Check odds

Check Odds

Get current prediction-market prices for a question. Four platforms are one no-auth curl each and Metaculus needs a free token — run them all in parallel (parallel Bash calls inline; no need for one subagent per platform). All endpoints below were live-verified 2026-07-08; if one 4xx/5xxs, note it in output rather than silently skipping.

Search with short keywords, not the full question. Derive 1–3 keyword variants from the framed question (entity + event core: "Fed December", "FOMC cut" — not "Will the Fed cut rates at the December 2026 FOMC meeting?"); most venues' fuzzy search returns zero hits on full sentences. Try the variants before declaring "no markets found".

This skill produces three things: p_market (the synthesized market prior), the synthesis depth (deep / shallow / none — how much real, independent information the venues carry), and the per-platform table (cross-venue divergences are themselves signal). Standalone, this is the fast path to "what are the odds?"; the superforecaster skill consumes all three outputs as its anchor — depth sets how strong the evidence must be to move off it, and how much research the question deserves.

Market structures — synthesize, don't rank dogmatically

Each venue is an incentive structure with characteristic strengths and distortions. Notably, the best-calibrated venue has no money on it at all — so "real money beats play money" is not a law. Weight venues by track record and structure, and read disagreements structurally:

Platform	Structure	Strengths	Characteristic distortions
Metaculus	Reputation + scored calibration	Best public track record; superforecaster norms	Slow to update on fast news; thin on pop culture
Manifold	Play-money economy (designed incentives)	Unmatched long-tail coverage; fast	No stakes on exotic questions; whale/meme distortion — judge each market on its crowd, not the currency
Kalshi, Polymarket	Real money, CLOB	Deepest books; skin-in-the-game; fast on news	Time-value: nobody corrects long-dated tail odds priced below the risk-free return, so rare-event prices sit too high; retail/narrative flow; manipulation incentive (below)
PredictIt	Real money, capped	US-politics crowd wisdom	$3,500 caps + 10%+5% fees structurally distort prices
Good Judgment Open	Reputation	Superb calibration	No API, not searchable — reference only

Two real-money distortions deserve explicit checks:

• Time-value floor. A trader who thinks a 5¢ year-out contract is worth 1¢ earns ~4% locking capital for a year — the risk-free rate. Long-dated tails on real-money books are systematically un-corrected; discount them, and lean on Metaculus/Manifold for rare-event odds.
• Manipulation. Real payoffs create incentives to move the settlement source, not just the price (Kalshi's June 2026 Spotify market settled on ~500k bot streams that cost the manipulator a few hundred dollars against a ~$3M market). When the settlement metric is cheap to move — streams, follower/post counts, a single sensor — a late one-sided price move near resolution reads as manipulation intent unless primary data explains it. This is a weight and a depth cap, not just a flag: such venues take the step-4 haircut below and can never make the synthesis deep by themselves.

Criteria validation — does this market measure the user's event? The anchor is only as good as this check. Before a venue enters the synthesis, read its actual resolution rules (contract terms / rules page — titles lie) and confirm it resolves on the event being modeled, not a correlated proxy: a "game delayed" market is not a "game rained out" market — delay bundles every cause, and the umbrella decision rides on rain alone. Per venue, one verdict: match (enters synthesis) · proxy (evidence only — name how it differs in caveats, expect divergence, never anchor on it) · different event (dropped, named in output). If no match survives, that IS the answer — p_market: none, however loudly the proxies quote a number.

Synthesis → p_market (mechanical recipe — the evidence-weighted median):

1. Drop venues that failed criteria validation above (a proxy informs caveats, not the number).
2. Convert each surviving venue to a midpoint probability on the 0–1 scale (Kalshi quotes integer cents — divide by 100; everyone else already returns 0–1).
3. Assign weights: revealed Metaculus CP (≥~20 forecasters) 4 · real-money book ≥$100k 2 · thin real-money 1 · play-money 1 · PredictIt 0.5.
4. Halve any venue that is visibly stale (no trades/updates in days on a moving story), priced below ~5¢ on a long-dated contract (time-value floor), or settles on a cheap-to-move metric (manipulation, above).
5. p_market = weighted median: sort by price, take the first price where cumulative weight reaches half the total.
6. Before accepting an outlier's exclusion or a wide spread, explain it from structure (criteria difference, fees, time-value, manipulability, stale crowd) — an explained divergence is an adjustment; only unexplained divergence is candidate mispricing. Cross-structure agreement (reputation + play-money + real-money within a few points) is the strongest market signal there is; note it.

Synthesis depth:

• Deep: (revealed Metaculus CP with forecaster_count ≥ ~20 or a ≥$100k real-money book) and corroborated within ~15 points by at least one other structure — both conditions, whichever venue leads.
• Shallow: anything less — one thin/uncorroborated venue, or venues still >15 points apart after structural explanation (flag it — usually a criteria mismatch, occasionally real mispricing).
• None: no market whose criteria match. "No markets found" is informative — say it explicitly per platform.

Venues change. New markets and aggregators appear (and die) regularly: when you discover one, live-verify its API and consider adding it to the table. If a valuable data source needs credentials or account setup, prompt the user to connect it rather than working around the gap — reliable data feeds are the system's foundation.

