Help Request from Someone Who Uses Fourier Transformations. Why is my Synthetic Data Positively Biased?

Basically, I'm using FFT to get magnitude/frequency values and then taking random draws between -pi and pi to create a synthetic dataset. I can attach code but I'm curious if anyone has seen this before. It's particularly bad for AMD.

Charts:

​

AMD

​

​

MSFT

I'm using log returns in the FFT and then using the previous close times the cumulative product of the exponential of the log returns. I use the same process with monte carlo and have not obtained results with anywhere near this level of bias.

Any help is appreciated!