We show that the performance gap between BSDE and PINNs methods stems from discretization bias in Euler-Maruyama, and propose a Stratonovich-Heun BSDE method that eliminates the bias and restores BSDE performance.