Quick answer
AI tools and agents help with prediction markets in four concrete ways: researching and summarising information, estimating probabilities, monitoring prices and news, and placing orders automatically through official APIs such as Polymarket's CLOB API and Kalshi's REST/WebSocket/FIX APIs [1] [2].
- On ForecastBench, several AI systems are now statistically indistinguishable from human superforecasters, and in July 2026 one system ranked above the superforecaster median on market questions for the first time [3].
- In Metaculus's Spring 2026 FutureEval, a team of 10 pro forecasters still beat the top-10 bot team by 1.25 points per question, but the gap was not statistically significant [4].
- Agents cannot guarantee profit, legally bypass geo-blocks or KYC, or trade on inside information — the CFTC has already settled an insider-trading case on event contracts [5].
New to the topic? Start with our prediction markets guide.
What can AI agents do in prediction markets?
| Task | What AI does | Typical inputs / tools | Main limitation |
|---|---|---|---|
| Research | Finds, filters and summarises news, filings and data | Search, retrieval (RAG), LLMs | Hallucination; stale or biased sources |
| Forecasting | Turns evidence into a probability | Ensembles of LLMs, base rates, time series | Calibration drifts; small samples |
| Monitoring | Watches prices, order books and news for changes | WebSocket feeds, odds APIs, alerts | Latency vs. professional market makers |
| Execution | Places and cancels orders under rules you set | Polymarket CLOB API, Kalshi API | Legal eligibility, key security, runaway losses |
| Cross-market comparison | Compares sportsbook odds with prediction-market prices | Odds APIs (e.g. Odds-API.io, SportsGameOdds) | Different rules, fees and settlement between venues |
How does AI help with research and data aggregation?
The strongest AI forecasters are pipelines, not single prompts. The Forecasting Research Institute describes the top ForecastBench system in July 2026 (from Cassi AI) as “a multi-stage pipeline that generates sub-questions and search queries, retrieves up-to-date context, and filters for relevance and recency”, then forecasts with an ensemble of models and a final LLM that reviews the reasoning [3].
For prediction markets, that means an agent can:
- read a market's resolution rules first (the rules, not the title, decide the outcome) [6];
- gather official sources named in those rules (agencies, league feeds, election authorities);
- pull market metadata and prices from official APIs [7] [2];
- log every source it used so a human can audit the conclusion.
Can AI forecast better than humans? What the benchmarks say
ForecastBench (Forecasting Research Institute)
ForecastBench compares AI systems with human superforecasters on two question types: dataset questions generated from time series (ACLED, DBnomics, FRED, Yahoo! Finance, Wikipedia) and market questions drawn from Manifold, Metaculus, Polymarket and the RAND Forecasting Initiative [3]. As of 16 July 2026:
- several submissions (from Cassi AI, xAI and Google DeepMind) were statistically indistinguishable from superforecaster accuracy on the tournament leaderboard;
- Cassi's system ranked above the superforecaster median on market questions — the first time any model had done so;
- 17 submissions ranked above superforecasters on the preliminary (dataset-question) leaderboard.
FRI's own caveats matter: superforecaster predictions were last elicited in 2024 and are extrapolated, confidence intervals overlap, and the results are “more consistent with superforecaster parity than with outperformance” [3]. Live rankings: [8].
Metaculus FutureEval (Spring 2026)
173 bots forecast 297 scored questions resolving between January and May 2026. Metaculus's 10 Pro Forecasters beat the top-10 bot team by an average of 1.25 head-to-head points per question, but the edge was not statistically significant (one-sided p = 0.247), and the bot team improved notably over the year [4].
What this means for traders
Matching superforecasters on a benchmark is not the same as beating a liquid market after fees. Market prices already aggregate many informed traders, and fees on Kalshi and Polymarket peak around 50% prices [9] [10].
Odds and data APIs that AI agents use
Agents need clean, structured data. Besides the official Polymarket and Kalshi APIs, two commercial odds APIs are widely used for sports and cross-market analysis. Details below are as published on each provider's site on 2 November 2026 — check current terms before buying.
Partner link — we may earn a commission. See our affiliate disclosure.
Odds-API.io Partner link Featured partner
Odds-API.io is our featured odds-data provider for AI and agent projects. It publishes the broadest bookmaker coverage of the APIs we reviewed and ships an official MCP server, so assistants such as Claude, Codex or Cursor can query odds directly. All details below are as published on its website on 2 November 2026 [11].
Coverage
- 450+ bookmakers, 34 sports, 12,000+ leagues and 500+ market types, including bet365, Betfair Exchange, Betano and 1xBet plus local books in Europe, Australia, the US and Latin America [11].
- Pre-match and live (in-play) odds, with live scores, match clock and status; finished events include final results for settlement [11].
- Player props from 104 bookmakers, plus corners, cards, Asian handicap, half-time lines and team totals where books offer them; esports such as CS2, Valorant, League of Legends and Dota 2 [11].
API features
- One schema across bookmakers: events, teams and markets are matched across books, so the same event returns the same id everywhere [11].