Access & auth

1. Metaculus — token required (free: metaculus.com account settings → API token)

curl -s -H "Authorization: Token $METACULUS_TOKEN" \
  "https://www.metaculus.com/api/posts/?search=QUERY&limit=5&statuses=open&forecast_type=binary&with_cp=true"

Use /api/posts/ — the older /api2/questions/?search= ignores search terms (verified 2026-07-08). Per result: .results[].title, CP = .results[].question.aggregations.recency_weighted.latest.centers[0] (if null, try the key named by .question.default_aggregation_method instead of recency_weighted), and forecaster count from ...latest.forecaster_count or the post's forecasts_count.

CP is gated by account API tier — nothing client-side fixes it (verified 2026-07-09): standard accounts show api_access_tier: "restricted" at /api/users/me/ and get all-null aggregations on every question — open, closed, even ones that account has forecast — while search, titles, and statuses work fine. Anonymous API calls are Cloudflare-blocked. Bot accounts do not help: a fresh bot token (is_bot: true) is equally restricted (verified) — bots exist to submit forecasts, not read the crowd. The bulk export /api/data/download/ hits the same production gate (live-tested; the permissive logic visible in Metaculus's open-source repo is not what production enforces). Tiers, from their source + API docs: restricted → benchmarking (CP on a few hundred benchmark questions, meant for bot training/eval) → unrestricted (written agreement). The sanctioned path up: the Metaculus Data Needs Form and/or [email protected], disclosing the use — their Terms restrict using CP to train or evaluate AI/ML models without written permission, and anchoring an AI forecaster to CP is exactly that, so ask rather than work around. Until granted: report "Metaculus: CP unavailable via API (restricted tier)" loudly, pull CP from WebSearch site:metaculus.com QUERY snippets (they routinely quote it) as table/caveat evidence only — snippet CP never enters the weighted median — and treat the synthesis as missing its strongest venue — degraded coverage, never silently skipped. If the token is missing entirely: mark Metaculus unavailable, continue with the other venues, and ask the user to add a token after delivering the result.

2. Manifold — no auth

curl -s "https://api.manifold.markets/v0/search-markets?term=QUERY&sort=liquidity&filter=open&limit=5"

Extract: question, probability (0–1), url, volume, uniqueBettorCount. Unmatched long-tail coverage.

3a. Polymarket — no auth

curl -s "https://gamma-api.polymarket.com/public-search?q=QUERY&events_status=active&limit_per_type=5"

.events[] | {title, slug, volume, liquidity,
             markets: [.markets[] | {question, outcomes: (.outcomes|fromjson),
                                     prices: (.outcomePrices|fromjson), volume}]}

outcomePrices/outcomes are stringified JSON — pipe through fromjson. Prices are 0–1. Event URL: https://polymarket.com/event/<slug>. URL-encode spaces. Do NOT use /events?title_contains= (unranked, includes closed markets).

3b. Kalshi — no auth

curl -s "https://api.elections.kalshi.com/v1/search/series?query=KEYWORD"

.current_page[] | {event_title, event_ticker,
                   markets: [.markets[] | {ticker, yes_bid, yes_ask, last_price, volume}]}

Integer prices are cents (yes_bid: 9 = 9%); *_dollars string fields exist in v2. Use short keywords — the fuzzy ranker degrades on long queries. api.elections.kalshi.com is canonical (api.kalshi.com does not resolve). Full catalog fallback: trade-api/v2/events?with_nested_markets=true&status=open (cursor-paginated).

4. PredictIt — no auth, US politics only

curl -s "https://www.predictit.org/api/marketdata/all/"

Returns all ~200 markets; no server-side search — filter client-side by whole-word match on name/shortName ("Maine" contains "ai"). Extract contracts[].lastTradePrice (0–1); keys are camelCase (live-verified 2026-07-08). Rate limit ~1 req/min. Skip for non-political questions.

Dropped (do not re-add without re-verifying): Betfair (auth-walled), Smarkets (no search), Adjacent News (dead) — all verified unusable 2026-07-06.

Output format

## Prediction Market Odds: <topic>

| Platform | Question | Probability | Liquidity | Criteria | Link |
|----------|----------|-------------|-----------|----------|------|
| Metaculus | ... | 11% | 150 forecasters | match | [link] |
| Manifold | ... | 11.9% | 95k mana | match | [link] |
| Polymarket | ... | 10.5% | $1.65M vol | match | [link] |
| Kalshi | ... | 9% | 8.3M contracts | match | [link] |
| PredictIt | (not political — skipped) | | | | |

**p_market: 0.11 (evidence-weighted median; Metaculus CP 0.11 leads, Polymarket 0.105 corroborates)** — depth: deep
Cross-structure spread: 9–12% — reputation, play-money, and real-money venues agree.
Caveats: <thin books, hidden predictions, criteria mismatches, time-value on long-dated tails, manipulable settlement metric>

No matching markets anywhere: the line reads p_market: none — depth: none (no number — don't invent one). Downstream, the base rate becomes the prior, and a ledger row logs without --market-p/--market-source.

Don't editorialize — prices, liquidity, caveats. Everything else this output needs (parallel calls, criteria matching, divergence explanation, per-platform "no market") is specified above.