At Deck, we build predictions using an approach we call “contextual inference.” This approach uses hard data on individuals’ past behaviors, and the context around those behaviors, to predict what new people in new contexts might do. In this documentation, we’ll discuss how we are using that approach to determine the likelihood a given person will make a contribution to a specific political campaign.

If you’ve received emails or text messages asking you to donate to a campaign you’ve never heard of, you’ve probably been identified as a likely Democratic campaign contributor by a fundraising data vendor. These vendors create lists of people from across the country who have contributed to other campaigns in the past and sell those lists to campaigns eager to get their fundraising operation off the ground. Often, these vendors will charge campaigns some percentage of the donations raised from prospects on these lists.

While these lists have been helpful, they also have drawbacks — such as exacerbating donor fatigue, shifting a campaign’s focus to out-of-district donors, and leaving prospective contributors who aren’t existing Democratic Party donors on the table.

For these reasons, we’ve built a model that generates person-level, campaign-specific probability scores. This approach (1) takes into account what makes each campaign unique, (2) learns from past donors without limiting future outreach to them, and (3) makes it easier to identify potential donors in your own district.

This model uses traits of voters, traits of candidates, election-related media coverage, campaign finance reports, and election results.

The first step in preparing this model is to assemble its training data. To do this, we combine a registered voter’s traits (including their demographics and socioeconomic details) with the details of each election (including whether it’s a primary or special election, how much media coverage the race was getting, and how financially competitive the race was), and the traits of the individual campaigns on the ballot. Campaign traits we use include candidate’s demographics, incumbency status, and history in office; the volume and type of coverage the campaign is getting in the media; and the demographic and socioeconomic traits of the campaign’s existing donors.

Next, we use our training data to identify the features most likely to have high predictive power -- either alone or in combination with others -- and those most likely to confuse a model into overfitting or diminish the impact of other features. At this stage, we (1) prune highly correlated features and features without meaningful variation, (2) use a technique called VSURF (variable selection using random forest) to better interpret how features will interact with each other, and (3) impute missing data.

Finally, we design a deep learning architecture to predict our outcome. In this case, we’ve built a six-layer neural network. The model uses a sigmoid activation function and the Adam optimization algorithm, optimizing for low binary cross-entropy.

To validate this model, we suppressed a randomly selected batch of campaigns from our training data. We then used the data associated with those campaigns to test the model’s accuracy.

When validated on over 45,000 testing samples, we found that the model’s area under the ROC curve was 0.97, indicating that the model’s ranking of contribution likelihood aligned with actual contributions in our testing samples 97% of the time. The overall binary accuracy of our testing predictions was 77%. And in a lift chart organized by decile, the top decile had a lift of 805 while the bottom had a lift of 0. This means that voters with the top decile of scores were 8 times more likely than a random voter to make a contribution to a specific campaign. Those in the bottom decile were less than 1% as likely to make a contribution.

While it’s difficult to measure variable importance in deep learning models, we can use the output of our VSURF runs to estimate which variables have the most predictive power. The most significant variables are described below, grouped by category.

Voter traits

Campaign traits