To raise money, recruit volunteers, and persuade and turn out your voters your campaign has to decide who to talk to. Your campaign has some information in VoteBuilder about voters like their party registration and vote history, but that isn't enough information to accurately predict if that voter is a good fundraising or persuasion target.

To help your campaign talk to the right people, Deck develops models that predict how likely someone is to vote, vote for you, donate, or volunteer. Models are not yes or no answers: they don’t tell us if a voter is for sure going to vote or vote for a Democrat or vote for a particular candidate. They do help campaigns focus energy and resources on the people who it is most important to talk to.

These models give each voter in your district a score. A voter with a turnout score of 80% means that our model gives them an 8 out of 10 likelihood of voting in your specific upcoming election. Or a voter with a contributor score of 10% means that of a group of 10 voters with similar traits, 1 of them donated to a campaign similar to yours. Our scores are specific to your campaign and your district.

Deck's support score is also campaign- and voter-specific. If a voter has a support score of 80% that means that of a group of 10 voters with similar traits, 8 of them will vote for your campaign: not a generic Democrat running in your race, but your campaign and candidate specifically. These candidate- and voter-specific scores are available to state and federal general election campaigns.

How does Deck build these models? You can find more in depth information about our data sources and model documentation here. The short answer is that Deck gets a lot of data from a lot of places and then asks a computer program to figure out patterns and similarities. We incorporate contextual information like past election results, socio-economic data, and census block level information. We also include voter specific traits and information as well as candidate-specific data like how the media is covering your campaign, your and your opponent's campaign finance data, and other factors like candidate demographics, endorsements, and political leanings.

After collecting all of this data for your campaign, we ask our model to look for similarities in a subset of all the data we've collected. For example, based on income, how have people voted in the past? This process is called training: the model is learning about the world and how to make predictions based on patterns that happened in the past.

Then Deck gives the model a test: it asks the model to predict behavior based on a new set of data the model hasn’t seen before, data that wasn't in the training subset. Then we see how good the model is and make adjustments from there. We go through a rigorous evaluation process while building our models and after each cycle. You can read more about our validation here.

Then using that model, Deck can predict how likely a specific person is to behave in a specific way. We use this process to build models to predict a voter's swinginess, turnout likelihood, support score, and donation likelihood for a specific person in a specific race.

You can find Deck's scores by selecting your universe: persuade, turnout, register, organize, or fundraise. Then click "Show filters" and you can view the distribution of scores in your district and use our scores to create lists. Learn more about our best practices for using Deck's scores here.