14
day interest rate series, for which the ICSS algorithm does not find any shifts in variance
prior to February 2001.
[Table 3]
The next step was to remove serial correlation and conditional heteroskedasticity
from the series by a GARCH(1,1) model, in which the mean equation includes one lag of
the dependent variable. Surprisingly, this time the ICSS algorithm did not find any
volatility breakpoints over 1997-2002 in any of the stock return series. Therefore, the
presence of conditional heteroskedascity would be driving the results reported in Table 4
for the most part. (In general, serial correlation in daily returns is low). Regarding interest
rates, the ICSS algorithm still detects some variance shifts: January 2001 and April 2001
for the 30-day interest rate, August 2001 for the 60-day interest rate, January 2000 and
March 2001 for the 90-day interest rate. For the 180-day interest rate, however, no breaks
are found. It is interesting to see that for the 90-day interest rate, the number of breakpoints
detected is substantially reduced. This again suggests that conditional heteroskedasticity
played an important role in our previous conclusions.
With respect to wavelet analysis, some evidence against the null hypothesis of
variance homogeneity is found in three indices for 1997-1998. Specifically, the null
hypothesis is rejected for Europe at scale 2 at the 10 percent significance level, for Latin
America at scales 1, 2, 5 and 6 at the 10 percent significance level, and for North America
at scale 4 at the 5 percent significance level. In other words, the Asian crisis would have
caused variance shifts elsewhere, after controlling for conditional heteroskedasticity. Over
1999-2002, no breaks are found in any filtered series at any scale (Table 4).
[Table 4]