On Tuesday, February 10, 2026, at 6:47 PM ET, the Fed released minutes from its January meeting with language that was slightly more hawkish than expected. Within three minutes, "Will the Fed cut rates in March?" contracts on Kalshi dropped 4 cents. Polymarket's equivalent moved 3.2 cents in the same direction.
Most traders found out about the minutes from a push notification on their phone. By the time they opened a browser, checked the relevant contracts, and decided whether to act, the initial move was done.
Previa's AI pipeline flagged the story, mapped it to fourteen active markets across both platforms, and pushed alerts to users watching those contracts — all within 90 seconds of the release.
Here's how that works.
Step 1: Ingestion
Previa monitors over 50 news sources continuously. Not on a schedule — continuously. We pull from AP, Reuters, Bloomberg headlines, major newspaper feeds, CoinDesk, ESPN, Politico, and a handful of specialty sources that cover the categories prediction markets care about: politics, economics, crypto, sports, culture, and science.
We also watch Twitter/X for high-signal accounts. Not the noise — the accounts that, based on historical accuracy, tend to surface information before it hits mainstream outlets.
Every piece of content passes through a deduplication layer. If Reuters and AP both run the same Fed minutes story (they always do), we process it once.
Step 2: Entity extraction and relevance scoring
This is where the AI earns its keep.
For each new piece of content, our pipeline extracts entities (people, organizations, events, policy actions), determines the core claim or development, and then asks: which active prediction market contracts does this plausibly affect?
This isn't keyword matching. "Fed" doesn't just map to every interest rate contract. The AI reads the substance of the article — what did the Fed actually say, and what does it imply about the March meeting specifically? — and scores relevance against each contract's specific resolution criteria.
We use a combination of LLM analysis and semantic similarity between article embeddings and market question embeddings. The LLM handles nuance and reasoning; the embeddings handle breadth and speed.
Step 3: Impact scoring
Relevance isn't enough. A story can be relevant to a market but low-impact (a routine poll showing no change) or highly impactful (a key endorsement in a tight race).
For each article-to-market mapping, Previa generates an impact score from 0-100 and a directional signal: bullish, bearish, or neutral for YES contracts. The impact score considers:
- Novelty: Is this genuinely new information, or a rehash?
- Source authority: AP newswire vs. a blog post vs. an anonymous tweet
- Magnitude: Does this shift the underlying probability meaningfully?
- Proximity to resolution: A strong signal matters more when the market closes in 48 hours than when it closes in six months
The AI also generates a brief reasoning paragraph — two to three sentences explaining why this article affects this market and in which direction. These aren't vibes. They cite specific facts from the article and specific contract terms.
What this enables
The news-to-market mapping pipeline feeds three user-facing features:
Confidence Scores. News signals are one of the four inputs to Previa's composite confidence score for each market. When a high-impact article maps to a contract, the score updates in real time.
AI Insights. When a mapping exceeds our impact threshold, Previa generates an insight card that appears in users' feeds — a plain-English summary of what happened, which markets are affected, and what the AI thinks the implications are.
Smart Alerts. Users who have "news catalyst" alerts enabled on their watchlist or portfolio positions receive push notifications within minutes of a high-impact mapping. The alert includes the article, the affected markets, and the AI's directional assessment.
The honest limitations
We're not pretending this is perfect. No AI system reads every nuance of geopolitical news correctly. Our accuracy rate on directional signals is good but not infallible. We show our reasoning specifically so users can evaluate it themselves — this is decision support, not decision making.
The speed advantage is real, though. When a story breaks, the difference between finding out in two minutes versus twenty can mean the difference between entering a position before the crowd and chasing a price that's already moved.
Under the hood
For the technically curious: the pipeline runs on a combination of Anthropic's Claude for analysis and reasoning, OpenAI embeddings for semantic similarity, and a custom scoring model we've trained on historical news-to-market-movement data. The whole thing is designed to run at ingestion speed — we process a new article from first byte to mapped-and-scored in under 30 seconds.
We'll be writing more about the technical architecture in upcoming posts. If you're building in the prediction market space and want to dig into how we approach these problems, follow along.