The forecast is now final and will no longer update.
The forecast is a complex model that relies on many factors to make its prediction. This text will try to walk you through how it works.
In it's beginning, the forecast makes five different predictions. They are based on the results of the last presidential election, the results of the last House election, the betting odds, the RealClearPolitics averages and the FiveThirtyEight forecast. Some of these data points are adjusted for this first prediction. An example for this is the prediction based on the results of the last presidential election which takes into account the shift it anticipates to occur in the national popular vote. It is important to note that some of these first five predictions do not exist in every state, as for example the RCP polling averages.
Based on an average of this first five predictions, a state importance index is calculated for every state. The index determines how important a state is for the outcome of the election and is dependent from both a states number of electoral votes and anticipated vote margin. This state importance index is then used to calculate a weighted national popular vote prediction for every one of the first five predictions.
Based on these weighted national popular vote predictions, the five predictions are then separated into an relative and absolute prediction of the results of the election. Based on how these specific relative and absolute predictions performed in the last four elections, they are given different weights by an advanced preditcion algorithm. With these weights, a final prediction is calculated for every state.
Based on the error size of predicting the 2020, 2016, 2012 and 2008 elections with the data from the other three elections respectively, the forecast develops a simulation of the presidential election, which, when applied a multitude of times, can offer insight into how likely specific outcomes are. During the development of the forecast, the algorithms to create the original five predictions, the state importance index and the weights for relative and absolute predictions have been improved to minimize the error by whose assumption they simulate the results.