VotePredictor

VotePredictor is an artificial intelligence forecasting project that predicts outcomes for United States elections, congressional voting behavior, and Supreme Court decisions. The platform generates daily win probabilities for Senate, House, and gubernatorial races using a machine learning model trained on historical polling and electoral results. By providing transparent performance metrics and detailed political data, it serves as a public research tool for understanding American political dynamics. The system utilizes a TabPFN model, detailing its analytical methodology, trained on 25 years of polling and results spanning from 1998 to 2025. For the upcoming midterm elections, it calculates win probabilities for 435 House seats, 100 Senate seats, and 36 governorships to evaluate competitive federal and state campaigns. It tracks the balance of power, calculating the exact likelihood of party control in both chambers based on national and state data. To ensure accountability, the project maintains a comprehensive leaderboard that grades its own past predictions against those of other major forecasters, polling firms, and prediction markets. Across 627 historically resolved contests from 2016 to 2024, the base model correctly called 86.8 percent of the winners. The platform evaluates predictive entities using a win-call Brier score adjusted for race difficulty, preventing analysts from boosting their rankings by exclusively rating safe seats. Beyond calling electoral campaigns, the project provides data on the individuals currently serving in government. It maintains detailed profiles for 976 current and former politicians who have served in the United States Congress from the 114th through the 119th sessions. The database scores lawmakers across 14 distinct issues based on their actual roll-call votes, assigning an overall ideological lean to each representative and senator. Users can explore a political compass, review issue breakdowns, and analyze financial disclosures to determine the declared wealth of elected officials. Visitors can filter the legislative database to find specific members, isolate sitting politicians from former ones, or sort the entire legislative body from most progressive to most conservative. The platform applies its machine learning capabilities to the judicial branch by forecasting decisions made by the United States Supreme Court. The model predicts the exact votes of the 9 sitting justices on pending cases based solely on facts known before a final decision is issued. In a rigorous walk-forward backtest covering 54 terms from 1970 to 2023, the system evaluated 53,291 individual justice votes and achieved a 67.2 percent accuracy rate, outperforming the statistical baseline. The platform profiles every member of the court, analyzing their historical share of conservative versus liberal votes, how often they side with the majority, and their distinct behavior in closely divided 5 to 4 decisions. It also evaluates prediction markets alongside crowd-sourced human forecasts to establish a consensus expectation for high-profile pending litigation.