I am currently making a delta predictor, but I stumbled on a big issue:
The delta prediction during is off by a lot from the actual delta, that's because after a contest, many people get hacked (on Div3/4/Edu) and get banned too, so it is really hard to get your actual rank at the rating change time
How did extensions like carrot calculate delta? I actually want to know how
I honestly have no idea. I recently tried to use API to scrape the rating history of every user for my next contrib farm blog and it runs at like 2 users per second. I think carrot uses just the
ratedListendpoint that gives the raw ratings of everybody but that is flawed because it doesn't include the "ramp up" (if you have done one contest and your rating is 550 then it's actually 1450 internally).Also another thing is are you removing yourself in the placements? Because if you beat yourself then you will be inflating the placings up a bit. I think adding 0.5 (feels right to me, haven't proved it) or removing yourself from the list should work.
Personally, i use carrot and even it's not very accurate, as you mentioned, due to hacking phase and people getting banned. A good rule of thumb is your estimate delta is around +20 of what carrot shows.