Plural Vote's state-by-state forecasts are informed both by polls and a metric of "media interest". Our predictions are based two-thirds on polling, and the rest on this measurement. We created a comprehensive model using real-time (past 3 months) Google Trends data to gauge media interest. The volume of interest in polarizing or partisan media outlets reflect the partisan mood of the electorate. Our predictions informed by media interest are accurate in past elections, including 2016.
The following are our predictions for 2016 using our "Median Partisanship Mode", applying a regression (regressed on the accuracy of 2012 search trends) to 2016 search trends.
Applied to 2016, our Media Partisanship model would have correctly predicted 46 out of 50 states. The correlation coefficient between the predictions and actual vote margins by state is r = 0.945.
Current 3-month forecast for 2020, based solely on Media Partisanship model:
Note that our measurement of "Media Partisanship" is only 1/3rd of our 2020 forecast prediction, the rest being a polling average.