Matthew Yglesias has a good discussion of why the poll-based models that give Biden a high probability of winning are probably right, despite the well-known polling errors in 2016. Nonetheless, it seems reasonable to believe that the poll-based models (538, The Economist) are overstating Biden’s chances, for several reasons.
Turnout this year will be unusually difficult to predict. How will the surge in mail in balloting affect turn out? Will it lead to a large increase in voting, likely favoring Democrats, or are many voters likely to leave their ballots on the dining room table, or mail them in too late to be counted? Will weather effects on voting have a partisan slant (in either direction, potentially), given that Republicans are more likely to vote in person? How will COVID affect in person turn out? Any increase in uncertainty favors Trump, given that he is behind in projections.
We don’t know how effective voter suppression efforts will be. How will long lines affect turnout? How widespread and effective will outright intimidation be? Will efforts to intimidate backfire and increase Democratic turnout? Voter suppression tactics have changed enormously in the past few years due to Shelby County, so the current state of affairs may not be reflected in data from prior elections.
Finally, we don’t know what the Courts will do, and how their rulings will affect the vote. The biggest uncertainty is probably what happens to late mail in ballots, but other issues will arise.
As Andrew Gelman (creator of The Economist model) says, poll-based models are of “vote intentions, not of votes as counted.”