It's very hard to come up with reliable likely voter models, not least because there are significant changes in people who vote every time. The opinion polls in the 2017 British general election were out, for example, because in the event unusual numbers of unlikely voters came out to vote Labour. Things were obviously going to be different that time (Labour had elected a deeply unlikey leader and repositioned itself) but no-one had any solid evidence how turnout patterns would change, or to what degree.

So yes, inaccurate likely voter modeling is a serious problem. But it's really hard if things are changing, politically.