By the Hawkfish Tech Team
Hawkfish was created to help our clients’ efforts to defeat Donald Trump and push forward progressive values. From building cutting-edge ad data technology to innovative content strategy and creative production, our work proved to be a pivotal part of Democrats’ success in 2020. While Hawkfish won’t continue, we want to share an overview of some of the tech accomplishments, in the hopes that what we’ve learned can be leveraged for future election cycles.
Post-primaries, we successfully supported 27 clients, ran approximately $20 million worth of media across 11 platforms, and developed more than 3,000 pieces of unique creative moving the needle on the Presidential election and driving the conversation on key issues like climate change and Covid response. Together with our clients, our efforts helped drive just over 60,000 voter registration sign-ups in the key battleground states of Arizona, Wisconsin and Georgia, where the Democrats’ margins of victory were under 21,000 votes.
To enable that success, the Hawkfish tech team developed a number of innovative approaches, tools, and products.
What we did
Over the course of the presidential campaign, our tech team focused on the development of high-quality voter data and sophisticated media impact analysis, as well as an advanced machine learning infrastructure. That led to several key accomplishments, including:
- Accurate election forecasting with massively parallelized and novel machine learning
We achieved more accurate election forecasts than any other publicly available prediction in nine of 16 battleground states, including Florida, Michigan and North Carolina, via our novel national-level bagged neural network candidate support model. To meet our massive training, testing, and batch scoring needs, we parallelized on a 2,000 core Google Kubernetes Engine cluster using Kubeflow Pipelines, Kubeflow Katib and Argo.
- Cost-efficient GOTV campaigns via our support and turnout models and early vote data
For our clients’ get out the vote (GOTV) efforts, our work netted the addition of over 80,000 early and absentee votes in Blue Wall states at a cost-per-action (CPA) of less than $10. We optimized spend mid-campaign by analyzing the individual-level early vote results associated with our randomized control experiment for ad performance measurement, then adjusting both our audience and respective creative strategy to effectively reach likely Biden supporters needing mobilization. We verified our results using regression-based analysis on final precinct data. This overall campaign was given an Expy Award from the Analyst Institute.
- Highly optimized persuasion campaigns through our Looking Glass tool
As an example, we targeted a specific demographic in battleground states and drove a 3% increase in Biden support via successful in-flight optimization of our ad campaigns. A key factor was our Hawkfish Looking Glass tool — a rapid, multi-mode, in-field persuasion measurement capability that interfaces with audiences as they are internet browsing to measure lift.
- All of the above enabled by high-quality voter data and infrastructure
We built one of the most expansive, complete and accurate available pictures of people in the U.S through the aggregation and merging of several population-level voter files and smaller commercial data sets, using advanced entity resolution and machine learning methods. To measure success, we benchmarked against one of the standard voter files in the Democratic ecosystem and found:
- Identification of 7 million previously untracked unregistered voters
- An increase of 10 million phone numbers and 40 million email addresses to improve voter contactability and platform match rates
- A 10% increase in demographics accuracy for race and education and a 25% increase in sub-ethnicity classification accuracy
- Identification of 31 million Democratic donors and 69 million progressive issues donors, each representing vast increases over individual input data sources.
All of our successes in data and analysis rest on the machine learning infrastructure we built, tailored for ease of use and support in rapid model development and deployment. Here are some details:
- We balanced flexibility with an opinionated framework. By using Docker and Kubeflow Pipelines (KFP), data scientists had the ability to both choose their stack or use a default suite of reusable KFP components (via Sci-kit Learn Estimator API).
- We developed a Model Quality Control Dashboard in Plotly to provide a suite of diagnostics including: User-defined evaluation metrics, HiPlot parallel plots, Partial Dependence Plots, SHAP Plots, and Conditional Score Distributions.
This work represents the tip of the iceberg in terms of what is possible by applying commercial best practices to political and issue campaigns. We hope to see progress in this area continue.
We’re exceptionally proud of these achievements, and the talented people who made them possible. We know many on our team will continue the vital work of innovating the Democratic and progressive ecosystem’s digital infrastructure, defining industry best practices and spawning innovations that will be the next generation of tech advances for making real change. We look forward to the contributions our Hawkfish tech team members will make in the future.
(Further inquiries can be directed to us via our LinkedIn page.)
