Yes, detrending when training can make sense when you have a strong trend in the simulation period and a long/short asymmetry in the system.