QuoteOriginally posted by: AlanThanks again -- two good linksYou're welcome, Alan.There was another thread as well, but I can't find it anymore, don't know the title but it might come back. Search does not really help.. I posted it in July on Student?Title; Stochastic, Bayes??

QuoteOriginally posted by: rezaMCMC vs MLEany reason to perfer one to the other?advantages/ disadvantages ... philosophical differences?for an unusual distribution, both would require simulation and would be computationally expensive ...I wonder why recently most researchers use MCMC for SV problems?This is how I see things though with the caveat that this is a contentious issue:Maximum Likelihood Estimation is the first best approach when trying to approximate the true value of your models parameters. Why? Because it is consistent, efficient and easy to communicate. There is only one parameter value that you need to argue about (academia?s obsession with standard errors is overdone IMHO) and other people can easily replicate your results with other tools and data. For most applicable models (i.e. > 90%) MLE is the best way to go. However if the likelihood function of your model is not well behaved (i.e. positive, semi-definite) you will not have a tractable the analytic solution and an optimizer (i.e. Newton-Raphson) will most likely not converge to the best fit solution if it con verges at all. This happens with non-linear models such as Kalman filters and Hidden Markov Models. In this case you have to be more creative about ensuring that the parameters that you have converge to the real solution. MCMC is one solution to this problem, EM algorithms are another. For SV problems, you need to jointly estimate the separate distributions of the volatility and the returns from the same underlying data. This problem is not solvable from via MLE without highly (in my opinion not credible) restrictive assumptions.