Prediction markets are inefficient for sports. The crowd is smart, but beatable. We combine the #1 ranked forecasting model with an agentic intelligence harness to extract alpha where retail traders cannot.
Kalshi attracts retail traders, not sharp sports bettors. The prices are decent but far from efficient. Our analysis of 21,039 real games with resolved outcomes proves there is room for a superior model.
The obvious approach: find where sportsbook consensus disagrees with Kalshi, and bet the discrepancy. This loses money. We tested it on 21,039 real games.
| Metric | Sportsbook Consensus | Kalshi Price |
|---|---|---|
| Accuracy | 70.0% | 72.7% |
| Brier Score | 0.1901 | 0.1810 |
| Edge ROI (3%+ threshold) | -13.24% | -- |
Conclusion: Kalshi is more accurate than sportsbook consensus. Betting the discrepancy is a losing strategy. The edge must come from a superior forecasting model, not from arbitraging existing price sources.
The #1 ranked sports forecasting model on ProphetArena, the UChicago prediction benchmark. A 32-billion parameter model trained not on human labels, but on actual outcomes.
UChicago's benchmark for prediction quality. Foresight-v3 beats GPT-5, Gemini 3 Pro, and all other frontier models on sports forecasting tasks.
Trained with Future-as-Label methodology: outcome-based reinforcement learning. The model learns from what actually happened, not what humans predicted.
Brier score optimized. When Foresight says 70%, it means 70%. Calibration is the foundation of profitable prediction market trading.
A production-grade intelligence harness built on 11 native Rust modules. Cognitum does not predict outcomes. It turns predictions into profitable, risk-managed bets.
Each component solves the other's weakness. Foresight generates calibrated probabilities. Cognitum validates, sizes, executes, and learns. Together they form a complete trading system.
Every outcome feeds back through SONA. The system learns which market conditions produce the highest edge, which coherence thresholds filter noise vs. signal, and how to adjust sizing over time. This is not a static model. It is a system that improves with every resolved bet.
We follow a disciplined validation path. No real money until the data proves positive edge. Every phase has measurable gates that must be passed before proceeding.
Analyzed 21,039 games. Resolved 302 Kalshi outcomes via API. Proved consensus edges lose money. Established Brier benchmarks. Identified the opportunity.
Foresight-v3 API connected and tested. 15 real predictions made at $0.22 total cost. OpenAI-compatible API. Partnership with Lightning Rod AI established.
Running Foresight on 50+ Kalshi sports markets daily. Tracking predictions vs. outcomes. Computing real Brier score after 200+ resolved predictions. Gate: Foresight Brier must be below 0.181.
Small bets ($1-5 each) on Kalshi. 10 bets/day maximum. Track every prediction, bet, and outcome. Gate: Positive P&L after 100 bets.
Increase to $10-50 per bet. Automated scanning and execution via PM2. Gate: Sharpe ratio above 1.0 over 4 consecutive weeks.
Most Kalshi traders are retail. An AI system with calibrated probabilities, intelligent risk management, and a continuous learning loop has a structural advantage. Selective betting -- only when Foresight + coherence gate agree -- is our edge.