wilmott.com

Hi all,At the moment i am given the option to study bayesian inference however the course is centred around programming on R (a stats package) and the major area is pretty much on MCMC (markov chain monte carlo simulations).Given the course the Gibbs sampling algo and the Metropolis-Hastings algo is taught to us.Question is do you think the course will be useful? Or is it a complete waste of time and money and i should look for something else?Thanks!

JohannesandErakerare just some of the finance professors who apply MCMC methods to the estimation of financial models. It is an important area.

Statisticians are more likely to be recruited at hedge funds than ibanks, and the use of MCMC methods in HFT is somewhat limited by the fact that algorithms have to run extremely fast. There will be places where you can use it though, but it really depends what you're working on.

QuoteOriginally posted by: acastaldoJohannesandErakerare just some of the finance professors who apply MCMC methods to the estimation of financial models. It is an important area.What academics find interesting is only loosely correlated with what a bank/fund will actually pay you to do.

First, Bayesian approaches are somewhat limited in IB. Secondly, MCMC methods are sometimes incredibly slow and it is often hard to tell if the invariant distribution has been reached (or the algorithm is still in the "burn-in" period). Indeed, one can see that as an opportunity for investigating these MCMC algorithms and make them more suitable for financial applications. People (in academia) do use a lot of MCMC techniques when trying to estimates diffusion models.

Depends on:Where the course is and who teaches it.What you actually want to do.What your background is. As others have said, they are interesting useful things, but rather niche in application.

QuoteOriginally posted by: acastaldoJohannesandErakerare just some of the finance professors who apply MCMC methods to the estimation of financial models. It is an important area.Nice links.Newbie question:Here is also an article on volatility fitting and GARCH model Kim et al Is this a good approach?I have no insights here (learning by osmosis but could this approach be an alternative to other methods, e.g. methods used in combination with PDEs? If the only impediment is performance but how applicable is the method in general?

The procedure in the Kim et al paper is fast and efficient I would say. You can also check their follow up paper.However I think the sequential Monte Carlo methods are more interesting eg this paper.

I would also say that there is now quite a gap between what is done inside academia and inside IBs/funds. Perfect sampling of diffusions, very advanced sequential MC and particle filters methods, particle filter inside MCMC schemes, ABC methods, etc... All these techniques are well known in academia, but unfortunately not very spread inside IBs. I think that there is definitely an opportunity here, so go for it...