This is The Washington Post's general election night model. The model was created in conjunction with Decision Desk HQ/0ptimus.
The model uses quantile regression for point prediction and quantile regression + conformal prediction to generate prediction intervals. You can read about the methods employed here or for a less technical read see here.
The model in production uses data collected by Decision Desk HQ/0ptimus and The Washington Post.
To see how the model works we have included county-level 2012 and 2016 election results for Georgia, Kansas, Kentucky, Missouri and Texas. The data was taken from the MIT Elections Lab.
We were unable to include the county level data we use in the actual model. Instead this model uses a state level fixed effect and a national fixed effect only. You will see, however, that it still performs well.