How It Works
Many Worlds Research uses AI agents and Monte Carlo simulation to produce independent probability estimates for sporting events, then compares them against prediction market prices to surface potential edges.
1. Event Discovery
We continuously crawl Polymarket for upcoming events. We're starting with sports — NBA, NFL, MLB, NHL, and NCAA basketball — with plans to expand to other domains as the platform matures.
2. Agent Debate
For each event, we use Intellidimension's Mesh — a multi-agent debate platform — which orchestrates a panel of AI agents that debate the matchup in depth. They analyze offensive and defensive matchups, pace, shooting, guard play, coaching tendencies, injuries, rest, travel, and more. A chair agent synthesizes the debate into a scenario-weighted analytical report.
3. Many-Worlds Decomposition
Rather than producing a single point estimate, the system decomposes the game into distinct worlds — named game scripts representing different ways the event could play out. Each world has a tilt: an expected margin of victory in that scenario.
For example, a basketball game might decompose into worlds like "Home team interior-possession avalanche" (tilt: +12 points) and "Away team guard-and-rhythm control" (tilt: −8 points), each weighted by how likely that script is to unfold.
4. Monte Carlo Simulation
We run 2,000,000+ Monte Carlo simulations across the worlds using Dirichlet sampling to generate a full probability distribution over outcomes. This distribution is what you see in the histogram on each card — not a single number, but the complete shape of possible results.
5. Market Comparison
The simulation's implied probabilities are compared against live Polymarket odds. The edge indicator on each card shows where our model disagrees with the market:
- Arrow toward a team — the sim thinks that side is more likely than the market does. The number (e.g., "12pt") is the gap between our estimate and the market price.
- No arrow — the sim and market roughly agree, or the sim is less confident than the market (no actionable edge).
For spread markets, the simulation distribution is used to compute cover probabilities for each spread line, giving an independent view of whether a spread is priced correctly.
6. Live Tracking
Market prices are refreshed automatically before game time. Once a game starts, odds are frozen at their last pre-game snapshot and the card shows a LIVE badge — in-game market prices reflect real-time scoring information our pre-game model doesn't have, so we don't compare against them.
After the event resolves, the card shows the actual outcome with a checkmark next to the winning side.
Boosting Events
We can't run every event on Polymarket — each analysis costs real compute. Logged-in users receive credits they can spend to boost events they care about. The more boosts an event has, the more likely it is to be picked up for processing. This lets the community shape which matchups get analyzed.
Accounts & Access
Anyone can browse live and past research without logging in. Upcoming events — the ones you'd want to see before placing a bet — are only visible to logged-in users.
MCP Access
Want to use the data via MCP? See the setup guide at Many Worlds MCP: How To.
The histograms, probabilities, and edge indicators on this site are research outputs — not financial advice.