysis as well as the theory of

FAR operators. Finally, Section 5 describes the main empirical results whereas

Section 6 concludes.

2 Previous research

The literature on yield curve modeling has grown tremendously over the last couple

of decades. Existing proposals can be broadly described in terms of three categories.

A first set of yield curve models is based on modeling the instantaneous short rate

which is the theoretical interest rate prevailing over an infinitesimal time interval

[t, t+dt]. An early work in this field is provided by Vasicek [14] who models the short

rate as a mean-reverting stochastic process and derives a corresponding yield curve

model. Despite analytical tractability of this model, its validity has been questioned

partly due to its inability to reproduce strictly non-negative interest rates. Cox,

Ingersoll and Ross [2] further add to this class of models by modeling the variance of

the stochastic short-rate process as a function of the short rate itself, such that non-

negative interest rates are ensured. Other important contributions in the short-rate

class of models include Ho and Lee [6] and Hull and White [8].

4

Page 13: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

A drawback of these models, however, resides within their inability to model yield

curves such that arbitrage opportunities can be excluded. A second branch of the

literature, generalized by Heath, Jarrow and Morton [5], therefore proposes models

which are consistent with no-arbitrage restrictions. This is done by modeling the

instantaneous forward curve, rather than the instantaneous short rate, such that

arbitrage opportunities can be ruled out.

A third category of yield curve literature aims at modeling yield curve shapes by

fitting suitable basis functions to observed yield curves. An early piece of work

in this class is provided by Nelson and Siegel [11] who propose a parsimonious

three-factor model that tends to describe the yield curve rather efficiently. A four-

factor extension of this model has been proposed by Svensson [13]. Furthermore,

Diebold and Li [3] propose a dynamic factor model based on the Nelson and Siegel

[11] framework. A similar contribution is provided by Hays et al [4]. This model

however uses the theory of FDA which improves the accuracy of the functional form

used to express the yield curve function.

This paper adds to existing yield curve literature by expanding the use of FDA tech-

niques in yield curve modeling. While Hays et al [4] mainly study the predictability

performance of their model, this paper rather focuses on the distribution of yield

curve shifts from a risk management perspective.

3 Data

3.1 Yield curve data

In order to evaluate the applicability of FDA methods (described in detail in Sec-

tion 4) we apply these on empirical observations of German sovereign bond yields.

The data set spans weekly observations between November 18th, 2002 and March

10th, 2014, equivalent to 590 consecutive observations. At each date, we use the

mid values between quoted closing ask and bid yields of 14 different maturities re-

spectively between 3 months and 30 years. All observations are downloaded from

5

Page 14: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

the Bloomberg platform and the full data set is illustrated in Figure 1.

May01 Feb04 Nov06 Aug09 May12 Feb15 0

10

20

30

−1

0

1

2

3

4

5

6

Maturity [years]

Yie

ld [%

]

Figure 1: German sovereign yield curves in the period November 18th, 2002 - March 10th, 2014,corresponding to 590 consecutive observations.

3.2 Construction of yield curve shifts

Let yn(τ0), yn(τ1), ..., yn(τm) denote the set of m+ 1 observed yields with maturities

τ0, τ1, ..., τm at each date n = 1, 2, ..., N , where τ0 < τ1 < ... < τm. Furthermore, let

xn(τi) = yn(τi)− yn−1(τi) (1)

denote the change in the yield with maturity τi between week n−1 and n. Ultimately,

we are interested in understanding the joint behavior of xn(τ0), xn(τ1), ..., xn(τm).

However, the FDA techniques described in Section 4 requires yield curve shifts to

be defined on the interval [0, 1]. To enforce this, each maturity is transformed into

the dimensionless quantity

ti =τi − τ0τm

∈ [0, 1] (2)

6

Page 15: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

and the remainder of the paper will analyze the properties ofXn(t0), Xn(t1), ..., Xn(tm),

where

Xn(ti) := xn(τi). (3)

A summary of the maturities τi included in the data set and their transformed

values ti is provided in Table 1. Furthermore, estimated cross-sectional correlations

of (Xn(t0), Xn(t1), . . . , Xn(tm)) are provided in Table 2 whereas Figure 2 illustrates

histograms of yield changes across all 14 maturities.

Maturity Abbreviation i τi ti3 months 3M 1 0.25 0.006 months 6M 2 0.5 0.011 year 1Y 3 1 0.032 years 2Y 4 2 0.063 years 3Y 5 3 0.094 years 4Y 6 4 0.135 years 5Y 7 5 0.166 years 6Y 8 6 0.197 years 7Y 9 7 0.238 years 8Y 10 8 0.269 years 9Y 11 9 0.2910 years 10Y 12 10 0.3320 years 20Y 13 20 0.6630 years 30Y 14 30 1.00

Table 1: Summary of observed maturities τi along with transformed values ti = τi−τ1τm

∈ [0, 1].

7

Page 16: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

−2

02

050100

150

200

03M

−2

02

020406080100

120

140

06M

−2

02

01020304050607001

Y

−2

02

0510152025303502

Y

−2

02

0510152025303503

Y

−2

02

051015202504

Y

−2

02

051015202505

Y

−2

02

051015202506

Y

−2

02

05101520253007

Y

−2

02

05101520253008

Y

−2

02

051015202509

Y

−2

02

051015202510

Y

−2

02

05101520253020

Y

−2

02

051015202530

Y

Fig

ure

2:H

isto

gra

ms

of

obse

rved

wee

kly

yie

ldch

anges

for

14

diff

eren

tm

atu

riti

esin

the

range

3m

onth

s-30

yea

rs.

8

Page 17: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

3M6M

1Y2Y

3Y

4Y

5Y

6Y

7Y

8Y

9Y

10Y

20Y

30Y

3M1.

000.

660.

430.

