Most polls apply their own “special sauce”.
Individually, they all need to be taken with a grain of salt.
It is more important to observe the trending.
Most poll ‘sauces’ aren’t special, they’re standard.
RV, LV by response, or LV by past history, or historical turnout.
Here, they mixed two standard methods and added a turnout model, so indeed it’s ‘special’.
It’s VERY special:
“The sample was balanced to match the demographic and political characteristics of active registered voters in the L2 voter file by age, race, gender, party registration, region and a modeled turnout score. The voter file data on respondents, not the self-reported information provided by respondents, was used for weighting.
Likely voters were determined by averaging a self-reported likely-voter screen and a modeled turnout score.
Self-reported likely voters were those who indicated that they were “almost certain” or “very likely” to vote, or rated their chance of voting as a “9” or “10” on a scale from 1 to 10.
The turnout score was based on a model of turnout in the 2012 presidential election. The probabilties were applied to 2016.
The probability that a registered voter would turn out was based on the average of whether they were a self-reported voter and their modeled turnout score.”
“... 59 percent of interviews were completed on cellphones”
I wish pollsters would settle on a set ratio of cell/land lines. They’re all over the place on this.