- REST v3 endpoints for odds (up to 30 bookmakers per event), multi-event odds, updated odds since a timestamp, line movements, events, live events, event search, participants, value bets, arbitrage, dropping odds, historical events, historical odds and closing lines (for CLV tracking) [11].
- WebSocket streaming (
wss://api.odds-api.io/v3/ws) pushes odds, score and status changes; it is a paid add-on [11]. - Direct bet links to each bookmaker's event page, and official Python (
pip install odds-api-io) and TypeScript (npm i odds-api-io) SDKs [11].
MCP server for AI agents
- Official, free, open-source (MIT) MCP server,
odds-api-mcp-serveron npm, with 25 tools covering sports, bookmakers, leagues, markets, events, live events, odds, odds movements, dropping odds, value bets, arbitrage, historical odds, settlements, bet grading and docs [12]. - Works with MCP clients over stdio, including Claude Code, Claude Desktop, Codex, Cursor, VS Code with GitHub Copilot, Windsurf and Zed. It runs on your machine and sends your key only to api.odds-api.io [12].
- Read-only: the MCP server cannot place bets. Docs are also published as llms.txt for assistants without MCP [12] [11].
- Example setup:
claude mcp add odds-api --env ODDS_API_KEY="your-api-key" -- npx -y odds-api-mcp-server[11].
Pricing (as published on 2 November 2026)
| Plan | Bookmakers | Price |
|---|---|---|
| Solo | 2 | £49/mo |
| Starter | 5 | £99/mo |
| Growth | 10 | £179/mo |
| Pro | 15 | £229/mo |
| Enterprise | Any number | On request |
- Every plan: 5,000 requests/hour with no monthly cap, every sport and market, pre-match and live; annual billing gives 2 months free [11].
- WebSocket costs the plan price again; request add-ons raise the limit to 10,000–30,000/hour. New free keys are paused [11].
- Pay by card, PayPal, Apple/Google Pay or crypto on monthly plans. You may display the data in your own products; reselling the raw feed needs an enterprise agreement [11].
Get an Odds-API.io API key → Partner link
Full guide: Odds-API.io — endpoints, coverage, pricing and code examples.
Worked example: Odds-API.io case study: API key, plans, AI agents, market analysis and placing bets.
Alternative: SportsGameOdds Partner link
SportsGameOdds says it covers 85+ bookmakers across 67+ leagues, including Kalshi and Polymarket prices alongside sportsbooks, over REST and WebSocket [13]. It bills per event (“object”) rather than per market; as of 2 Nov 2026 there is a free Amateur tier (2,500 objects/month, 10 requests/min), Rookie at $99/month and Pro at $299/month [14].
How an agent uses them
A typical agent compares a sportsbook's implied probability (after removing the bookmaker margin) with a prediction-market price for the same event, flags large gaps, and asks a human before acting. Remember that venues define outcomes differently — a sportsbook “match winner” and an exchange contract may settle on different rules.
How do automated trading bots work on Polymarket and Kalshi?
Polymarket
- Orders are EIP-712-signed messages submitted to the CLOB operator, which checks the signature, balance, allowances and tick size, then matches or rests the order; matched trades settle atomically on Polygon [1].
- CLOB V2 has been live since 28 April 2026, with new SDKs (
@polymarket/clob-client-v2,py-clob-client-v2), fees set by the protocol at match time and pUSD as collateral [15]. - Builders must check geographic eligibility (
GET https://polymarket.com/api/geoblock) before placing orders; orders from blocked regions are rejected [16]. - Polymarket's open-source agents framework showed how LLM agents can research and trade, but the repository is now archived [17].
Kalshi
- Kalshi offers REST, WebSocket and FIX connectivity, with separate demo and production environments and credentials — test strategies in demo first [2] [18].
Market making
Polymarket charges no fees to makers and redistributes part of taker fees to market makers through a rebate programme — one reason automated liquidity provision is common [10].
How is AI changing prediction markets themselves?
- Faster price discovery: automated research and monitoring shorten the time between news and price moves.
- More liquidity: fee structures that reward makers encourage automated quoting [10].
- Markets as AI benchmarks: ForecastBench uses questions from Polymarket, Manifold and Metaculus to test AI forecasters [3].
- More scrutiny: regulators are focusing on insider trading, manipulation and wash trading on event contracts [19].
What are the risks of using AI agents for prediction markets?
- Overfitting and false confidence: a backtest that “works” may just fit noise; benchmark confidence intervals are wide [3] [4].
- Rule misreading: an agent that trades on the market title rather than the resolution rules can be confidently wrong [6].
- Manipulation and insider trading: the CFTC says it has authority over fraud, manipulation, wash trading and insider trading on event contracts, and in 2026 settled a case against a trader who used advance knowledge of a speech; he paid $172,539.02 and received a three-year ban [19] [5].
- Terms of service and geo-rules: automation does not change eligibility; using VPNs to bypass restrictions breaks Polymarket's terms [20].
- Security: API keys and wallet keys give real spending power — store them securely, use least privilege and set hard limits.
- Runaway losses: bugs can place many orders quickly. Use position limits, loss limits and a kill switch.