22

0.2

00.1

70.1

50.1

30.1

20.1

30.1

10.1

30.1

40.1

36M

0.66

1.00

0.74

0.46

0.4

00.3

70.3

40.3

20.3

00.2

80.2

50.2

60.2

60.2

51Y

0.43

0.74

1.00

0.80

0.7

20.6

80.6

40.6

00.5

60.5

30.4

90.4

80.3

90.3

62Y

0.22

0.46

0.80

1.00

0.9

30.9

10.8

80.8

40.8

10.7

70.7

40.7

20.6

00.5

53Y

0.20

0.40

0.72

0.93

1.0

00.9

80.9

40.8

90.8

60.8

30.8

00.7

90.6

80.6

34Y

0.17

0.37

0.68

0.91

0.9

81.0

00.9

80.9

40.9

30.9

10.8

80.8

70.7

60.7

15Y

0.15

0.34

0.64

0.88

0.9

40.9

81.0

00.9

70.9

60.9

40.9

20.9

10.8

10.7

66Y

0.13

0.32

0.60

0.84

0.8

90.9

40.9

71.0

00.9

90.9

80.9

60.9

50.8

50.8

07Y

0.12

0.30

0.56

0.81

0.8

60.9

30.9

60.9

91.0

00.9

90.9

80.9

70.8

80.8

38Y

0.13

0.28

0.53

0.77

0.8

30.9

10.9

40.9

80.9

91.0

00.9

90.9

80.9

00.8

69Y

0.11

0.25

0.49

0.74

0.8

00.8

80.9

20.9

60.9

80.9

91.0

00.9

90.9

20.8

810

Y0.

130.

260.

480.

72

0.7

90.8

70.9

10.9

50.9

70.9

80.9

91.0

00.9

30.9

020

Y0.

140.

260.

390.

60

0.6

80.7

60.8

10.8

50.8

80.9

00.9

20.9

31.0

00.9

830

Y0.

130.

250.

360.

55

0.6

30.7

10.7

60.8

00.8

30.8

60.8

80.9

00.9

81.0

0

Tab

le2:

Est

imate

dcr

oss

-sec

tional

corr

elati

on

coeffi

cien

tsfo

robse

rved

wee

kly

yie

ldch

anges

.

9

Page 18: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

3.3 Construction of continuous yield curve shifts

As outlined in Section 1, FDA enables the risk analyst to model yield curve shifts

as continuous stochastic functions. Therefore, before applying FDA methods, the

observed (discrete) data needs to be interpolated such that continuous objects are

obtained. This is done by fitting the observations Xn(ti) to natural spline functions

Si,n(t) for i = 0, 1, ...,m − 1, where t (without subscript) refers to the continuous

maturity variable on the interval [0, 1].

For each observation date n and for each sub-interval [ti, ti+1], i = 0, 1, ...,m we fit

the third degree polynomial Si,n(t) = ai,n + bi,n(t − ti) + ci,n(t − ti)2 + di,n(t − ti)3

such that the conditions

Si,n(ti) = Xn(ti), i = 0, 1, ...,m− 1

Si,n(ti+1) = Xn(ti+1), i = 0, 1, ...,m− 2

Si+1,n(ti+1) = Si,n(ti+1), i = 0, 1, ...,m− 2

S′i+1,n(ti+1) = S′i,n(ti+1), i = 0, 1, ...,m− 2

S′′i+1,n(ti+1) = S′′i,n(ti+1), i = 0, 1, ...,m− 2

S′′i,n(t0) = S′′i,n(tm) = 0

(4)

are fulfilled. The interpolated continuous yield curve shifts are then given by

Xn(t) =

S0,n(t), t ∈ [t0, t1]

S1,n(t), t ∈ [t1, t2]

...

Sm−1,n(t), t ∈ [tm−1, tm].

(5)

The first two spline conditions in Eq. (4) ensure that the fitted splines pass through

the observed discrete points Xn(ti). Furthermore, the third condition guarantees

continuity in Xn(t) across the full interval [0, 1], whereas the fourth and fifth con-

ditions generate smooth shapes through enforcement of continuous first and second

10

Page 19: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

derivatives. Finally, the last two conditions are boundary conditions which have

limited effects on the overall shape of the functions.

In order to determine the coefficients ai,n, bi,n, ci,n, di,n for i = 0, 1, ...,m and n =

1, 2, ..., N Burden and Faires [1] propose the following steps:

1. Define αi,n = 3hi

(Xn(ti+1) − Xn(t1)) − 3hi−1

(Xn(ti) − Xn(ti−1)), where hi =

ti+1 − ti.

2. Define the variables li, µi and zi,n such that li = 2(ti+1 − ti−1) − hi−1µi−1,

µi = hi/li and zi,n =αi,n−hizi−1,n

lifor i = 1, 2, ...,m − 1. Furthermore, set

l0 = lm = 1 and z0,n = zm,n = 0 as well as cn,m = 0.

3. The polynomial coefficients are given as ai,n = ti, bi,n =ai+1,n−ai,n

hi−hi ci+1,n+2ci,n

3 ,

ci,n = zi,n − µici+1,n and di,n =ci+1,n−ci,n

3hiwhich are obtained by solving the

corresponding linear system of equations.

Interpolated functions Xn(t) satisfying the spline conditions are illustrated in Figure

3. The remaining analysis will mainly focus on these observations and the methods

for doing this are described in the next section.

4 Methodology

4.1 Functional data analysis

This section provides an overview of the mathematical framework underlying FDA.

The theory described here is based on the book by Horvath and Kokoszka [7] and

it relies on the mathematics of Hilbert spaces. In particular, the analysis will con-

sider the space L2 = L2([0, 1]) which is the set of measurable real-valued functions

u = {u(t), t ∈ [0, 1]} with the square integrable property∫ 10 u

2(t)dt < ∞. L2 is a

separable Hilbert space with the inner product

〈u, v〉 =

∫ 1

0u(t)v(t)dt (6)

11

Page 20: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

0 0.2 0.4 0.6 0.8 1−1.5

−1

−0.5

0

0.5

1

t

Xn(t

)

Figure 3: Continuous yield curve shifts.

and the norm

‖u‖2 = 〈u, u〉 =

∫ 1

0u2(t)dt. (7)

Furthermore, the analysis makes use of operators in L which is the space of bounded

linear operators on L2.

