FiveThirtyEight explaining their election forecast model changes.

They simplified it so it is easier to understand.

Instead of a full Bayesian model, they are running a polling forecast and a fundamentals forecast separately. And then averaging the two with weights according to how close the election is.

That way the weighting can be clearer to see and you don't get the counterintuitive results of the old model.

Good decision as if you didn't read the methodology you'd be quite confused at how the model was generating its results. In some cases for example a state forecast was outside either the state polling or state fundamentals. This was because it was incorporating data from other states.

I am curious to see how the old model performs though.