What AI agents can and cannot do (summary)
| Agents can | Agents cannot (or must not) |
|---|---|
| Summarise news and resolution rules in seconds | Guarantee profits or “beat the market” reliably after fees |
| Produce calibrated probability estimates competitive with skilled humans on benchmarks | Know the future — benchmark parity is statistical, not certainty |
| Monitor many markets and odds feeds 24/7 | Legally bypass geo-blocks, KYC or age checks |
| Place and cancel orders through official APIs within limits you set | Use insider or non-public information (illegal) |
| Compare sportsbook odds with prediction-market prices | Assume two venues settle the same event identically |
A responsible blueprint for building a prediction-market agent
- Check eligibility first: confirm the venue is legal for you and check geo-restriction endpoints [16].
- Start read-only: pull markets, prices and odds; no trading keys.
- Forecast and log: record every forecast with sources and a timestamp, and score it after resolution (Brier score).
- Paper-trade or use demo: Kalshi provides a demo environment [18].
- Add hard limits: max stake per market, daily loss limit, kill switch, human approval for large orders.
- Review regularly: compare your forecasts against market prices and benchmarks before scaling.
Frequently asked questions
Can AI predict prediction-market outcomes?
AI can estimate probabilities, and on ForecastBench several systems are now statistically comparable to human superforecasters. That is not certainty, and it does not guarantee beating market prices after fees.
Can I build an AI trading bot for Polymarket?
Technically yes: Polymarket offers a CLOB API with signed orders and official SDKs (CLOB V2 since 28 April 2026). You must be in an eligible jurisdiction, check the geoblock endpoint, and follow the terms of service.
Does Kalshi have an API for bots?
Yes. Kalshi provides REST, WebSocket and FIX APIs, with a separate demo environment for testing.
What odds APIs work with AI agents?
Odds-API.io offers REST, WebSocket and an official read-only MCP server with 25 tools; SportsGameOdds offers REST and WebSocket feeds that include sportsbooks plus Kalshi and Polymarket prices.
Is it legal to use AI bots on prediction markets?
Automation itself is generally allowed through official APIs, but you must be eligible to use the venue, follow its terms, and never trade on inside information or manipulate markets — the CFTC actively polices this in the US.
Are AI forecasts better than the market price?
Not reliably. Market prices already aggregate many traders. Benchmarks show AI matching skilled humans, not consistently beating liquid markets.
About this guide
Pred-Data publishes independent, source-linked explainers about prediction markets and the AI tools used around them. Every factual claim on this page links to a primary or reputable secondary source, checked on the date shown above. We do not run a trading venue, take bets, or give personal advice.
Author: Ken Lin · Corrections: if you spot an error, we review it and date-stamp the fix.
Affiliate disclosure
Some links on this site (marked “Partner link”) may be affiliate links: if you sign up through them we may earn a commission at no extra cost to you. Partner links use rel="sponsored nofollow" and never change our assessment. Platform links in our comparison tables are plain, non-affiliate links to official sites.
Sources
- Order Lifecycle — Polymarket Documentation, accessed 2 Nov 2026
- API Environments and Endpoints (REST, WebSocket, FIX, demo) — Kalshi API Documentation, accessed 2 Nov 2026
- AI models have likely reached parity with superforecasters on ForecastBench (16 Jul 2026) — Forecasting Research Institute, accessed 2 Nov 2026
- Spring 2026 FutureEval analysis — Metaculus, accessed 2 Nov 2026
- CFTC Orders Gabriel Perez to Pay $172,000 for Insider Trading of Mention Market Event Contracts — U.S. CFTC, accessed 2 Nov 2026
- Resolution — Polymarket Documentation, accessed 2 Nov 2026
- Polymarket Documentation — overview — Polymarket, accessed 2 Nov 2026
- ForecastBench leaderboards — ForecastBench, accessed 2 Nov 2026
- Fee Schedule — Kalshi, accessed 2 Nov 2026
- Fees — Polymarket Documentation, accessed 2 Nov 2026
- Odds-API.io — homepage (coverage, endpoints, MCP, plans) — Odds-API.io, accessed 2 Nov 2026
- Odds API MCP server — Odds-API.io, accessed 2 Nov 2026
- SportsGameOdds — homepage (coverage, features, FAQ) — SportsGameOdds, accessed 2 Nov 2026
- SportsGameOdds — pricing & rate limits — SportsGameOdds, accessed 2 Nov 2026
- Migrating to CLOB V2 (live 28 Apr 2026; pUSD collateral) — Polymarket Documentation, accessed 2 Nov 2026
- Geographic Restrictions (API reference) — Polymarket Documentation, accessed 2 Nov 2026
- Polymarket/agents (archived open-source AI agent framework) — GitHub, accessed 2 Nov 2026
- Test In The Demo Environment — Kalshi API Documentation, accessed 2 Nov 2026
- CFTC Enforcement Division Issues Prediction Markets Advisory (Feb 2026) — U.S. CFTC, accessed 2 Nov 2026
- Geographic Restrictions (VPN prohibition, ToS §2.1.4) — Polymarket Help Center, accessed 2 Nov 2026