recent developments in high frequency financial ... - Index of
recent developments in high frequency financial ... - Index of
recent developments in high frequency financial ... - Index of
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Dynamic modell<strong>in</strong>g <strong>of</strong> large-dimensional covariance matrices 297<br />
The optimal shr<strong>in</strong>kage constant is def<strong>in</strong>ed as the value <strong>of</strong> α which m<strong>in</strong>imizes the<br />
expected value <strong>of</strong> the loss function <strong>in</strong> expression (5):<br />
α ∗ t<br />
= argm<strong>in</strong> E [L(αt)] . (6)<br />
αt<br />
For an arbitrary shr<strong>in</strong>kage target F and a consistent covariance estimator S, Ledoit<br />
and Wolf (2003) show that<br />
α ∗ =<br />
�Ni=1 �Nj=1 � � � � ��<br />
Var sij − Cov fij,sij<br />
�Ni=1 �Nj=1 � � �<br />
Var fij − sij + (φij − σij) 2�,<br />
where fij is a typical element <strong>of</strong> the sample shr<strong>in</strong>kage target, sij – <strong>of</strong> the covariance<br />
estimator, σij – <strong>of</strong> the true covariance matrix, and φij – <strong>of</strong> the population shr<strong>in</strong>kage<br />
target �. Further they prove that this optimal value is asymptotically constant over<br />
T and can be written as 3<br />
(7)<br />
κt = πt − ρt<br />
. (8)<br />
νt<br />
In the formula above, πt is the sum <strong>of</strong> the asymptotic variances <strong>of</strong> the entries <strong>of</strong><br />
the sample covariance matrix scaled by √ T : πt = �N � �√Tsij,t �<br />
Nj=1<br />
i=1 AVar ,<br />
ρt is the sum <strong>of</strong> asymptotic covariances <strong>of</strong> the elements <strong>of</strong> the shr<strong>in</strong>kage target<br />
with the elements <strong>of</strong> the sample covariance matrix scaled by √ T : ρt =<br />
�Ni=1 � �√Tfij,t, Nj=1<br />
ACov<br />
√ �<br />
Tsij,t , and νt measures the misspecification <strong>of</strong><br />
� Nj=1 (φij,t −σij,t) 2 . Follow<strong>in</strong>g their formulation<br />
the shr<strong>in</strong>kage target: νt = �N i=1<br />
and assumptions, �N � �√T(fij �<br />
Nj=1<br />
i=1 Var − sij) converges to a positive limit,<br />
and so �N �Nj=1 i=1 Var � � √<br />
fij − sij = O(1/T). Us<strong>in</strong>g this result and the T convergence<br />
<strong>in</strong> distribution <strong>of</strong> the elements <strong>of</strong> the sample covariance matrix, Ledoit<br />
and Wolf (2003) show that the optimal shr<strong>in</strong>kage constant is given by<br />
α ∗ �<br />
1 πt − ρt 1<br />
t = + O<br />
T νt T 2<br />
�<br />
. (9)<br />
S<strong>in</strong>ceα ∗ is unobservable, it has to be estimated. Ledoit and Wolf (2004) propose<br />
a consistent estimator <strong>of</strong> α ∗ for the case where the shr<strong>in</strong>kage target is a matrix <strong>in</strong><br />
which all pairwise correlations are equal to the same constant. This constant is<br />
the average value <strong>of</strong> all pairwise correlations from the sample correlation matrix.<br />
The covariance matrix result<strong>in</strong>g from comb<strong>in</strong><strong>in</strong>g this correlation matrix with the<br />
sample variances, known as the equicorrelated matrix, is the shr<strong>in</strong>kage target. The<br />
equicorrelated matrix is a sensible shr<strong>in</strong>kage target as it <strong>in</strong>volves only a small<br />
number <strong>of</strong> free parameters (hence less estimation noise). Thus the elements <strong>of</strong> the<br />
sample covariance matrix, which <strong>in</strong>corporate a lot <strong>of</strong> estimation error and hence<br />
can take rather extreme values are ‘shrunk’ towards a much less noisy average.<br />
3 In their paper the formula appears without the subscriptt. By add<strong>in</strong>g it here we want to emphasize<br />
that these variables are chang<strong>in</strong>g over time.