Throughout the remainder of the paper, the continuous yield curve shifts Xn(t),

n = 1, 2, ..., N will be viewed as realizations of the stochastic function X = {X(t), t ∈

[0, 1]} which is a random element in L2, equipped with the Borel σ-algebra. Fur-

thermore, it will be assumed that X is square integrable in the sense that

E‖X‖2 = E[〈X,X〉] = E

[∫ 1

0X2(t)dt

]<∞. (8)

Provided that this assumption holds, there is a unique function in L2

µ(t) = E[X(t)] (9)

12

Page 21: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

which is known as the mean function of X. Furthermore, conditional on the square

integrable Condition (8), the covariance operator of X is defined as

C(u) = E[〈(X − µ) , u〉 (X − µ)] ∈ L (10)

for u ∈ L2. It holds that

C(u) =

∫ 1

0c(t, s)u(s)ds (11)

where

c(t, s) = E[(X(t)− µ(t)) (X(s)− µ(s))] (12)

is known as the covariance function. The sample counterparts of the mean function,

covariance operator and covariance functions are defined respectively as

µ(t) =1

N

N∑j=1

Xj(t), (13)

C(u) =1

N

N∑j=1

〈Xj − µ, u〉 (Xj − µ) (14)

and

c(t, s) =1

N

N∑j=1

(Xj(t)− µ(t)) (Xj(s)− µ(s)) , (15)

where u ∈ L2.

In the remaining analysis, it will be assumed that µ(t) = 0 which is supported by

the economic intuition that yields are drift-less in the sense that they tend to remain

within certain bounds over time. This assumption is verified by checking the sample

mean function µ(t) for the observed data set. Computations show that this function

is close to zero for all t ∈ [0, 1].

Provided that E‖X‖4 < ∞, which we have implicitly assumed by Condition (8),

the covariance operator is a symmetric positive-definite so called Hilbert-Schmidt

13

Page 22: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

operator in L. As such, it can be decomposed as

C(u) =∞∑k=1

λk〈u, vk〉vk (16)

where vk and λk are the eigenfunctions and eigenvalues of C respectively, such that

C(vk) = λkvk. (17)

The eigenfunctions (also known as principal components) form an orthogonal basis

in L2 which allows the yield curve shifts X(t) to be expressed as

X(t) =∞∑k=1

ξkvk (18)

where the ξk = 〈X, vk〉 are known as the principal component scores of X. This

expansion is known as the Karhuen-Loeve expansion and provides the foundation of

functional principal component analysis (FPCA) described next.

4.2 Functional principal component analysis

In the context of modeling yield curve shifts, we are interested in expressing these

functions through an efficient basis expansion in L2. Ideally, we want this expan-

sion to capture as much as possible of the characteristics inherent in the functions

with as few basis functions as possible. For this purpose, let u1, u2, ..., up, denote

the basis functions of an orthonormal basis. We want to choose this basis such

that the observed yield curve shifts can be closely approximated by the expansion∑pk=1〈X,uk〉uk for a given value of p. This problem is equivalent to that of mini-

mizing

S2 =N∑n=1

‖Xi −p∑

k=1

〈Xn, uk〉uk‖2. (19)

Horvath and Kokoszka [7] show that this expression is minimized if u1 = v1, u2 =

v2, ..., up = vp, where v1, v2, ..., vp are the principal components corresponding to the

14

Page 23: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

p largest eigenvalues λ1 > λ2 > ... > λp of the empirical covariance operator C.

The basis v1, v2, ..., vp can hence be viewed as the optimal orthonormal basis with

which we can describe the data. In particular, this result suggests that the Karhuen-

Loeve expansion in Eq. (18) can be well approximated by its finite counterpart

X(t) =

p∑k=1

ξkvk. (20)

For a given choice of p, this basis captures as much as possible of the observed

variance. To see why this is the case, consider the statistic

1

N

N∑n=1

〈Xn, u〉2 = 〈C(u), u〉 (21)

which can be interpreted as the sample variance along the direction of the function

u. Since v1, v2, ..., vN constitutes an orthogonal basis, it holds that

1

N

N∑n=1

‖Xn‖2 =1

N

N∑n=1

N∑q=1

〈Xn, vq〉2 =N∑q=1

1

N

N∑n=1

〈Xn, vq〉2 =N∑q=1

λq. (22)

Hence, the principal component vk explains a fraction of the total empirical variance

equivalent to λk∑Nq=1 λq

. Since λ1 > λ2 > ... > λp, the basis v1, v2, ..., vp therefore

captures the highest possible fraction of the variance for a given choice of p.

In order to compute this basis numerically, we follow the procedure suggested by

Ramsay and Silverman [12]. Let s1, s2, ..., sT denote discretized values of t. Further-

more, let

X =

X1(s1) X1(s2) . . . X1(sT )

X2(s1) X2(s2) . . . X2(sT )...

......

XN (s1) XN (s2) . . . XN (sT )

. (23)

The discretized sample covariance matrix can then be defined as

C =1

NX′X. (24)

15

Page 24: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

Furthermore, let uk and ρk denote the eigenvector and eigenvalues of C such that

Cuk = ρkuk. An approximate discrete form of the empirical functional eigenequation

C(vk) = λkvk (25)

is then given by1

TCvk = λvk (26)

where the discretized eigenfunctions can be calculated as

vk =√Tuk (27)

provided that uk is a normalized eigenvector of C. In particular, vk contains the

elements vk(s1), vk(s2), ..., vk(sT ) which approximately replicate the true functions

vk(t) if the discretization is sufficiently dense.

4.3 Functional autoregression estimation

This section describes the theory of FAR processes and the derivation of an empirical

estimate of the FAR(1) operator as suggested by Horvath and Kokoszka [7]. Let

{Xn,∞ < n <∞} be a sequence of random yield curve shifts in L2. These are said

to follow a FAR(1) model if

Xn(t) = Ψ(Xn−1) + εn(t) (28)

where Ψ ∈ L is the FAR(1) operator and {εn,∞ < n < ∞} is a sequence of mean

zero iid errors in L2 such that E‖εn‖2 <∞.

