This might a somewhat silly question, but how does one estimate the covariance matrix? I know the formulae; however, whatever one measures depends heavily on the data that is used (particularly in terms of the look-back period), and the actual data set (i.e. daily, weekly, monthly returns). What the lookbackperiod does not address is the stability through time of the matrix; if I "measure" the matrix today, there is no guarantee that it will give the correct covariances for tomorrow, etc. and I haven't found anything to go on. Any thoughts on this? (Most textbooks don't mention the difficulty of actually coming up with a co-variance matrix, which is really frustrating.)

for what purpose do you want covariance matrix... finding VaR, monte-carlo, or calibration.. need for accuracy depends..

VaR analysis and Monte-Carlo in fx...

Well what are you doing? single-period analysis or multiperiod? If you believe your lookback frame is an unbiased estimator for the future, then that will be your expected covariance going forward.