recent developments in high frequency financial ... - Index of
recent developments in high frequency financial ... - Index of
recent developments in high frequency financial ... - Index of
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 227<br />
likelihood approach (see Sever<strong>in</strong>i and Wong (1992) and Sever<strong>in</strong>i and Staniswalis<br />
(1994)).<br />
In this paper we propose a new method to jo<strong>in</strong>tly estimate the parametric<br />
dynamic and nonparametric seasonal components <strong>in</strong> an ACD framework. Estimation<br />
is based on generalized pr<strong>of</strong>ile likelihood techniques. To make <strong>in</strong>ference<br />
on the parameters <strong>of</strong> <strong>in</strong>terest, we need to extend some previous results on i.i.d.<br />
data, Sever<strong>in</strong>i and Wong (1992), to a dependent data setup. Our estimation method<br />
presents several advantages aga<strong>in</strong>st other methods <strong>in</strong> the literature: (1) It presents<br />
closed form seasonal estimators that are very <strong>in</strong>tuitive. The result<strong>in</strong>g nonparametric<br />
estimator <strong>of</strong> the seasonal component is a simple transformation <strong>of</strong> the<br />
Nadaraya–Watson estimator. (2) The statistical properties <strong>of</strong> both the parametric<br />
and nonparametric estimators are well established. We present the asymptotic<br />
distribution <strong>of</strong> both the nonparametric and the parametric estimator. This enables<br />
us to make correct <strong>in</strong>ference <strong>in</strong> the different components. (3) This methodology<br />
provides a data driven method for comput<strong>in</strong>g the bandwidth. On the contrary,<br />
polynomial spl<strong>in</strong>e techniques do not have a method for choos<strong>in</strong>g the number and<br />
location <strong>of</strong> nodes and the proportionality coefficients for the end-po<strong>in</strong>t restrictions.<br />
(4) Multivariate extensions <strong>of</strong> the seasonal estimator are straightforward.<br />
For example, consider<strong>in</strong>g a multivariate exponential distribution, the estimator<br />
is easily adapted to the multivariate case. (5) The decomposition presented <strong>in</strong><br />
the paper can be easily extended to cope with other specifications that are frequently<br />
used <strong>in</strong> the econometric analysis <strong>of</strong> tick-by-tick data. For example, to<br />
capture nonl<strong>in</strong>ear relationships between f<strong>in</strong>ancial durations and market microstructure<br />
variables (see, for <strong>in</strong>stance, Spierdik et al. (2004), and references there<strong>in</strong>).<br />
Likewise, it is also possible to replace the dependent variable by any other<br />
tick-by-tick market microstructure variable and make it a function <strong>of</strong> its own<br />
lags, through the parametric component, and any other variable through the nonparametric<br />
component. In sum, although focused on durations and its nonl<strong>in</strong>ear<br />
dependency with the time-<strong>of</strong>-the-day, the potential applications <strong>of</strong> this model are<br />
very ample.<br />
As an illustration, we apply our method to price and volume durations <strong>of</strong> two<br />
stocks traded on the NYSE. We show that the model is able to correctly capture<br />
the seasonal pattern and it is able to adjust to changes on this pattern.<br />
The structure <strong>of</strong> the paper is as follows. Section 2 develops a general ML<br />
framework for analyz<strong>in</strong>g tick-by-tick data, and proposes the new estimator for<br />
seasonality. Its asymptotic properties are also analyzed. Second, it develops the<br />
same method but <strong>in</strong> a generalized l<strong>in</strong>ear model (GLM) framework, which allows<br />
us to use quasi maximum likelihood (QML). Section 3 is devoted to the empirical<br />
application compar<strong>in</strong>g the results with others exist<strong>in</strong>g <strong>in</strong> the literature. We use<br />
density forecast to evaluate the out-<strong>of</strong>-sample goodness <strong>of</strong> fit. Section 4 concludes.<br />
The assumptions and pro<strong>of</strong>s <strong>of</strong> the ma<strong>in</strong> results are relegated to the Appendix.<br />
2 Econometric model and estimators<br />
Let ti be the time at which the ith event occurs and let di = ti − ti−1, where<br />
ti−1