In order to find an empirical estimator of Ψ, let

C1(u) = E[〈Xn, u〉Xn+1] (29)

denote the lag-1 operator which can be expressed in terms of the covariance operator

16

Page 25: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

as C1 = ΨC. It follows that the FAR(1) operator can be obtained as

Ψ = C1C−1 (30)

provided that the inverse of the covariance operator C−1 exists. The representation

in Eq. (16) suggests that this may be expressed as

C−1(u) =

∞∑k=1

λ−1k 〈u, vk〉vk. (31)

However since ‖C−1(vk)‖ = λ−1k → ∞ as k → ∞ it is unbounded and therefore

badly suited for estimation. It is only defined in the subspace spanned by the first

q principal components such that λ1 ≥ λ2 ≥ ... ≥ λq > λq+1 = 0. Therefore, the

inverse of the empirical covariance operator may be estimated with the first p ≤ q

principal components as

ˆIC(u)p =

p∑k=1

λ−1k 〈u, vk〉vk. (32)

This calculation becomes a trade-off between choosing p small enough to avoid re-

ciprocals of small eigenvalues yet large enough to resemble as much as possible of

the information inherent in the data set.

An empirical estimator of C1 is given by

C1(u) =1

N − 1

N∑n=1

〈Xn, u〉Xn+1. (33)

Along with Eq. (30) and Eq. (32), this suggests that an empirical estimator of Ψ

17

Page 26: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

can be expressed as

Ψp(u) = C1ˆICp(u) = C1

(p∑

k=1

λ−1k 〈u, vk〉vk

)=

=1

N − 1

N−1∑n=1

⟨Xn,

p∑k=1

λ−1k 〈u, vk〉vk

⟩Xn+1 =

=1

N − 1

N−1∑n=1

p∑k=1

λ−1k 〈u, vk〉〈Xn, vk〉Xn+1.

(34)

In addition, Horvath and Kokoszka [7] suggest smoothing the operator by using

the approximation implied by Eq. (20) such that Xn+1 ≈∑p

i=1〈Xn+1, vi〉vi. The

empirical operator then becomes

Ψp(u) =1

N − 1

N−1∑n=1

p∑k=1

p∑i=1

λ−1k 〈u, vk〉〈Xn, vk〉〈Xn+1, vi〉vi. (35)

The estimator given in Eq. (35) is a kernel operator in the sense that

Ψp(u) =

∫ 1

0ψp(t, s)u(s)ds (36)

where the kernel can be expressed as

ψp(t, s) =1

N − 1

N−1∑n=1

p∑k=1

p∑j=1

λ−1k 〈Xn, vk〉〈Xn+1, vi〉vkvi. (37)

The predictor of Xn+1 is now given by Eq. (28), Eq. (35) and Eq. (37) as

Xn+1 = Ψp(Xn) =

∫ 1

0ψp(t, s)Xn(s)ds =

=

p∑k=1

(p∑l=1

ψkl〈Xn, vl〉

)vk(t),

(38)

where

ψji = λ−1i1

N − 1

N−1∑n=1

〈Xn, vi〉〈Xn+1, vj〉. (39)

18

Page 27: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

4.4 Model formulation

In order to evaluate whether FDA techniques can be used for risk management

purposes, we propose two different models and compare them with empirical yield

data. The first model uses the finite approximation of the Karhuen-Loeve expansion

given in Eq. (20) and has the form

Xn(t) =

p∑k=1

σkZnk vk(t) (40)

where σk for k = 1, 2, ..., p are the standard deviations of the principal component

scores ξnk = 〈Xn(t), vk(t)〉 and Znk are standard normally distributed variables such

that Cov(Znk , Znl ) = 0 for k 6= l and Cov(Znk , Z

mk ) = 0 for m 6= n. The model

implicitly assumes that the principal component scores ξnk are normally distributed

with mean zero and standard deviation σk and that yield curve shifts are independent

across time.

The second model uses the theory underlying FAR(1) operators as described in

Section 4.3 and has the form

Xn(t) = Ψp(Xn−1) +

p∑k=1

σ∗kZnk v∗k(t) (41)

where Ψp is the empirical FAR(1) operator given in Eq. (35). The second term

in this expression represents the error term εn(t) in Eq. (28). In particular, v∗k(t)

denotes the principal components of the error term and σ∗k denotes the standard

deviations of the corresponding scores ξn∗k = 〈εn(t), v∗k(t)〉. In contrast to the model

described by Eq. (40), this model does not assume independent yield curve shifts

across time as implied by the FAR(1) operator. The model however assumes that

the errors are temporally independent.

19

Page 28: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

4.5 Methods for evaluating model performance

Let X∗j (ti) denote ordered observations of yield changes X(ti) such that X∗1 (ti) <

X∗2 (ti) < ... < X∗N (ti). Furthermore, let Y ∗j (ti) denote corresponding simulated

(model) yield curve shifts such that Y ∗1 (ti) < Y ∗2 (ti) < ... < Y ∗N (ti). It is assumed

that X∗j (ti) and Y ∗j (ti) are outcomes of the stochastic variables X∗(ti) and Y ∗(ti)

with probability distribution functions FX∗(ti) and FY ∗(ti) respectively. In order

to evaluate whether the artificial distributions FY ∗(t1), FY ∗(t2), ..., FY ∗(tm), appear

similar to the true empirical distributions FX∗(t1), FX∗(t2), ..., FX∗(tm), we use two

different evaluation techniques.

Firstly, we visually inspect the marginal distributions of yield changes for each ma-

turity by illustrating them in quantile-quantile plots (henceforth referred to as qq-

plots). This is done by plotting the points {F−1X∗(ti)

(N−k+1N+1

), F−1Y ∗(ti)

(N−k+1N+1 ) : k =

1, 2, ..., N} in a diagram. If FY ∗(ti) = FX∗(ti), the points in the qq-plots should form

a straight line in the diagram.

