538 approach is not bad, but I still suspect some pro-D bias there. It is virtually impossible not to introduce your own preferences in the models after tinkering with them for months. Also, 538 and similar tend to paddle dramatically towards Republicans in the days before election so that their final prediction is closer to the final result. At least, it is my recollection. It would be interesting to test these sites about their early prediction – that is when they affect voters, not their last minute position.
As to this election, they went from 35% R- chance in the Senate to 60% at the end during the last month, so the trend would hit 100% by the time they count the votes 🙂