We document the systematic failure of four algorithmic trading strategies on Polymarket weather prediction markets. Over six weeks (February–March 2026), we deployed 509 live trades across four strategy generations—ensemble forecasting, smart-money copy-trading, elite wallet replication, and multi-wallet convergence signals—collectively losing USD 401 on a USD 372 starting balance. Meanwhile, 482 paper trades under backtested configurations earned USD 172, exposing a structural backtest-to-live gap. Key findings: - An information-theoretic limit: forecast MAE (2.03°F) equals contract bucket width (2°F), rendering probabilistic edge irreducible regardless of ensemble size - The profit-lock paradox: in binary markets, profit-taking strategies destroy value because exiting a USD 0.20 position at USD 0.30 captures 12.5% of maximum gain while retaining 100% of maximum loss - A two-tier efficiency structure: 276 elite wallets (0.73% of 37,562) maintain 97.8% win rates—likely through latency arbitrage rather than superior forecasting—while the market remains efficient for retail participants - Rapid market crowding: 144% wallet growth in 10 days eroded smart wallet win rates from 84.4% to 81.1%- The paper includes complete trade-level data, two strategy-evolution tables, and visualizations of the forecast-precision bound and crowding dynamics. All findings generalize beyond weather markets to binary prediction markets broadly.
Wenth Andreas (Mon,) studied this question.