prediction_market_research v1.0.0

name: "prediction_market_research" description: "Read prediction markets as research evidence: decode implied probabilities, audit resolution rules before trusting prices, score calibration with Brier, and spot wash trading and manipulation. Use when an agent needs to interpret Polymarket or Kalshi prices as forecasts, compare markets across venues, or assess whether a market's odds can be trusted."

Prediction Market Research

Purpose

Turn a prediction market page into reliable evidence. A market price is a crowd forecast with money behind it, but only if you read the resolution rules first, check who is actually trading, and know the documented calibration quirks. This skill gives an agent the full read workflow: price to probability, rules audit, calibration scoring, manipulation checks.

Workflow

1. Read the rules before the odds. Open the market's resolution text and extract: exact question, observation window, named resolution source, time zone and cutoff, threshold language (strictly above vs reaches), and edge-case handling (cancellations, postponements, data revisions, conflicting sources). If the research question and the market's rules do not mean the same thing, the price tells you nothing about your question. Stop and say so.
2. Convert price to probability. A binary contract trading at $0.62 implies ~62%. Use the order-book mid between best bid and best ask as the point estimate. A wide spread means a thin, uncertain market; a tight spread means real conviction. Do not vig-remove a single contract price. It is already a probability.
3. Check the venue and liquidity. Note which exchange (Polymarket international vs Polymarket US vs Kalshi), whether the venue is geo-restricted for the audience, and the volume behind the price. A 90c price on $200 of volume is noise. See references/manipulation-tells.md for the wash-trading checklist.
4. Assess the resolution path. For Polymarket: a proposer stakes a bond and asserts an outcome, a ~2 hour dispute window opens, and disputed markets go to a UMA token-holder vote taking 48-96 hours. Disputes are rare (~1.3%) but a disputed resolution is a vote, not a fact. See references/resolution-and-disputes.md.
5. Apply calibration knowledge. Markets are well calibrated near resolution but compressed toward 50% at long horizons. Adjust for that before quoting a long-dated price as someone's belief. See references/calibration-and-biases.md.
6. Score your own forecasts. If the agent records probability estimates over time, run bin/brier.py on a CSV of (forecast, outcome) rows to get a Brier score, the always-50% baseline, and a calibration table. Under ~100 resolved forecasts, the numbers are noise. Do not over-interpret them.

Output Contract

A market read reports:

• the exact proposition as written in the resolution rules, plus the named resolution source and cutoff
• implied probability from the book mid, with spread width and volume as a confidence note
• any rule quirks that change what the price means (announcement vs event, revision handling, 50/50 tie clauses)
• manipulation flags from the tells checklist, or an explicit "no flags found"
• for multi-market or cross-venue comparisons: which pairs resolve on the same condition and which do not (non-comparable pairs get named, not silently merged)

Operating Rules

1. The resolution text is the contract. Never quote a price without the rule it settles under.
2. Never compare two markets that resolve differently and call it a disagreement.
3. Volume is not demand. Check for wash-trading tells (post-decision volume spikes, long-shots out-trading favorites) before citing volume as conviction.
4. A resolved-but-disputed market is unsettled. Only finalized markets count as ground truth.
5. This skill is read-only research. It does not place trades, hold positions, or recommend bets. If a workflow drifts toward execution, stop and ask.
6. No unverified numbers. Every figure in a market read traces to the market page, a study, or a named source, or it gets cut.