20.11.2012 Views

recent developments in high frequency financial ... - Index of

recent developments in high frequency financial ... - Index of

recent developments in high frequency financial ... - Index of

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.

Semiparametric estimation for f<strong>in</strong>ancial durations 229<br />

ACD-type <strong>of</strong> model whereas ϕ(u, v) = exp(u + v) = exp(u) × exp(v) represents<br />

a Log-ACD type <strong>of</strong> model with<br />

ψ(¯di−1; ϑ1) = ω +<br />

J�<br />

αj ln di−j +<br />

j=1<br />

L�<br />

βℓψi−ℓ. (4)<br />

With respect to the seasonal component, φ(¯di−1, ¯yi−1,ϑ2), several alternatives<br />

are available. In this class <strong>of</strong> models it is usually assumed that the seasonal term is<br />

somehow related to ti. In order to make this dependence more explicit, we def<strong>in</strong>e<br />

a rescaled time variable, t ′ i , such that<br />

t ′ i =<br />

�<br />

ti −<br />

ti<br />

� ti−to<br />

tc−to<br />

ℓ=1<br />

�<br />

(tc − to) if ti >tc,<br />

otherwise,<br />

where ⌊x⌋ is the <strong>in</strong>teger part <strong>of</strong> x. If the seasonal <strong>frequency</strong> is daily, to and tc<br />

stand for the stock market open<strong>in</strong>g and clos<strong>in</strong>g (<strong>in</strong> seconds), respectively; i.e.<br />

to = 9.5 × 60 × 60 and tc = 16 × 60 × 60 if it opens at 09:30 and closes at<br />

16:00. Note that def<strong>in</strong>ed <strong>in</strong> this way, t ′<br />

i ∈[to,tc] is <strong>of</strong> bounded support and is the<br />

time-<strong>of</strong>-the-day (<strong>in</strong> seconds) at which the event has occurred. By chang<strong>in</strong>g the<br />

values <strong>of</strong> to and tc, other possible types <strong>of</strong> seasonality (hourly, weekly, monthly,<br />

etc.) can be considered.<br />

In the context <strong>of</strong> regularly spaced variables, several functional forms for<br />

φ(·; θ2) have been proposed <strong>in</strong> literature. If the function is assumed to fall with<strong>in</strong><br />

a known class <strong>of</strong> parametric functions then, by substitut<strong>in</strong>g (3) <strong>in</strong>to (2), we obta<strong>in</strong><br />

the follow<strong>in</strong>g log-likelihood function<br />

Ln (d; ϑ1,ϑ2,ξ) =<br />

(5)<br />

n�<br />

log p � di| ¯di−1, ¯yi−1,t ′ i−1 ; ϑ1,ϑ2,ξ � , (6)<br />

i=1<br />

and the parameters ϑ1, ϑ2, and ξ can be estimated as <strong>in</strong> the one component case<br />

(2). If the error density is correctly specified then standard ML estimation methods<br />

apply.<br />

But if very little <strong>in</strong>formation is available about the functional form that relates<br />

seasonality and time-<strong>of</strong>-the-day, the risk <strong>of</strong> misspecification <strong>in</strong> choos<strong>in</strong>g φ(·; ϑ2)<br />

is <strong>high</strong>. Consequently, it is worth the use <strong>of</strong> nonparametric methods to approximate<br />

the unknown seasonal function. Let φ(t ′ i ) be the determ<strong>in</strong>istic seasonal component<br />

at time t ′ i . Given the proposed specification for the components, (3) becomes<br />

� �<br />

= ϕ ψ(¯di−1, ¯yi−1; ϑ1), φ(t ′ i−1 )� , (7)<br />

E � di| ¯di−1, ¯yi−1,t ′ i−1<br />

where ϑ1 and the function φ(·), evaluated at time po<strong>in</strong>ts t ′ 1 ,...,t′ n , have to be esti-<br />

mated. Follow<strong>in</strong>g the standard approach (as <strong>in</strong> (2) or (6)), one would be tempted to<br />

obta<strong>in</strong> estimators for ϑ1, ξ, φ(t ′ 1 ),...,φ(t′ n ), by choos<strong>in</strong>g the values that maximize<br />

the follow<strong>in</strong>g log-likelihood function<br />

Ln (d; ϑ1,ξ) =<br />

n�<br />

i=1<br />

log p � di| ¯di−1, ¯yi−1; ϑ1,φ � t ′ � �<br />

i−1 ,ξ . (8)

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

Saved successfully!

Ooh no, something went wrong!