Secondly, we complement the qq-plot inspection by conducting the Kolmogorov-

Smirnov [10] test of the null hypothesis H0 : FX∗(ti) = FY ∗(ti) against the alternative

H1 : FX∗(ti) 6= FY ∗(ti). This is done by considering the test statistic

D =

√N

2+ 0.12 +

0.11√N2

d (42)

where d = max(FX∗(ti)(s)− FY ∗(ti)(s)

)is the maximum difference between the

empirical distribution functions

FX∗(ti)(s) =1

N

N∑j=1

I{X∗j (ti)<s}(s) (43)

and

FY ∗(ti)(s) =1

N

N∑j=1

I{Y ∗j (ti)<s}(s) (44)

20

Page 29: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

of X∗ and Y ∗ respectively. Here I{A}(s) =

1 if s ∈ A

0 if s 6∈ Adenotes the indicator

function. Under H0, the probability to observe the value d or larger is given by

p(D) = 2∑∞

k=1(−1)k−1e−2k2D2

. Hence, H0 may be rejected if this probability is

small.

The qq-plot and the Kolmogorov-Smirnov test allow us to evaluate whether any

of the two simulation models capture the marginal distribution of observed yields.

However, a successful simulation model should also capture the joint behavior of

yields with different maturities. In order to analyze this property, we compare

cross-sectional correlation matrices between empirical and simulated samples.

5 Results

5.1 Empirical functional principal components

The first 10 principal component functions vk(t), k = 1, ..., 10 are illustrated in

Figure 4, whereas Table 3 reports estimated eigenvalues λk as well as cumulative

fractions of total variance explained up to each principal component. Results suggest

that 86.9% of the observed variance is explained by the first principal component,

implying that it is possible to capture much of the observed yield curve dynamics

with only one stochastic factor. However, if 10 principal components are used, we

may explain as much as 99.9% of the observed variance as shown in Table 3.

Histograms of the corresponding principal component scores ξnk = 〈Xn, vk(t)〉 are

illustrated in Figure 5 which suggests that the distributions of ξnk appear symmetrical

and centered around zero.

In order to analyze the severity of the estimation error of the finite Karhuen-

Loeve approximation in Eq. (20), we plot the estimation errors en(t) = Xn(t) −∑pk=1 ξ

nk vk(t) for p = {2, 6, 10} in Figure 6. As expected, the basis expansion ap-

pears to better approximate yield curve shifts for large values of p implied by the

21

Page 30: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

00.

10.

20.

30.

40.

50.

60.

70.

80.

91

−10−

5051015

t

vk(t)

k=1

k=2

k=3

k=4

k=5

k=6

k=7

k=8

k=9

k=10

Fig

ure

4:

Funct

ional

pri

nci

pal

com

ponen

ts,

equiv

ale

nt

toth

efirs

t10

eigen

funct

ions

of

the

cova

riance

op

erato

r.

22

Page 31: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

−0.

50

0.5

05101520k=

1

−0.

050

0.05

01020304050k=

6

−0.

50

0.5

0510152025k=

2

−0.

050

0.05

05101520253035k=

7

−0.

50

0.5

0510152025303540k=

3

−0.

050

0.05

05101520253035k=

8

−0.

50

0.5

01020304050k=

4

−0.

050

0.05

01020304050k=

9

−0.

50

0.5

0510152025303540k=

5

−0.

050

0.05

01020304050k=

10

Fig

ure

5:F

unct

ional

pri

nci

pal

com

ponen

tsc

ore

sco

rres

pondin

gto

the

firs

t10

eigen

funct

ions

of

the

cova

riance

op

erato

r.N

ote

the

diff

eren

tsc

aling

of

thex

axes

bet

wee

nth

eupp

erpanel

and

the

low

erpanel

.

23

Page 32: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

k λk Fraction of total variance1 0.009384363 86.9%2 0.000838669 94.7%3 0.000234191 96.9%4 0.000160627 98.4%5 0.000072049 99.0%6 0.000039878 99.4%7 0.000016059 99.5%8 0.000013559 99.7%9 0.000013211 99.8%10 0.000010272 99.9%

Table 3: The first 10 eigenvalues λk and the total cumulative variance explained up to each k.

smaller errors. In addition, it seems that the estimation error is higher for shorter

maturities compared to the medium-term and long-term counterparts, regardless of

the choice of p.

To gain further understanding of the error characteristics we plot means, standard

deviations, maximum values and minimum values of e(t) for different maturities and

different choices of p in Figure 7. Results suggest that the mean error is virtually

zero for all choices of p and for all maturities, implying that the method does not

introduce any systematic estimation bias. Furthermore, error standard deviations

and maximum errors (in absolute terms) decrease with p which confirms that the

approximation improves with p. Figure 7 however, again, suggests that the approxi-

mation error is more severe for short-term maturities implied by the larger standard

deviations and maximum errors for the 3 month rate.

5.2 FAR(1) predictions

Given the results above, the empirical FAR(1) operator Ψp can be calculated using

Eq. (38). Predicted functions Xn and the corresponding prediction errors εn(t)

are plotted in Figure 8 using p = 10 eigenfunctions and the full data set of 590

observations. Visual inspection of Figure 8 suggests that Ψp captures much of the

observed curvature in yield curve shifts around short maturities. In particular, the

resulting errors εn(t) appear more flat compared to the original data shown in Figure

3. Furthermore, results imply that some of the outliers observed in the original data

24

Page 33: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

00.

10.

20.

30.

40.

50.

60.

70.

80.

91

−1012

t

e(t)

p=2

00.

10.

20.

30.

40.

50.

60.

70.

80.

91

−0.

50

0.51

t

e(t)

p=6

00.

10.

20.

30.

40.

50.

60.

70.

80.

91

−0.

2

−0.

10

0.1

0.2

t

e(t)

p=10

Fig

ure

6:E

stim

ati

on

erro

rari

sing

from

repla

cing

the

infinit

eK

arh

uen

-Loev

eex

pansi

on

wit

hit

sfinit

eco

unte

rpart

forp

=2,6,1

0.

Note

the

diff

eren

tsc

ale

son

they-a

xes

.

25

Page 34: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

23

45

67

89

10−

10−8

−6

−4

−2024

x 10

−18

p

Mean error

23

45

67

89

100

0.02

0.04

0.06

0.080.

