Two classes of tests designed to detect changes
in volatility are proposed. Procedures based on squared model
residuals and on the likelihood ratio are considered. The tests are
applicable to parametric nonlinear models like GARCH. Both asymptotic
and bootstrap tests are investigated by means of a simulation study
and applied to returns data. The tests based on the likelihood ratio
are shown to be generally preferable. A wavelet based estimator of
long memory is applied to returns data to shed light on the interplay
of change points and long memory.
Key words: GARCH model, change-point, likelihood ratio,
parametric bootstrap, squared residuals, size-power curves, wavelets.
We propose two multivariate long-memory ARCH models: we first consider
a long-memory extension of the restricted constant conditional
correlations (CCC) model introduced by Bollerslev (1990), and we
propose a new unrestricted conditional covariance matrix model which
models the conditional covariances as long-memory ARCH processes. We
apply these two models to two daily returns on foreign exchanges (FX)
rates series, the Pound-US dollar, and the Deutschmark-US dollar. The
estimation results for both models show: (\(i\)) that the unrestricted
model outperforms the restricted CCC model, and (\(ii\)) that all the
elements of the conditional covariance matrix share the same degree of
long-memory for the period April 1979-January 1997, i.e., after the
European Monetary System inception in March 1979. However, this result
does not hold for the period September 1971-January 1997, and the
floating period before March 1979. Semiparametric methods confirm that
the volatilities and co-volatility of the two FX rates share the same
long-range component, and that the break in the long-term structure is
likely to be caused by the European Monetary System inception.
Key words: long-memory processes, conditional
heteroskedasticity, multivariate long-memory ARCH models,
multivariate FIGARCH models, semiparametric estimators.
Researchers have been investigating the long-memory properties of financial time series
either in the mean or in the conditional variance. We show that some financial time
series display long-memory in both their conditional mean and their
conditional variance: we refer to such time series as double long-memory
time series. We model these series by combining a fractionally integrated regression
function and a fractionally integrated skedastic function. We examine the case of a
double long-memory model, the ARFIMA-FIGARCH, and we propose an estimation
procedure by QML which requires pre-sample values. We show, by using Monte Carlo simulations,
that the QML estimator of this model has some nice properties: it is root-\(n\) consistent,
asymptotically normal, and the level of bias is negligible.
We study for this model some specification tests, i.e., the Ljung-Box statistics based on standardized
residuals, squared standardized residuals and absolute standardized residuals.
Disregarding the long-memory component in the regression function, even for moderate
values of the degree of long-memory which are not detected by appropriate statistical
tests, results in (\(i\)) an over-parameterization of the regression function,
(\(ii\)) a bias in the estimation of the ARCH parameters of the conditional variance
function, whilst the long-memory parameter of the conditional variance is unaffected
by this misspecification, and (\(iii\)) a rejection of the null hypothesis of the
Ljung-Box statistic based on standardized residuals, whilst the Ljung-Box statistics based on
absolute and squared standardized residuals are unaffected.
Key words: fractionally integrated processes, double long-memory, heteroskedasticity, second generation
models, Monte Carlo methods.