Hi all!I have a problem with implementation numerical schemes for backward equations (FBSDE). I try to describe it hoping that maybe some of you came across similar problems and got rid of them.I try to price a contingent claim by means of the abovementioned technic. The solution should by the function of the forward equations - there are well described schemes for this (see papers of Jianfeng Zhang) that seem to be implemented easilly. The procedure is recursive with respect to time variable - from the maturity backwart to time 0. The key point there is generating (some) forward trajectories of portfolio value process from point t_{i-1} to t_i at each node (possible value of portfolio process) and calculate a certain empirical expectation. But it is very sensitive - as i spotted - to the random numbers used. If i use (pseudo) normal distribution (generated for example by Mersenne Twister gen) i hardly obtain series of numbers with mean 0. That is why - as i think - i obtain completely different values of the expectation for to consecutive nodes - these small differences from 0 are magnified. Increasin the numbr of trajectories generated does not help much. An the expectactions should change "smoothly" from one node to another. Did you have similar problems in fact with calculating, say, expectations on the grid?thanks (i hope not too long dicription and not too vague)

How many iterations ?

1000 iterations

If I've got you right, this is one of the common problems with MC.You may be modelling something with a smooth curve, but any of the partial differentials will be very jagged, and your steps represent an approximation to the differential.Have you tried Sobol sequences ?

No, i have not heard about this method - could you give me the reference to article about it (or maybe c++ code). looking thru the web did not give good results. Thanks