1

p

Error standard deviation

03

M01

Y05

Y10

Y30

Y

23

45

67

89

100

0.2

0.4

0.6

0.81

1.2

1.4

p

Maximum errors

23

45

67

89

10−

0.5

−0.

4

−0.

3

−0.

2

−0.

10

p

Minimum errorsF

igu

re7:

Mea

n,

standard

dev

iati

on,

maxim

um

valu

esand

min

imum

valu

esfo

rth

eappro

xim

ati

on

erro

rof

sele

cted

yie

lds

ari

sing

from

repla

cing

the

Karh

uen

-Loev

eex

pansi

on

by

its

finit

eco

unte

rpart

.R

esult

sare

plo

tted

for

diff

eren

tva

lues

ofp.

26

Page 35: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

set are captured by Ψp since the dispersion of the errors seems to be somewhat

smaller than the original yield curve shifts.

0 0.2 0.4 0.6 0.8 1−1

−0.5

0

0.5

t

Ψp(X

n(t))

FAR(1) predictions

0 0.2 0.4 0.6 0.8 1−0.5

0

0.5

t

ε n(t)

FAR(1) errors

Figure 8: The upper plot shows FAR(1) predictions Xn(t) = Ψp(Xn−1(t)), whereas the lower plot

shows the corresponding estimation errors εn(t) = Ψp(Xn−1(t))− Xn(t).

5.3 Simulation model performance

In this section, we evaluate the performance of the models given in Eq. (40) and Eq.

(41). Results from simulating 590 yield curve shifts for both models are illustrated

in Figure 9. The σk and σ∗k coefficients in these simulations were set to their corre-

sponding sample estimates. To facilitate comparison, Figure 9 also shows empirical

yield curve observations already shown in Figure 3.

In order to evaluate the in- and out-of-sample performance of each simulation model,

we split the sample into one estimation sample containing the first 295 observations

of the data set and one validation sample containing the remaining 295 observations.

We thereafter extract the first 10 principal components of the estimation sample and

estimate the corresponding parameters σk and σ∗k. Finally, we simulate 295 yield

27

Page 36: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

0 0.2 0.4 0.6 0.8 1−0.5

0

0.5Simulated independent yield curve shifts

t

Xn(t

)

0 0.2 0.4 0.6 0.8 1−0.5

0

0.5

t

Xn(t

)

Simulated FAR(1) yield curve shifts

0 0.2 0.4 0.6 0.8 1−0.5

0

0.5Real yield curve shifts

t

Xn(t

)

Figure 9: Comparison between simulated and observed shifts. The upper plot shows curvessimulated according to Eq. (40) which assumes that shifts are independent across time. Themiddle plot shows curves simulated according to Eq. (41) which assumes that yield curve shiftsobey the FAR(1) process with independent errors. The lower plot shows real yield curve shifts.

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

t

p−va

lue

Kolmogorov−Smirnov test (in−sample)

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

t

p−va

lue

Kolmogorov−Smirnov test (out−of−sample)

Independent modelFAR(1) model

Figure 10: Significance levels of the Kolmogorov-Smirnov test.

28

Page 37: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

curve shifts of each model and compare these with the estimation sample (in-sample

comparison) and the validation sample (out-of-sample comparison).

Figure 11 shows in-sample qq-plots for a selection of maturities for both simulation

models. Visual inspection of these plots suggests that distributions of simulated

yield shifts appear similar to the empirical distributions for medium-term and long-

term maturities. In particular, the FAR(1) model performs slightly better than the

independent yield curve shift for the 5-year, 10-year and 30-year maturities. None

of the models, however, perform satisfactory for the two shortest maturities.

Figure 12 shows the corresponding out-of-sample qq-plots. While both models ap-

pear to more or less replicate the distributions of 10-year yields and 30-year yields,

they perform somewhat worse for 5-year yields compared to their in-sample per-

formances. In addition, the poor performance for short maturities remains in the

out-of-sample comparison.

Significance levels obtained from Kolmogorov-Smirnov tests are plotted for t =

0, 0.01, ..., 1 in Figure 10 which confirms the poor distributional fits between simu-

lated and empirical data at the short end of the yield curve. The results show small

p values for short-term maturities both in- and out-of-sample, suggesting that the

null hypothesis of equal distributions can be rejected at this segment of the yield

curve.

Finally, Table 4 displays the differences in estimated sample cross-sectional correla-

tions between the empirical sample and each of the two simulation samples. Small

values of these differences suggest that the simulation samples appear to replicate

the cross-sectional dependence of yield changes. Results show that this error ranges

in the interval [−0.047, 0.121] for the independent yield curve model, whereas the

corresponding interval for the FAR(1) model is [−0.009, 0.0078]. Hence, the FAR(1)

model appears slightly better at capturing cross-sectional correlations, although both

models can be said to perform satisfactory in this respect.

29

Page 38: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

−2

−1

01

−6

−5

−4

−3

−2

−1012

Independent model

03M

−2

−1

01

−5

−4

−3

−2

−101

FAR(1) model

03M

−1

−0.

50

0.5

−1.

5

−1

−0.

50

0.51

01Y

−1

−0.

50

0.5

−1.

5

−1

−0.

50

0.51

01Y

−0.

50

0.5

−0.

50

0.5

05Y

−0.

50

0.5

−0.

50

0.5

05Y

−0.

50

0.5

−0.

4

−0.

3

−0.

2

−0.

10

0.1

0.2

0.3

0.4

10Y

−0.

50

0.5

−0.

4

−0.

3

−0.

2

−0.

10

0.1

0.2

0.3

0.4

10Y

−0.

50

0.5

−0.

50

0.5

30Y

−0.

50

0.5

−0.

6

−0.

4

−0.

20

0.2

0.4

30Y

Fig

ure

11:

In-s

am

ple

qq-p

lots

bet

wee

nth

edis

trib

uti

ons

of

sim

ula

ted

yie

ldch

anges

and

true

yie

ldch

anges

of

sele

cted

matu

riti

es.

The

upp

erpanel

corr

esp

onds

tosi

mula

ted

yie

ldch

anges

acc

ord

ing

toE

q.

