[8note]

The pieces that it does predict some things better:

Is one player happier than the other right now?

How long will it take to make the next move?

How long will the game last?

There's lots and lots of observable things other than who wins that the coin flip model does a worse job at. Thus, the example makes just as good an argument that making better predictions is still what makes a model better.

The missing question is "what are you using the model for?"

Are you trying to learn chess? Understand human psychology? Predict the outcomes of chess games? The same model isn't going to be the best for all of them. Like, the super GM level stockfish is going to suck at predicting beginner chess games

Whether one model is better than another is all about the purpose it's trying to serve, and how well it does that.