Estimation in Financial Models - RiskLab
Estimation in Financial Models - RiskLab
Estimation in Financial Models - RiskLab
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
When the transition densities of X are unknown, the usual alternative estimator<br />
is dened by approximat<strong>in</strong>g the log-likelihood function for based on<br />
cont<strong>in</strong>uous observations of X. For this log-likelihood function to be dened,<br />
the diusion coecient (t; x; ) =(t; x) has to be known (see Section 3.1.1,<br />
p.17). Under some assumptions (see [51] x7) the log-likelihood function for <br />
based on cont<strong>in</strong>uous observations of X <strong>in</strong> [0;t n ] can be written <strong>in</strong> terms of<br />
<strong>in</strong>tegrals<br />
l c t n<br />
() =<br />
Z tn<br />
0<br />
, 1 2<br />
b (s; X s ; ) T (s; X s )(s; X s ) T ,1<br />
dXs<br />
Z tn<br />
0<br />
b (s; X s ; ) T (s; X s )(s; X s ) T ,1<br />
b (s; Xs ; ) ds; (3.16)<br />
and the usual approximation of these <strong>in</strong>tegrals leads to the approximate loglikelihood<br />
function for based on discrete observations of X<br />
~ ln () =<br />
nX<br />
i=1<br />
, 1 2<br />
with the notation<br />
b (t i,1 ;X ti,1 ; ) T i,1 (X ti , X ti,1 )<br />
nX<br />
i=1<br />
b (t i,1 ;X ti,1 ; ) T i,1 b (t i,1 ;X ti,1 ; )(t i , t i,1 ); (3.17)<br />
i,1 (t i,1 ;X ti,1 ) (t i,1 ;X ti,1 ) T ,1<br />
:<br />
When the diusion coecient also depends on an unknown parameter, the<br />
question arises how the parameter is to be estimated. If divides <strong>in</strong>to two<br />
parts =( 1 ; 2 ), such that b(; ; ) =b(; ; 1 ) and (; ; ) is known up to<br />
the scalar factor 2 , that means (; ; ) = 2 ~(; ), we may avoid the problem<br />
of parameter dependence. In this case 2 2<br />
can be estimated <strong>in</strong> advance by<br />
a quadratic variance-like formula (see Florens-Zmirou [25]), and by <strong>in</strong>sert<strong>in</strong>g<br />
this estimate <strong>in</strong> (; ; ) = 2 ~(; ), the diusion term can be assumed<br />
"known". Then the estimate of the approximate log-likelihood function ~ l n<br />
can be used to estimate 1 .<br />
In the case where the diusion coecient (; ; ) depends on more generally,<br />
Hutton and Nelson [37] show that the discretized score function correspond<strong>in</strong>g<br />
to lt c n<br />
can under certa<strong>in</strong> regularity conditions still be used to estimate<br />
. That means <strong>in</strong> this case we are able to estimate based on discrete<br />
observations of X as well.<br />
However, estimation methods for discrete observations that arise from the<br />
theory of cont<strong>in</strong>uous observations have the undesirable property that the<br />
estimators are strongly biased unless max 1<strong>in</strong> jt i , t i,1 j is "small". If the<br />
23