I am trying to model 2x and 3x leveraged ETFs (SSO, UPRO, etc.) to generate historical data before the inception of those funds. I thought it would be simple: I'm tracking the change to ^GSPC, multiplying it by the factor of x2 and should be getting something close to the actual NAV (or closing price) of SSO before the daily expense fees. I am seeing something strange: that, for the first years the ideal (no fees) synthetic SSO performs better than the actual fund. However, in more recent years the actual SSO perform way better than the ideal (no fees/expenses) synthetic fund. How can it be? How can I model the ETF so it matches the actual fund closely?
I see this guy matches SSO very closely. Although it should be quite simple, I, for some reason, fail to match the ETF for the whole periods of its existance.
The second option that comes to mind (the first being asking "that guy"--assuming "that guy" hasn't already explained what he did in that link you posted) would be to plug the data into a machine learning app and letting it figure out the missing data. Keep us posted.