(40)

wher

eas

the

low

erpanel

corr

esp

onds

tosi

mula

tions

of

Eq.

(41).

30

Page 39: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

−0.

50

0.5

−1.

5

−1

−0.

50

0.51

Independent model

03M

−0.

50

0.5

−1.

5

−1

−0.

50

0.51

FAR(1) model

03M

−0.

50

0.5

−1.

5

−1

−0.

50

0.51

01Y

−0.

50

0.5

−1.

5

−1

−0.

50

0.51

01Y

−0.

50

0.5

−0.

8

−0.

6

−0.

4

−0.

20

0.2

0.4

0.6

05Y

−0.

50

0.5

−0.

8

−0.

6

−0.

4

−0.

20

0.2

0.4

0.6

05Y

−0.

50

0.5

−0.

4

−0.

20

0.2

0.4

0.6

10Y

−0.

50

0.5

−0.

4

−0.

3

−0.

2

−0.

10

0.1

0.2

0.3

0.4

10Y

−0.

50

0.5

−0.

4

−0.

3

−0.

2

−0.

10

0.1

0.2

0.3

0.4

30Y

−0.

50

0.5

−0.

4

−0.

3

−0.

2

−0.

10

0.1

0.2

0.3

0.4

30Y

Fig

ure

12:

Out-

of-

sam

ple

qq-p

lots

bet

wee

nth

edis

trib

uti

ons

of

sim

ula

ted

yie

ldch

anges

and

true

yie

ldch

anges

of

sele

cted

matu

riti

es.

The

upp

erpanel

corr

esp

onds

tosi

mula

ted

yie

ldch

anges

acc

ord

ing

toE

q.

(40)

wher

eas

the

low

erpanel

corr

esp

onds

tosi

mula

tions

of

Eq.

(41).

31

Page 40: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

3M6M

1Y2Y

3Y

4Y

5Y

6Y

7Y

8Y

9Y

10Y

20Y

30Y

Sim

ula

ted

ind

epen

den

tyie

ldcu

rve

shif

ts3M

0.00

00.

057

0.07

50.

121

0.0

87

0.0

85

0.0

86

0.0

74

0.0

63

0.0

49

0.0

35

0.0

28

0.0

06

-0.0

14

6M0.

057

0.00

00.

009

0.05

10.0

26

0.0

23

0.0

24

0.0

40

0.0

31

0.0

20

0.0

06

0.0

04

-0.0

07

-0.0

17

1Y0.

075

0.00

90.

000

0.01

90.0

07

0.0

00

-0.0

07

0.0

01

-0.0

07

-0.0

17

-0.0

30

-0.0

31

-0.0

39

-0.0

47

2Y0.

121

0.05

10.

019

0.00

00.0

03

-0.0

04

-0.0

06

-0.0

05

-0.0

09

-0.0

12

-0.0

17

-0.0

21

-0.0

27

-0.0

27

3Y0.

087

0.02

60.

007

0.00

30.0

00

0.0

07

0.0

01

0.0

02

0.0

01

0.0

01

-0.0

01

-0.0

03

-0.0

01

-0.0

01

4Y0.

085

0.02

30.

000

-0.0

040.0

07

0.0

00

0.0

08

0.0

03

0.0

02

0.0

02

0.0

00

-0.0

01

0.0

00

0.0

03

5Y0.

086

0.02

4-0

.007

-0.0

060.0

01

0.0

08

0.0

00

0.0

01

-0.0

01

-0.0

02

-0.0

05

-0.0

05

-0.0

07

-0.0

01

6Y0.

074

0.04

00.

001

-0.0

050.0

02

0.0

03

0.0

01

0.0

00

0.0

01

0.0

02

0.0

00

-0.0

01

-0.0

01

-0.0

04

7Y0.

063

0.03

1-0

.007

-0.0

090.0

01

0.0

02

-0.0

01

0.0

01

0.0

00

0.0

02

0.0

01

0.0

00

0.0

00

-0.0

02

8Y0.

049

0.02

0-0

.017

-0.0

120.0

01

0.0

02

-0.0

02

0.0

02

0.0

02

0.0

00

0.0

02

0.0

02

0.0

00

0.0

01

9Y0.

035

0.00

6-0

.030

-0.0

17-0

.001

0.0

00

-0.0

05

0.0

00

0.0

01

0.0

02

0.0

00

0.0

00

-0.0

03

0.0

03

10Y

0.02

80.

004

-0.0

31-0

.021

-0.0

03

-0.0

01

-0.0

05

-0.0

01

0.0

00

0.0

02

0.0

00

0.0

00

0.0

01

0.0

07

20Y

0.00

6-0

.007

-0.0

39-0

.027

-0.0

01

0.0

00

-0.0

07

-0.0

01

0.0

00

0.0

00

-0.0

03

0.0

01

0.0

00

0.0

07

30Y

-0.0

14-0

.017

-0.0

47-0

.027

-0.0

01

0.0

03

-0.0

01

-0.0

04

-0.0

02

0.0

01

0.0

03

0.0

07

0.0

07

0.0

00

Sim

ula

ted

FA

R(1

)yie

ldcu

rve

shif

ts3M

0.00

00.

009

0.00

90.

025

0.0

05

0.0

18

0.0

23

0.0

18

0.0

23

0.0

32

0.0

42

0.0

38

0.0

61

0.0

53

6M0.

009

0.00

00.

004

0.02

00.0

31

0.0

38

0.0

36

0.0

43

0.0

46

0.0

48

0.0

54

0.0

44

0.0

68

0.0

74

1Y0.

009

0.00

40.

000

0.01

70.0

43

0.0

55

0.0

51

0.0

51

0.0

56

0.0

61

0.0

69

0.0

68

0.0

77

0.0

78

2Y0.

025

0.02

00.

017

0.00

00.0

13

0.0

08

0.0

06

0.0

05

0.0

07

0.0

09

0.0

10

0.0

10

0.0

03

0.0

05

3Y0.

005

0.03

10.

043

0.01

30.0

00

0.0

06

-0.0

01

-0.0

03

-0.0

02

-0.0

03

-0.0

03

-0.0

03

-0.0

09

-0.0

09

4Y0.

