Our 2020 forecast is generated by a model that blends data from the past with data from the present. The parts of the model that learns from the past is what we call the Historic Model.
How it works:
- Our Historic model is made from 2 sub-models that detect voting patterns in all 275 constituencies over time, and then uses those patterns to simulate the election more than 10,000 times to generate the 2020 forecast.
- The first of the sub-models is a machine-learning model that is fed more than 2000 data points per political party. The model is finely tuned to detect voter patterns across the country by learning from all past elections from 1996 to date.
- After the AI model learns, it then generates baseline predictions for 2020 that encapsulate a set of key variables or conditions that are likely to have the biggest impact in this year’s election.
- That set of key variables are then fed to a statistical which uses it to simulate the election more than 10,000 times. The goal is to generate a distribution of possible outcomes on Election Day.
- It is from this distribution of likely outcomes that the election forecast is generated. This also explains why our forecast is probabilistic and not deterministic: the chance of winning is based on the number of simulations in which a given party received 50% + 1 vote.
- One big drawback of our historic model is that it’s ignorant of current conditions by design. This is why we combine the output of the model with that of the Polls-Model, a statistical model that is tuned to capturing the present mood of the electorate.
