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.
176<br />
order serial correlation for price changes <strong>in</strong> the same directions (upper left and<br />
lower right panel), which underp<strong>in</strong>s the existence <strong>of</strong> a bid-ask bounce. For JBX an<br />
analogue pattern is not observable. Moreover, for HAL the positive autocorrelations<br />
at longer lags <strong>in</strong>dicate that the bounce effect will be compensated <strong>in</strong> later<br />
periods. F<strong>in</strong>ally, note that the negative serial correlation caused by the bid-ask<br />
bounce is a short-run phenomenon.<br />
In order to capture the dynamics <strong>of</strong> the price direction variable, the vector <strong>of</strong><br />
log–odds ratios Λi =(Λ −1i, Λ 1i)′ = (ln[π −1i/π 0i], ln[π 1i/π 0i])′ is specified as a multivariate<br />
ARMA process. The f<strong>in</strong>al form <strong>in</strong>clud<strong>in</strong>g possible explanatory variables<br />
is:<br />
i ¼ Pm<br />
i ¼ þ Pp<br />
l¼0<br />
GlZ D i l þ i<br />
Cl i 1 þ<br />
l¼1<br />
Pq<br />
Al i l<br />
l¼1<br />
(2.13)<br />
with {Cl, l: 1→p} and {Al, l: 1→q} be<strong>in</strong>g matrices <strong>of</strong> dimension (2×2) with the<br />
elements {c (l)<br />
hk} and {a (l)<br />
D<br />
hk} and μ =(μ1, μ2)′. The vector Zi conta<strong>in</strong>s additional<br />
explanatory variables captur<strong>in</strong>g other marks <strong>of</strong> the trad<strong>in</strong>g process (market microstructure<br />
variables) with {Gl, l: 0→m} as the correspond<strong>in</strong>g coefficient matrix<br />
(l)<br />
and typical element {ghk}. The vector <strong>of</strong> log–odds ratios is driven by the mart<strong>in</strong>gale differences:<br />
i¼ð 1i; 1iÞ<br />
0 ; with ji ¼ xji ji<br />
q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; j ¼ 1; 1; (2.14)<br />
ji 1 ji<br />
R. Liesenfeld et al.<br />
(2.13)<br />
which is the standardized state vector xi. In this ACM-ARMA(p,q) specification,<br />
the conditional distribution <strong>of</strong> the direction <strong>of</strong> price changes depends on lagged<br />
conditional distributions <strong>of</strong> the process and the lagged values <strong>of</strong> the standardized<br />
state vector. 4 The process is stationary if all values <strong>of</strong> z that satisfy ∣I−C1z−<br />
C2z 2 −⋯−Cpzp∣=0 lie outside the unit circle. Furthermore note, that the existence<br />
<strong>of</strong> a bid-ask bounce would imply that a (1)<br />
12 > a (1)<br />
11 and a (1)<br />
21 > a (1)<br />
22, which means that<br />
the probability <strong>of</strong> an immediate reversal <strong>of</strong> the price direction is <strong>high</strong>er than that <strong>of</strong><br />
an unchanged price direction. 5 The log likelihood <strong>of</strong> the logistic ACM model, the<br />
4 Accord<strong>in</strong>g to the classification by Cox (1981), our ACM model belongs to the class <strong>of</strong><br />
observationally driven models where time dependence arises from a recursion on lagged<br />
endogenous variables. Alternatively, our model could be based on a parameter driven specification,<br />
<strong>in</strong> which the log–odds ratios Λi are determ<strong>in</strong>ed by a dynamic latent process. However,<br />
the estimation and the diagnostics <strong>of</strong> the latter approach results <strong>in</strong> a substantially <strong>high</strong>er<br />
computational burden than for the ACM model. On the other hand, models driven by latent<br />
processes are usually more parsimonious than comparable dynamic models based on lagged<br />
dependent variables. A comparison <strong>of</strong> the two alternatives should be the subject <strong>of</strong> future<br />
research.<br />
5 See Russel and Engle (2002) for a more detailed discussion <strong>of</strong> the stochastic properties <strong>of</strong> the<br />
ACM-ARMA(p,q) model.