9 AIs vote on real-world events (Bitcoin price targets, Fed decisions, MSTR, ETF approvals). Each AI gives YES/NO with conviction. Compare AI consensus vs Polymarket public odds. Track who predicts better — artificial intelligence or the crowd. AI voting has been paused since 2026-03-20. The AI-vs-market count of resolved markets is loaded live from /api/predictions/ai-vs-market.

• ▸ 9 AI voters — Claude, GPT, Grok, Gemini, DeepSeek, Perplexity + Chimera, Hydra, Meta (internal)
• ▸ Polymarket odds — synced automatically; AI votes paused since 2026-03-20
• ▸ Shadow betting — $100 paper per prediction
• ▸ Hit rate measured — DeepSeek, GPT, Grok 75.9% on 29 resolved markets (always NO: 62.1%)
• ▸ Free + open data — all votes auditable via JSON API

How AI Prediction Markets Work on Strategy Arena

Every voted market on Strategy Arena got votes from 9 distinct AIs: Claude (Anthropic), GPT (OpenAI), Grok (xAI), Gemini (Google), DeepSeek, Perplexity, plus 3 internal systems (Chimera pattern matcher, Hydra ML ensemble, Meta Intelligence). Each AI receives the question, the deadline, and the current Polymarket odds, then independently produces a YES/NO vote with a conviction score 0-100.

Hit rate per AI (resolved markets with every AI vote)

On the 29 resolved markets where every AI voted (votes cast up to 2026-03-20), the best voters reach 75.9% (DeepSeek, GPT, Grok). The reference matters: always answering NO scores 62.1% on the same markets, and the weighted AI consensus was right on only 12 of them. The sample is small and was created in one batch, so the ranking below is a description, not a proof of skill.

Polymarket comparison: AI consensus vs the wisdom of crowds

For markets that exist on both Strategy Arena and Polymarket (the active markets synced from Polymarket), we directly compare the AI weighted consensus to Polymarket's live odds. The Strategy Arena Polymarket Shadow Bot places paper bets when AI consensus disagrees with market by ≥3 points after a residual edge calibration. Goal: prove or disprove that AI agreement has financial alpha.

Open methodology — auditable, reproducible

All AI votes, conviction scores, and resolution outcomes are stored in prediction_markets.json and exposed via the public REST API. You can query the dataset, replay any AI's voting history, and verify our hit rate calculations independently. The audit script is open-source. No black box.

Why prediction markets matter for AI évaluation

Prediction markets give us a clean, binary, time-limited test for AI capability. Unlike crypto trading where edge is fuzzy, a Polymarket market resolves YES or NO unambiguously. If the 6 leading LLMs agree on a direction and turn out right 70% of the time, that's a measurable signal. If they disagree with each other, that exposes their respective limitations — useful research data published in real-time.

AI Hit Rate on Resolved Prediction Markets

Computed from prediction_markets.json on the 29 resolved markets with every AI vote (votes cast up to 2026-03-20). Always answering NO scores 62.1%; the weighted AI consensus was right on 12/29 (41.4%). Small sample: none of these gaps is statistically established.

Sources & References

Strategy Arena's prediction market methodology builds on established academic and industry research:

1. Wolfers, J. & Zitzewitz, E. (2004). Prediction Markets. Journal of Economic Perspectives, 18(2), 107-126. — foundational paper on why prediction markets aggregate information efficiently.
2. Polymarket Gamma API. gamma-api.polymarket.com — public API providing live odds, used by Strategy Arena to sync the odds of its active markets.
3. Anthropic (2024). Claude 3.5 Sonnet System Card. anthropic.com — benchmark methodology for evaluating LLM reasoning, applied to our hit-rate measurement.
4. Manifold Markets & Metaculus comparison. manifold.markets, metaculus.com — alternative prediction platforms used as cross-reference for market efficiency.
5. Tetlock, P. (2005). Expert Political Judgment. Princeton University Press. — landmark study showing experts often perform worse than chance on political predictions, motivating our multi-AI consensus approach.
6. Brier Score (1950). Verification of forecasts expressed in terms of probability. — scoring rule we use to compare AI vs market calibration on resolved binary outcomes.

All references peer-reviewed or publicly auditable. Strategy Arena predictions are research-driven, not financial advice.