Spreadwire

Methodology

Spreadwire finds sharp election-market moves, checks available headline evidence and records an AI assessment. It preserves the original assessment, its inputs and what happened afterwards. This is not a guarantee of causation, prediction accuracy or investment performance.

What we track

Polymarket’s Gamma API resolves markets and its CLOB WebSocket supplies live last trades and tight-book midpoints. A midpoint is shown only when the book spread is at most 10 cents. A last trade and a midpoint are different measurements, not interchangeable evidence.

How to read the two venues

Spreadwire watches two markets and never mixes them. Polymarket is the detection venue: the deepest 24/7 book on these races, sampled continuously, so a sharp move shows up there first and our engine catches it within minutes. Kalshi, a CFTC-regulated exchange, is the independent second reading taken at the moment of each move: if Kalshi moved too, the move has cross-venue weight; if it did not, we say that too.

Each card names the venue, the actual percentage, the quote type, the observation time and that market’s exact question. Each percentage is an implied chance of a YES outcome in that specific market. A last trade is an executed price; a midpoint sits between the available buy and sell quotes. The two venues get equal visual prominence, not presumed equal meaning. Gaps in the data stay gaps: nothing is averaged, nothing is backfilled, and a missing reading is never treated as zero.

Audit details: data sources, gates and equivalence reviews

Polymarket’s Gamma API resolves markets and its CLOB WebSocket supplies live last trades and tight-book midpoints. A midpoint is shown only when the book spread is at most 10 cents.

Current Kalshi observations come from a candidate-scoped, read-only public API after exact ticker, event, candidate-label and active-status checks. Streaming helpers remain isolated from the engine and outbound alerts. An older observation stays labelled stale; stale or missing input never becomes zero. Market and trade identifiers remain in the underlying audit data and comparison API, rather than cluttering the reader-facing cards.

There is no automatic claim that two contracts are equivalent. A calculated difference additionally requires explicit owner review of the outcome/resolution terms, an exact pairing of both current market IDs, matching quote types and fresh compatible timestamps. Revocation or a changed market invalidates that comparison. The history graph currently focuses on the actual Polymarket move. No Kalshi historical line is manufactured from one quote.

How a move is detected

The exact windows, thresholds and cooldown values are part of the engine’s calibration. They are applied identically to every candidate and kept in the audit data rather than listed here, so they cannot be tuned around.

The 90-second confirmation

A currently detected move is revalidated against actual Polymarket inputs about 90 seconds later. A failed check becomes reverted, not a public alert. A held low-severity move can be explained on demand. Medium/high moves enter the ordinary narrator path. Original trigger and confirmation timestamps, sample counts and failed outcomes are retained. Reversion is not proof that the first detection was a bad forecast.

Severity & verdicts

What the AI sees

Explanations are generated by an LLM chain grounded with: Google News RSS and GDELT headlines (12h/4h windows), general press RSS headlines, every other tracked candidate's price and recent move, Wikipedia pageview spikes, and a 7-day price baseline. Prompts include available headline titles, sources and URLs; they do not contain full articles. Text is AI-generated and a listed source is captured news context, not proof of causation or that the full article was read. If no provider can produce a usable analysis, the event remains without one rather than receiving invented text.

Read a worked case or inspect observed follow-through. Later price observations describe what happened after a move; they do not grade whether an explanation was correct.

Independence & source attribution

Spreadwire takes no affiliate or referral money from any trading venue. We don't care where anyone trades — only what a move means. Revenue is subscriptions, period.

Price data: publicly observable prediction-market prices. Spreadwire is independent and is not affiliated with, endorsed by, or sponsored by any prediction-market platform. Platform price data remain subject to their respective terms of use.

Citation & machine access

Spreadwire verdicts may be cited with attribution and a link. Every narration has machine-readable versions: /n/{id}.md (Markdown) and /n/{id}.json (JSON), plus a chart image at /og/{id}.png.

Spreadwire analyses publicly observable prediction-market data. These platforms are not authorised in France; Spreadwire does not promote, link to or recommend them. Political-risk analysis, not investment advice.