Hello,I'm pricing a path-dependent swap which has the following features:leg 1: vanilla (ex:euribor6m)leg 2: structured ==> if the euribor is < strike (ex.1,5%) then the rate is euribor 6mif the euribor 6m is >= strike then the rate is a fixe rate (ex.3,5%) for the remaining periodsI did use this method (just in case, here is how I proceeded): (see graph obtained below)For each path: simulating short rate process using RANDOMNUMBER and another one (mirror) using -RANDOMNUMBER and then, I just calculate (r+r mirror)/2 for each step(I could have used z and -z, it should be the same no ?)Then I use the resulting paths for the next step of the monte carlo.When I use the antithetic variates, the pay-off is lower than when I don't use it.1) It could be the pay-off structure ? I read we cannot use this reduction technique when the payoff depends "only" or "a lot" on a a highest ? In this case, it could be true because the pay-off is like a digital2) it could be a problem with my methodology ? As you read, I don't calculate the NPV for both normal and mirror path and then /2 => I just calculate (r+r mirror)/2 for each stepThank you for your timeLink to the graph of Convergence

I don't think you can just calculate (r+r mirror)/2 for each step. Wouldn't the numeraires be different on the antithetic and non-antithetic paths?

see for example this linkyou will need to calculate (payoff(r) + payoff(r mirror))/2 for each simulation run.