The answer(s) could come in at almost any step in their forecasting process, probably several. By breaking down the inputs, the process, and the outputs, we can at least get a sense of the possible answers to the question.
These models primarily use polls as input data. They may include other sources like prediction markets. The Economist model may use different sources or put more trust in some polls than others, compared to e.g. 538.
There is also historical data, which may be relied on differently by different models.
The Economist model simulates votes randomly to see what might happen. To do that, it needs to make a lot of assumptions about the process, including the distributions of votes for each state, how the states are correlated, and what the sources of error in their forecasts could be and how much impact they could have.
- Maybe the polls don't represent the true population of "likely voters".
- Maybe the likely voters don't represent the "actual voters", e.g. maybe people who work tough jobs systematically turn out less even though they intended to vote.
- Maybe there are other impacts on vote totals, such barriers to vote stemming from too few drop boxes or polling places and long lines. Maybe a large number of mail-in votes will be rejected for signature mismatches, or a state court will decide to invalidate them for other reasons.
- Maybe people lied to the polls in the first place (or were lying to themselves at the time).
You have to decide your sources of error around all these things. When I look at the Economist's forecast for Michigan, they predict vote totals that match the polls on average, and they predict a very high probability that the outcome will match the majority in the polls. That indicates they think their sources of uncertainty above are low (for some reason). This was a big concern after the 2016 election, and with 538's prominent place in the discussion, they've focused a lot on this issue.
But this is not enough. You also need to decide your sources of uncertainty around correlation between states. If Michigan is an upset, what is the chance that Wisconsin is also an upset? The Economist has a lot of details on their beliefs here, but it could be that they put less uncertainty on these factors than other models.
The Economist model may be predicting something different than other models. Consider just the following two examples:
- If nothing newsworthy happens before election day, who (if anyone) would get a majority of cast ballots in a set of states totalling 270+ electoral votes?
- Who (if anyone) will record a majority of the officially-announced vote total in a set of states totalling 270+ electoral votes?
- Who (if anyone) will be sworn in as President on January 20, 2021?
Between 1 and 2 are many sources of uncertainty: How does turnout among poll respondents compare to general population? How is that biased by party? Will there be any game-changing events such as a scandal unfolding in the next week? Etc. There are also sources of uncertainty between 2 and 3, such as whether the Electoral College votes actually match the state votes.
I want to add another point: incentives. It's not clear that the Economist optimizes their own goals by given the best possible forecast (whatever that means). People often evaluate forecasts in a winner-take-all manner: If you predicted Biden at 60% and he loses, you're heavily criticized for being "wrong". And if he wins in a landslide, you might still be considered "wrong". If you predict 90%, then at least if it does end up being a landslide for Biden then people will praise you. Another more nefarious possibility is that a forecaster biases their prediction in hopes of swaying the actual election outcome.
I don't actually think these are happening here, but it is something to be concerned about in election forecasting.