02.12.2012 Views

Applications of state space models in finance

Applications of state space models in finance

Applications of state space models in finance

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.4 Maximum likelihood estimation 31<br />

score vector with respect to problems formulated <strong>in</strong> <strong>state</strong> <strong>space</strong>, the reader is referred to<br />

Durb<strong>in</strong> and Koopman (2001, ¢ 7.3.3). In order to avoid numerical or analytical computation,<br />

the Hessian can usually be approximated. In Ox, H(ψ) is approximated us<strong>in</strong>g the<br />

quasi-Newton method accord<strong>in</strong>g to Broyden-Fletcher-Goldfarb-Shanno (BFGS), which<br />

ensures negative def<strong>in</strong>iteness <strong>of</strong> the Hessian; details on this nonl<strong>in</strong>ear optimization techniques<br />

are discussed by Greene (2003, ¢ E.6.2).<br />

3.4.3 The EM algorithm<br />

The unknown elements <strong>of</strong> the system matrices can alternatively be estimated via the<br />

expectation-maximization (EM) algorithm by Dempster et al. (1977). The EM algorithm,<br />

named after its two steps <strong>of</strong> maximiz<strong>in</strong>g the expectation <strong>of</strong> the loglikelihood, is<br />

an iterative algorithm for maximum likelihood estimation. It has orig<strong>in</strong>ally been developed<br />

for deal<strong>in</strong>g with miss<strong>in</strong>g observations and can be employed for maximiz<strong>in</strong>g the<br />

loglikelihood <strong>of</strong> many <strong>state</strong> <strong>space</strong> <strong>models</strong> (cf. Shumway and St<strong>of</strong>fer 1982; Watson and<br />

Engle 1983).<br />

As <strong>in</strong> this thesis the likelihood will be generally maximized us<strong>in</strong>g numerical maximization<br />

procedures, both <strong>in</strong> the context <strong>of</strong> Kalman filter based <strong>state</strong> <strong>space</strong> <strong>models</strong> as<br />

well as the basic hidden Markov model to be <strong>in</strong>troduced below, the reader is referred<br />

to the well-established literature for details on the EM algorithm. For an <strong>in</strong>troductory<br />

outl<strong>in</strong>e, see, for example, Bulla (2006, Appendix A) and the references provided there<strong>in</strong>.<br />

A general demonstration <strong>of</strong> how to use the EM algorithm to compute the ML estimates<br />

<strong>in</strong> some common statistical applications is provided by McLachlan and Krishnan (1997);<br />

a Bayesian treatment is given by Tanner (1993).<br />

3.4.4 Parameter restrictions<br />

Sometimes the parameters <strong>of</strong> a model are not allowed to take any value <strong>in</strong> R, the set <strong>of</strong><br />

real numbers. In this case, it may become necessary to impose parameter constra<strong>in</strong>ts.<br />

For example, the elements <strong>of</strong> the covariance matrices Q t and Ht <strong>in</strong> (3.1) and (3.2)<br />

are restricted to positive values. While it is theoretically possible to <strong>in</strong>troduce such constra<strong>in</strong>ts<br />

<strong>in</strong>to the numerical maximization procedure directly, it is not practically feasible.<br />

In order to employ Newton-type maximization rout<strong>in</strong>es implemented <strong>in</strong> standard statistical<br />

s<strong>of</strong>tware packages, it is recommendable to perform any maximization with respect<br />

to unconstra<strong>in</strong>ed quantities. This can be achieved by transform<strong>in</strong>g the parameters <strong>in</strong> an<br />

appropriate way. In case <strong>of</strong> a positive variance term σ 2 , the constra<strong>in</strong>t will be imposed<br />

by def<strong>in</strong><strong>in</strong>g<br />

ψσ = log σ 2 , −∞ < ψσ < ∞. (3.63)<br />

Once the loglikelihood is maximized us<strong>in</strong>g the transformed but unconstra<strong>in</strong>ed parameter<br />

ψσ, the constra<strong>in</strong>ed parameter can be calculated by back-transformation:<br />

ˆσ 2 = exp( ˆ ψσ), ˆσ 2 > 0. (3.64)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!