Scribd
Upload
173 views

Chapter 2: Estimating The Term Structure: 2.4 Principal Component Analysis

The document discusses principal component analysis (PCA) as a technique to reduce the dimensionality of multivariate data like yield curves. PCA works by finding orthogonal principal components that successively maximize the explained variance in the data. The first few principal components often explain most of the variance and can be used to approximate the original data. As an example, the first three principal components of monthly Swiss yield curve changes explain over 98% of the variance.

Uploaded by

Nghiem Xuan Hoa
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
173 views

Chapter 2: Estimating The Term Structure: 2.4 Principal Component Analysis

The document discusses principal component analysis (PCA) as a technique to reduce the dimensionality of multivariate data like yield curves. PCA works by finding orthogonal principal components that successively maximize the explained variance in the data. The first few principal components often explain most of the variance and can be used to approximate the original data. As an example, the first three principal components of monthly Swiss yield curve changes explain over 98% of the variance.

Uploaded by

Nghiem Xuan Hoa
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 15

Chapter 2: Estimating the Term Structure

2.4 Principal Component Analysis

Interest Rate Models


Damir Filipovic
2.4 Principal Component Analysis

Major problem in term structure


estimation: high dimensionality
Aim: find basis shapes of the yield
curve (increments)
Principal component analysis (PCA):
dimension reduction technique in
multivariate data analysis

Interest Rate Models


Spectral Theorem

Key mathematical principle underlying PCA:

Spectral theorem: Any real symmetric n n matrix Q can be decomposed

Q = ALA>

where
L is the diagonal matrix of eigenvalues 1 2 n of Q;
A is an orthogonal matrix whose columns a1 , . . . , an are the normalized
eigenvectors of Q (Qai = i ai ), which form an orthonormal basis of Rn .

Interest Rate Models


Spectral Decomposition of Covariance Matrix

Let X be an n-dimensional random X2


vector with
a2
mean = E[X ]
covariance matrix Q = cov[X ] a1
(symmetric and positive semi-definite)
Spectral theorem: Q = ALA> with
eigenvalues 1 n 0 0 X1
eigenvectors a1 , . . . , an

Interest Rate Models


Principal Components and Loadings of X

The principal component transform of X


is given by recentering and rotation X2
>
Y = A (X )
a2
where
a1
Yi = ai> (X ) is the ith principal

component of X
ai is the ith vector of loadings of X
0 X1
Principal component decomposition

X = + AY = + ni=1 Yi ai
P

Interest Rate Models


Properties of the Principal Components

The principal components Y are uncorrelated and have variances Var[Yi ] = i :

E[Y ] = 0
Cov[Y ] = A> QA = A> ALA> A = L

Y1 has maximal variance among all standardized linear combinations of X :

Var[a1> X ] = max Var[b > X ] | b > b = 1




For i = 2, . . . , n: Yi has maximal variance among all such linear combinations


orthogonal to first i 1 linear combinations

Interest Rate Models


Explained Variance

Observe that
n
X n
X n
X
Var[Xi ] = trace(Q) = i = Var[Yi ]
i=1 i=1 i=1

Hence the amount of variability in X explained by the first k principal


components Y1 , . . . , Yk is given by
Pk

Pni=1 i .
i=1 i

Interest Rate Models


Application

Let X be n-dimensional stationary model for (daily changes of) the yield curve.

If the first k  n principal components Y1 , . . . , Yk explain a significant amount


of variability in X then approximate
k
X
X + Yi ai .
i=1

Loadings a1 , . . . , ak are main components of stochastic yield curve movements.

Interest Rate Models


Sample Principal Components

Assume multivariate data observations x = [x(1), . . . , x(N)] where each column


x(t) = (x1 (t), . . . , xn (t))> is a sample realization of a random vector X (t).

Assume X (t) is identically distributed as X with


mean = E[X ]
covariance matrix Q = cov[X ]

Interest Rate Models


Empirical Mean and Covariance

Estimate by the empirical mean


1
PN
= N t=1 x(t)

and Q by the empirical n n covariance matrix (positive semi-definite)

Qij = Cov[xi , xj ] = N1 Nt=1 (xi (t) i )(xj (t) j ).


P

Spectral theorem: Q = ALA> with


eigenvalues 1 n 0
eigenvectors a1 , . . . , an

Interest Rate Models


Sample Principal Components

The sample principal component decomposition of x is given by

x = + ni=1 yi ai
P

with
empirical principal components y = A> (x )
loadings A = [a1 | | an ].

The empirical principal components are uncorrelated:


N
(
1 X i , if i = j,
Cov[yi , yj ] = yi (t)yj (t) =
N t=1 0, else

Interest Rate Models


Discussion

The empirical mean and covariance matrix Q are standard estimators for the
true parameters and Q, if observations X (t) are serially uncorrelated,

Cov[X (t), X (t + h)] = 0 for all h 6= 0.

If this kind of stationarity of time series X (t) is in doubt, the standard practice is
to differentiate and to consider the increments

X (t) = X (t) X (t 1).

Interest Rate Models


PCA of the Yield Curve

Example: monthly changes (t = 1/12) of the yield curve

xi (t) = y (t + t; t + t + i ) y (t, t + i )

of Swiss government bonds from August 2005 until July 2015 for n = 8 times to
maturities i = 2y, 3y, 4y, 5y, 7y, 10y, 20y, 30y.

Interest Rate Models


Yield Curve Loadings

First three yield curve loadings:


Level Slope Curvature
0.6
1st loading is roughly flat: parallel
0.4
shifts of the yield curve (level).
0.2
2nd loading is upward sloping: tilting

Loading
0
of the yield curve (slope).
-0.2
3rd loading is hump-shaped: flexing
-0.4
of the yield curve (curvature).
-0.6
0 5 10 15 20 25 30
Time to maturity

Interest Rate Models


Explained Variance of Principal Components

Explained variance of the principal components:


PC Explained variance (%)
1 76.97
2 18.39
3 3.25
sum 98.61

The first three principal components explain more than 98% of the variance.
Consequence: yield curve (movements) can be approximated by linear
combination of first three loadings, with small relative error.

Interest Rate Models

You might also like