What Do Political Prediction Markets Add? Evidence from Five U.S. Elections, On-Chain Trader Data and an Interpretable Rule-Based Forecasting Framework (replication package and registered 2026 forecasts)

Replication package, manuscript and registered forecasts for What Do Political Prediction Markets Add? Evidence from Five U.S. Elections, On-Chain Trader Data and an Interpretable Rule-Based Forecasting Framework. The study compares FiveThirtyEight's forecasts with political prediction markets (Polymarket, Kalshi, PredictIt, IEM and aggregators) over the 2016-2024 U.S. federal elections: lead-lag and cointegration tests with simulation nulls for bounded forecasts, an audit of archived 538 vintages, the markets' Republican lean, wallet-level Polygon trade data, and an interpretable rule-based forecasting framework validated on held-out elections. 2026 forecasts (probability that Democrats win control). All inputs were retrieved before any polls opened on Election Day, November 3, 2026. Registered forecast, inputs of October 10, 2026: Senate (rule model V5) 0.728, 80% interval 47-58 Democratic seats; House (rule model H2) 0.639, 80% interval 207-241 Democratic seats. Pre-election update, inputs retrieved 2026-11-03 05:01 UTC (00:01 ET): Senate 0.773, 80% interval 48-60 seats (rules fitted on polls 0-7 days out); House 0.675, 80% interval 208-243 seats (generic ballot D+8.75). Published-model and market benchmarks, exploratory variants and race-level probabilities are in Section 9 and Appendix C of the paper and in out/. Files. prediction_markets_paper.pdf (manuscript; every number generated from the outputs); rules_2026.html (self-contained interactive explainer of the 2026 rules with frozen snapshots); prediction_markets_study.zip (code, outputs, LaTeX source, most raw data, both 2026 input vintages: raw/live2026_20261010/ registered, raw/live2026/ update); companion raw-data archives (crosscontest_raw_part1/2, hist1620_raw_part1/2, house_raw, traders2024_data). See README.md and DATA.md in the main archive for provenance and the rebuild procedure (SKIP_ACQUIRE=1 bash run_all.sh). Data sources are listed in Appendix D of the paper; source data remain subject to their providers' terms.