This is a financial & insurance problem we approach through our Data to Data service line.
The problem
Prediction markets list event contracts on elections, economic releases, sport, weather and company news. Each one has a price that reads as a probability, and each one has resolution rules that decide what counts as yes.
A research team, a trading desk or an operator listing these markets faces the same issue: too many contracts to read carefully, analysts who weigh evidence differently, and no record of why a view was taken.
Why it’s hard
Three things make it harder than it looks:
- Resolution rules. Many apparent mispricings are a misreading of what exactly will be judged, by whom and by when.
- Evidence from many sources. News, structured data and the prices of related contracts each need different handling, and they arrive at different times.
- Calibration. A probability is only useful if events given that probability happen about that often. Without scoring estimates against outcomes, overconfidence goes unnoticed.
And a price gap is not an edge until fees, spread and liquidity have been counted.
How we approach it
We build the analysis as a LangGraph graph:
- Read the resolution rules. An agent extracts what counts as yes, the source of truth and the deadline, and flags ambiguity.
- Gather evidence in parallel. Separate agents for news, structured data and related contracts, fanned out and joined in one shared state.
- Estimate the probability, with reasons. The estimate carries the evidence it used, so a reviewer can challenge a reason rather than a number.
- Compare with the price in code. Net of fees and spread. With no gap, the result is logged and the run ends.
- Analyst review. With a gap, the graph pauses through an interrupt for a person to accept or reject the reasoning.
- Calibrate. Every resolved contract scores the estimate that preceded it, and the scores feed back into how much weight the graph gives each source.
What we would not do: let the graph act on a gap without review, trust a single model’s estimate without its evidence, or describe the output as investment advice.
What it takes
- Access to the contracts, prices and resolution rules you work with, through the market’s published interfaces.
- The evidence sources you trust, and the ones you do not.
- A decision on what the output is for: research notes, pricing input for an operator, or a queue for analysts.
- Clarity on the licence you operate under. That shapes what the system may do.
Where this has been done
This is a capability. The probability modelling discipline behind it is delivered work: separating the prediction from the position, which we explain in Predicting an outcome and staking on it are two problems, and backtesting with point in time data. The same agent graph pattern is being built into AnkEDGE. See Agentic AI and LangGraph for other industries.