I dabbled around a bit in the NBA games and quickly saw some strategies that would likely work if tailored more. However, it was entirely clear that the way to win money in those contests would be to build a model that could spit out lineups. In the big tournaments, a model could easily create a large number of high-variance lineups. In head-to-head or smaller games, it could create lower-risk lineups.
I started to draw up a lot of ideas on what the model could do to try and predict things like playing time, share of points/rebounds/assists, etc.
When it came time to actually build it out, I shelved it, largely because it was going to take time and money to work with the type of stats I needed, but also because I'd assume there are already a lot of people on these sites doing this. In fact, I'm assuming everyone who is profitable on those sites is running model-generated lineups. There are still lots of fish pumping money in, but the people taking it out would be at least as, and likely more, sophisticated than what I could do.
There is an extremely popular NBA bettor and he runs a multi-million dollar operation, hiring several analysts to help him make his predictions. In fact, he's actually consulted privately with several NBA teams because his analysis was ahead of what many NBA teams were doing several years ago.
If those are the types of people you're up against, it's a battle of algorithms, not people.