I'm doing an implementation of Joshi and Rebonatos Stochastic volatility BGM model (Joshi and Rebonato, 2001, A displaced-diffusion extension of the LIBOR market model)The basic idea of the model is that volatility is made stochastic by letting the coefficients (a, b, c and d) in the volatility function evolve as Ornstein-Uhlenbeck processes.My problem is convergence.They write that "64 volatility paths were sufficient to ensure convergence" using low discrepancy numbers (LDN)I can't get my model to converge for any resonable number of pathsIn my model I employ a hybrid quasi-Monte Carlo method (as suggested by Marco Antonio Guimarães Dias (http://www.puc-rio.br/marco.ind/quasi_mc2.html#HQR)) to avoid dimensionality problems.The procedure is as follows:As the brownian motions of the four coefficients should be uncorrelated I use a different dimension for each.Each coefficient is evolved using a 2x2 matrix containing "hybrid quasi-random" variables.Across paths the first timestep is obtained by normal variables based on a sequence of LDN (van der Corput).This vector is permuted to give the remaining timesteps, whereby the normal variables along each path are uncorrelated, while across paths they are low discrepancy.Have anyone implemented the model or can anyone see were I'm off?I really need to crack this as calibration is not really feasible if the computation time of each option is not reduced.Thank you in advanceChristian

are you using the black formula conditional on a vol path?

Yes. Using the displaced diffusion Black Scholes formula.