How It Works

When a prediction market offers multiple brackets for an event (e.g., 11 temperature ranges for tomorrow's NYC high), exactly one bracket MUST win and pay $1.00. If the sum of all bracket prices is less than $1.00, buying every bracket guarantees a profit regardless of the outcome.

Real example:On March 21, 2026, NYC's March 22 temperature brackets on Polymarket summed to $0.9245 — an 8.2% risk-free return.

Academic backing: Saguillo et al. (2025) documented $40 million in realized arbitrage profit extracted from Polymarket through this exact mechanism.

Verification rules: Outcomes must be exhaustive (one must happen), mutually exclusive (only one can happen), no new outcomes added after purchase, all resolve from the same source on the same date, and no void/N/A conditions.

Compare prediction market prices to superior external data sources. When the market price diverges significantly from a well-calibrated external estimate, the market is mispriced and we trade the gap.

Weather: GFS, ECMWF, and ICON ensemble models (122 total members) provide probability distributions across temperature ranges. After per-city bias correction (30-60 days of calibration data), these ensembles outperform crowd pricing.

Sports: Consensus odds from 40+ sportsbooks via The-Odds-API. Sportsbooks have decades of calibration. After mandatory vig stripping (proportional method), realistic edge is 2-7 cents per trade.

Economics:Cleveland Fed Nowcast for CPI, CME FedWatch for rate decisions, Bloomberg consensus for GDP/jobs. The Federal Reserve itself found Kalshi's CPI forecasts improve on Bloomberg consensus, but individual brackets can still be mispriced.

People consistently overpay for longshots (cheap contracts, unlikely outcomes) and underpay for favorites (expensive contracts, likely outcomes). This has been documented for 50+ years across horse racing, sports betting, and prediction markets.

Evidence: The GWU paper (Bürgi, Deng, Whelan, 2026) analyzed 300,000+ Kalshi contracts and found contracts priced below $0.10 lose 60%+ of money invested. Contracts priced above $0.50 earn small positive returns for makers.

Platform-specific: This bias exists on Kalshi but NOT on Polymarket (per Reichenbach & Walther, 2025, analyzing 124M trades). We trade this only on Kalshi.

Why calibrating: We need 200+ resolved markets at each price level to build our own calibration curve before trading. Buy favorites at $0.85+ where our data confirms they win more often than the price implies.

Polymarket offers temperature bracket markets for 15+ cities daily (219+ markets). We compare market prices to multi-model weather ensemble predictions with per-city bias correction.

Why calibration matters: Raw GFS output has systematic cold bias that varies by city. Seoul GFS predicted 7.5°C vs actual 11°C (3.5°C error). You cannot trade raw model output — 30-60 days of per-city calibration data is required first.

Multi-model requirement: Using only GFS (31 members), standard error is ±9 cents — need 18-20¢ divergence to be confident. With GFS + ECMWF + ICON (122 members), error drops to ±4.5 cents, and 8-10¢ divergence becomes tradeable.

Resolution: Each city resolves at a specific Weather Underground station (e.g., Seoul = Incheon Airport RKSI). Airport stations can read 2-3°C different from downtown.

Sportsbooks employ professional odds-makers with decades of data. Prediction markets are traded by retail users who overreact to narratives. When vig-stripped sportsbook consensus diverges from prediction market prices by 5+ cents, we trade the gap.

Mandatory vig stripping: Raw sportsbook odds include 3-5% profit margin. Without stripping it, you overestimate edge by ~29%. We use the proportional method to convert raw implied probabilities to true probabilities.

Realistic edge after vig stripping: 2-7 cents. Made viable by Polymarket US's 0.10% fee (essentially free).

On economic data release days (CPI, GDP, FOMC, jobs report, weekly claims), Kalshi offers bracket markets for the exact numbers. We compare bracket pricing to the best available external models.

Data sources: Cleveland Fed Nowcast for CPI, CME FedWatch for interest rate decisions, Bloomberg consensus for GDP and employment. High edge when active (10-20 cents on specific brackets).

Frequency: Only active 8-12 days per month. The system monitors the economic calendar and activates scanning only on data release days.

Platform: Kalshi has the deepest economic markets. Idle cash earns ~4% APY. Fee structure favors maker orders on high-probability brackets.