traday data, similar to Hsieh (1991). Based on a continuous time diffusion
process Andersen and Bollerslev (1998) estimate the one-day Quadratic Vari-
ation (QVl) which is also called integrated volatility and defined as the sum
of the squared returns, for the intraday frequency m, to produce the
daily volatility measure: QVl, discussed in section 1.3, using 5-minute sam-
pling frequency the lag length is τn = 288 for financial markets open 24 hours
per day (e.g. FX markets). QVl can be considered as an efficient estimate
of the quadratic variation of a stock returns process. One reason for their
efficiency being that they utilize the high-frequency intraday data informa-
tion. The QVl filter is generalized in Andreou and Ghysels (2002) using the
results in Foster and Nelson (1996) to increase the window length к = 2,3
days in QVk and to suggest rolling instead of block sampling schemes. The
rolling estimation method yields the one-day Historical Quadratic Variation
(//QVlj defined as the sum of m rolling QV estimates, as discussed in sec-
tion 1.3, which is also extended to a к window length, HQVk. The rolling
estimation method yields smooth volatility filters which answers one of the
criticisms of the QVl filter (see for instance Barnidoff-Nielsen and Shep-
hard, 2001). The K&L and I&T tests are applied to these estimates of the
quadratic variation and compared with the results for (rt)2. The results in
Table 8 reveal the existence of a single change-point that is detected in all
the QV type filters by the Uma,x∕σyARHAc and IT even at the 1% signifi-
cance level as opposed to the mixed evidence of a change-point in (rt)2 and
∣rt∣. This change-point in the quadratic variation of the YN∕USS series is
consistently estimated by the high-frequency volatility filters to be located
on the 8/2/1993 and 9/2/1993 and is associated with the highest increase of
the YN vis-à-vis the US dollar since the 1970s and the possibility of Central
Bank interventions (as published in the Asian Wall Street Journal dated 23rd
February, 1993).
The empirical analysis so far applied single and multiple breaks test pro-
cedures and identified the common dates estimated by the above tests as
change-points. In an approach to verify that there was indeed a structural
change in the asset returns processes we examine the volatility character-
istics of the series in alternative subsamples - prior and after the breaks.
The results in Table 9 report the estimated MLE parameters from a Normal
GARCH(1,1). The varying estimated coefficients of volatility persistence
and unconditional variance over the subsamples can be considered as further
supportive empirical evidence that complements the change-point tests.
17