018

0.03

80.

055

0.00

80.0

06

0.0

00

0.0

07

0.0

00

0.0

00

0.0

02

0.0

02

0.0

02

-0.0

04

-0.0

02

5Y0.

023

0.03

60.

051

0.00

6-0

.001

0.0

07

0.0

00

0.0

01

0.0

01

0.0

03

0.0

03

0.0

03

-0.0

02

0.0

00

6Y0.

018

0.04

30.

051

0.00

5-0

.003

0.0

00

0.0

01

0.0

00

0.0

02

0.0

03

0.0

02

0.0

02

0.0

00

0.0

00

7Y0.

023

0.04

60.

056

0.00

7-0

.002

0.0

00

0.0

01

0.0

02

0.0

00

0.0

02

0.0

00

0.0

01

-0.0

02

-0.0

01

8Y0.

032

0.04

80.

061

0.00

9-0

.003

0.0

02

0.0

03

0.0

03

0.0

02

0.0

00

0.0

01

0.0

02

-0.0

02

-0.0

01

9Y0.

042

0.05

40.

069

0.01

0-0

.003

0.0

02

0.0

03

0.0

02

0.0

00

0.0

01

0.0

00

0.0

00

-0.0

04

-0.0

01

10Y

0.03

80.

044

0.06

80.

010

-0.0

03

0.0

02

0.0

03

0.0

02

0.0

01

0.0

02

0.0

00

0.0

00

-0.0

05

-0.0

04

20Y

0.06

10.

068

0.07

70.

003

-0.0

09

-0.0

04

-0.0

02

0.0

00

-0.0

02

-0.0

02

-0.0

04

-0.0

05

0.0

00

0.0

09

30Y

0.05

30.

074

0.07

80.

005

-0.0

09

-0.0

02

0.0

00

0.0

00

-0.0

01

-0.0

01

-0.0

01

-0.0

04

0.0

09

0.0

00

Tab

le4:

Diff

eren

ceb

etw

een

esti

mate

dco

rrel

ati

on

coeffi

cien

tsfo

rth

etr

ue

sam

ple

and

the

two

sim

ula

ted

sam

ple

s.

32

Page 41: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

6 Conclusion

This thesis provides a framework which can be used for yield curve risk manage-

ment purposes. By modeling yield curve shifts as distinct but continuous stochastic

objects defined over a continuum of maturities, we describe how the theory of FDA,

FPCA and FAR operators can be applied in this setting. Based on these tech-

niques, we provide two stochastic models which attempt to describe realistic yield

curve shifts based on observed historical data.

The first of these models assumes that yield curve shifts are independent across time,

whereas the second model incorporates temporal dependence through the introduc-

tion of the FAR(1) operator. We find that both models more or less replicate the

marginal distributions of yield changes for medium-term and long-term maturities,

although the FAR(1) model appears to perform slightly better for this segment of the

yield curve. None of the models however performs satisfactory at the short-maturity

segment of the yield curve.

This failure may be generated by the approximation error arising from replacing the

infinite Karhuen-Loeve expansion by its finite counterpart. Indeed, this is supported

by the in-sample assessment of the approximation error (described in Section 5.1)

which suggests that the short end of the yield curve is more sensitive to small values

of p. Non-normality of short-maturity yield changes may be another likely reason

for the failure to capture these dynamics. It remains a task for future research to

determine the exact marginal distribution of principal component scores and yield

curve shifts.

Both of the proposed models appear to capture much of the cross-sectional depen-

dence observed across the yield curve as measured by sample correlations. This

measure likely provides a general idea of how yields co-move across the curve, al-

though it only provides information about linear dependence. Hence, future research

may improve the analysis by considering more general measures of dependence.

33

Page 42: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

References

[1] Burden, R.L ; Faires, J. D. 2005 Numerical Analysis (8th ed), ThomsonBrooks/Cole

[2] Cox J. C.; Ingersoll J. E. ; Ross S. A. 1985 A Theory of the Term Structureof Interest Rates, Econometrica 53 385-407

[3] Diebold, F.X.; Li, C. 2006, Forecasting the term structure of government bondyields, Journal of Econometrics 130, 337-364

[4] Hays, S.; Haipeng S.; Huang, Z.J. 2012 Functional dynamic factor modelswith application to yield curve forecasting, The Annals of Applied Statistics6, 870-894

[5] Heath, D.; Jarrow R.; Morton, A. 1992 Bond Pricing and the Term Struc-ture of Interest Rates: A New Methodology for Contingent Claims Valuation,Econometrica 60, 77-105

[6] Ho, T.; Lee, S. 1986, Term Structure Movements and Pricing Interest RateContingent Claims, The Journal of Finance 41, 1011-1029

[7] Horvath, L.; Kokoszka, P. 2012, Inference for functional data with applications,Springer Series in Statistics

[8] Hull, J.; White, A. 1990, Pricing Interest-Rate-Derivative Securities, The Re-view of Financial Studies 3, 573-592

[9] Kokoszka, P. 2012, Review article: Dependent Functional Data, ISRN Proba-bility and Statistics 201

[10] Lovric, M. 2011 International Encyclopedia of Statistical Science, Springer

[11] Nelson, C.; Siegel, A. 1987, Parsimonious Modeling of Yield Curves, Journalof Business 60, 473-489

[12] Ramsay, J.O; Silverman, B.W. 2005, Functional Data Analysis (2nd ed),Springer Series in Statistics

[13] Svensson, L. E. O. 1994, Estimating and Interpreting Forward Interest Rates:Sweden 1992-1994, NBER Working Paper Series

[14] Vasicek, O. 1977, An equilibrium characterization of the term structure, Jour-nal of Financial Economics 5, 177-188

34

Page 43: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts
Page 44: Stochastic modeling of yield curve shifts using functional ...kth.diva-portal.org/smash/get/diva2:729542/FULLTEXT01.pdf · We approach this problem by modeling yield curve shifts

TRITA-MAT-E 2014:38 ISRN-KTH/MAT/E—14/38-SE

www.kth.se