# Classifying $900 Million in California Campaign Contributions CalMatters analyzed massive political donation data to reveal industry influences on ballot measures. - **Source:** [CalMatters](https://calmatters.org) - **Published:** 2026-11-10 - **Topic:** [politics](https://brifing.si/c/politics) - **Original article:** - **On Brifing:** https://brifing.si/s/cb3c470e-67cd-4367-bbda-ee4cce5727bb ## Mapping Political Funding Sources CalMatters analyzed nearly $900 million in campaign contributions during the 2025-2026 election cycle. The project aimed to clarify the origins of funds supporting gubernatorial candidates, insurance commissioners, and statewide ballot propositions. By categorizing these massive financial flows, the team provided transparency into the diverse institutions and individuals bankrolling California's political landscape. ## Methodology for Industry Classification The team focused on contributors donating at least $5,000 to specific campaigns. They developed a custom industry schema, largely aligning with NAICS standards while adding specific categories like technology. Contributors failing to meet clear industry criteria, such as retired individuals or those with multiple employers, were grouped into a Major Donor category. ## Leveraging External Data Sources To ensure accuracy, researchers cross-referenced contribution records with the H1B Employer Data Hub and California’s Employment Development Department. They also integrated data from OpenSecrets to align with established political finance coding systems. These external databases provided foundational industry codes, which were then normalized to match the specific needs of the project. ## Keyword-Based Categorization Strategy A custom list of keywords was developed to automatically assign contributors to specific industries based on their name, employer, or occupation. For instance, terms like cattle rancher or law office triggered specific classifications. This method served as an initial filter to process the tens of thousands of records collected during the election cycle. ## Machine Learning Model Implementation For entities not identified by keywords, a machine learning classifier was employed. Using Microsoft’s all-MiniLM-L6-v2 sentence transformer, the system converted entity names and occupations into embedding vectors. A logistic regression model then predicted the most likely industry category, providing a confidence score for each classification to ensure data integrity throughout the process. ## Large Language Model Integration CalMatters utilized Claude and Gemini to classify unique contributors by summarizing their business activities. These models performed web searches to identify industry affiliations, assigning confidence levels to their findings. LLMs were never the sole source of truth; their outputs were cross-referenced against other methods to flag potential discrepancies for human review. ## Rigorous Manual Review Process Human reviewers acted as the final arbiter for all flagged data points. Any disagreement between classification methods, such as when an LLM and a machine learning model provided conflicting results, triggered a manual audit. Reviewers verified information using business identity lookups, company websites, and nonprofit databases to ensure the final data was accurate. ## Handling Political Committee Complexity Campaign finance often involves interconnected committees donating to one another, obscuring the original source of funds. To address this, CalMatters categorized contributions from political committees by looking at their own donors who gave $5,000 or more. This approach helped trace the money back to the underlying companies, unions, or associations involved. ## Defining the Technology Category Because standard NAICS codes lack a distinct technology sector, CalMatters created a custom category for this industry. This classification includes companies centered on apps, websites, AI, cryptocurrency, and software. Major entities like Uber, Airbnb, Google, and Meta were grouped here to reflect the high reader interest in the state's tech sector influence. ## Addressing Data Normalization Challenges Data normalization was critical for linking records across different datasets. The team stripped extra whitespace, converted text to consistent formats, and removed corporate suffixes like and affiliated entities. These steps allowed for accurate fuzzy matching between internal contribution records and external sources like OpenSecrets, minimizing errors caused by inconsistent naming conventions. ## Contextualizing Individual Occupations For individual donors, the team used the employer listed in the contribution record as the primary signal. When an employer name was ambiguous, such as Altais, the model used the individual's occupation, like physician, to infer the correct industry. This dual-signal approach significantly improved the accuracy of categorizing individuals within the dataset. ## Transparency in Political Influence This comprehensive classification effort provides voters with a clearer picture of who funds California's political campaigns. By transforming raw, ambiguous data into structured industry insights, the project empowers the public to identify major influencers. The results, published on the Voter Guide, serve as a vital tool for informed civic participation today. ## Sources covering this story - https://calmatters.org/show-your-work/2026/11/how-we-classified-nearly-900-million-in-contributions-from-individuals-companies-and-pacs/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._