Series temporelles
WCDM04, Torino, 9-10 January 2004
Roberto Ren` o reno@unisi.it Dipartimento di Economia Politica, Universit` di Siena a
January 2004 – p.1/18
Introduction
Measuring linear correlation among assets is crucial in financial applications (e.g. Markovitz theory). The last decade witnessed the advent of high-frequancy data. High frequency data are very peculiar: Unevenly sampled Microstructure effects Huge quantity of data
New econometric techniques are needed!
January 2004 – p.2/18
Measuring linear correlation d We assume that p t is the solution of the following stochastic differential equation (SDE):
¡
¢
£
¤
dp t p0
µ t dt x
σ t dW t
¡
¢
¡
¢
¡ ¡
¢
Standard techniques need interpolation of the data to get an evenly spaced grid:
Sτn X Y
t
m 1
∑
©
¥
¢
¥
¦
©
¡ t ¢
Yτn m
Xτn m
1
t
Xτn m
t
Yτn m
§
1
¨
t ¡ ¥
¢
January 2004 – p.3/18
The Fourier method we adopt instead an estimator based on Fourier analysis. We define the Fourier coefficients of the i-th component d pi in the usual way:
ai0
dp
1 2π
2π
¡
¢
d pi t
0 2π
¥
¡
¢
aik d p
1 π
cos kt d pi t
0
bik d p
1 π
2π
¡
¢
¡
¢
§ ¢
¡
¡
¢
¡ sin kt d pi t
0
¢ bk Σi j sin kt
and similar formulas hold for ak Σi j bk Σi j ; from the Fourier coefficients of Σi j , Σi j t can be obtained pointwise by the Fourier-Fejer inversion formula:
§ ¢
¡
¡
Σi j t
n
lim
∞
k 0
∑
n
¢
1
k n
ak Σi j cos kt
¡
¢
¡
¢
¡
¡
¢
¢
¡
¢
!¢
¥
¦
¡
§ ¢
¥
¥
¡
January 2004 – p.4/18
Computing the coefficients
Theorem
N π 1 i j ak d p a k d p 1 n0 k∑0 2 n
"
a0 Σi j