What's interesting about the models is they only take into consideration polls. This year there are way more online polls, and automated polls than live caller, or in other words shitty polls and they all get thrown into the models. Once early voting starts someone needs to design a model that incorporates that information.
For example all the models, even Princeton, have NV at best 50% chance based off all the polls which show a minor Trump lead, minor Hillary lead. However a majority of the state votes early, like almost all of it, and there is information on that. So early voting 2012 Obama built a firewall of 71,000 votes from Clarke county and won by +7. This year Hillary has built a firewall of 61,000 and still has today's early voting which probably will push it further and very close to Obama.
So you know what's better for any model than hypotheticals which all polls are, is actual data. Feed that information into the models and let it adjust with actual tabulated results and historical information. I'd love to see how different peoples models adjust with real information fed into it.