recent developments in high frequency financial ... - Index of
recent developments in high frequency financial ... - Index of
recent developments in high frequency financial ... - Index of
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
294 V. Voev<br />
with the problem <strong>of</strong> how to reduce the noise <strong>in</strong>herent <strong>in</strong> simpler covariance estimators<br />
such as the sample covariance matrix. Techniques have been developed<br />
to ‘shr<strong>in</strong>k’ the sample covariance (SC) matrix, thereby reduc<strong>in</strong>g its extreme values<br />
<strong>in</strong> order to mitigate the effect <strong>of</strong> the so-called error maximization noted by<br />
Michaud (1989). One <strong>of</strong> the shr<strong>in</strong>kage estimators used among practitioners is the<br />
Black–Litterman model (Black and Litterman (1992)). This model uses a prior<br />
which reflects <strong>in</strong>vestor’s beliefs about asset returns and comb<strong>in</strong>es it with implied<br />
equilibrium expected returns to obta<strong>in</strong> a posterior distribution, whose variance is<br />
a comb<strong>in</strong>ation <strong>of</strong> the covariance matrix <strong>of</strong> implied returns and the confidence <strong>of</strong><br />
the <strong>in</strong>vestor’s views (which are reflected <strong>in</strong> the prior covariance). Further, Ledoit<br />
and Wolf (2003) and (2004) use shr<strong>in</strong>kage methods to comb<strong>in</strong>e a SC matrix with<br />
a more structured estimator (e.g. a matrix with equal pairwise correlations, or a<br />
factor model). The idea is to comb<strong>in</strong>e an asymptotically unbiased estimator hav<strong>in</strong>g<br />
a large variance with a biased estimator, which is considerably less noisy. So the<br />
shr<strong>in</strong>kage actually amounts to optimiz<strong>in</strong>g <strong>in</strong> terms <strong>of</strong> the well-known trade-<strong>of</strong>f<br />
between bias and variance.<br />
Recently, with the availability <strong>of</strong> <strong>high</strong>-quality transaction databases, the<br />
technique <strong>of</strong> realized variance and covariance (RC) ga<strong>in</strong>ed popularity. A very comprehensive<br />
treatment <strong>of</strong> volatility modell<strong>in</strong>g with focus on forecast<strong>in</strong>g appears <strong>in</strong><br />
Andersen et al. (2006). Andersen et al. (2001a), among others, have shown that<br />
there is a long-range persistence (long memory) <strong>in</strong> daily realized volatilities, which<br />
allows one to obta<strong>in</strong> good forecasts by means <strong>of</strong> fractionally <strong>in</strong>tegrated ARMA<br />
processes. At the monthly level, we f<strong>in</strong>d that the autocorrelations decl<strong>in</strong>e quite<br />
quickly to zero, which led us to choose standard ARMA models for fitt<strong>in</strong>g and<br />
forecast<strong>in</strong>g.<br />
The aim <strong>of</strong> this paper is to compare the forecast<strong>in</strong>g performance <strong>of</strong> a set <strong>of</strong> models,<br />
which are suitable to handle large-dimensional covariance matrices. Lett<strong>in</strong>g H<br />
denote the set <strong>of</strong> considered models, we have H ={s,ss,rm,rc,src,drc,dsrc},<br />
where the first two models are based on the sample covariance matrix, the third<br />
model is a RiskMetrics TM exponentially weighted mov<strong>in</strong>g average (EWMA) estimator<br />
developed by J.P. Morgan (1996), the fourth and the fifth represent simple<br />
forecasts based on the realized and on the shrunk realized covariance matrix,<br />
and the last two models employ dynamic modell<strong>in</strong>g <strong>of</strong> the RC and shrunk RC,<br />
respectively. We judge the performance <strong>of</strong> the models by look<strong>in</strong>g at their ability<br />
to forecast <strong>in</strong>dividual variance and covariance series by employ<strong>in</strong>g a battery <strong>of</strong><br />
Diebold–Mariano (Diebold and Mariano (1995)) tests. Of course, if we have good<br />
forecasts for the <strong>in</strong>dividual series, then the whole covariance matrix will also be<br />
well forecast. The practical relevance <strong>of</strong> a good forecast can be seen by consider<strong>in</strong>g<br />
an <strong>in</strong>vestor who faces an optimization problem to determ<strong>in</strong>e the weights <strong>of</strong> some<br />
portfolio constituents. One <strong>of</strong> the crucial <strong>in</strong>puts <strong>in</strong> this problem is a forecast <strong>of</strong><br />
future movements and co-movements <strong>in</strong> asset returns. Our contribution is to propose<br />
a methodology which improves upon the sample covariance estimator and<br />
is easy to implement even for very large portfolios. We show that <strong>in</strong> some sense<br />
these models are more flexible than the MGARCH models, although this comes at<br />
the expense <strong>of</strong> some complications.<br />
The rema<strong>in</strong>der <strong>of</strong> the paper is organized as follows: Section 2 sets up the<br />
notation and describes the forecast<strong>in</strong>g models, Section 3 presents the data set used<br />
to compare the forecast<strong>in</strong>g performance <strong>of</strong> the models, Section 4 discusses the<br />
results on the forecast evaluation and Section 5 concludes the paper.