One fake headline can move a herd of trading agents, so signed news gains value

The space is early. I found no commercial product that signs or provenance-scores financial headlines specifically so trading agents can check them before acting. I also found no vendor that sells stress tests injecting synthetic news to measure how far a firm's agents herd. What does exist falls into four groups. Incumbent news vendors like RavenPack and Dataminr compete on speed, novelty and sentiment rather than verified origin. General-purpose claim-checking APIs for agents, such as Webcite and Parallel, are neither finance-specific nor fast enough for trading. Signing standards like C2PA and IPTC exist, but Reuters has so far applied them only to images and video. The academic work, including AutoRedTrader, MarketSim and the ECB synthetic-trader study, shows the risk is real but has not been turned into products. For the feed, a new entrant could win by sitting between the wires, social media and agent frameworks. It would check signatures where publishers provide them, fall back to cross-confirmation across independent sources where they don't, and return a score fast enough for trading. The biggest threat is that Bloomberg, LSEG/Reuters, Dow Jones or RavenPack add a provenance field to feeds funds already pay for. Wire services that own the signing keys hold a structural advantage, so a startup has a better chance as a neutral aggregator across sources (social, regulatory filings, smaller outlets) than as a replacement for the wires. The stress test is the emptier and more defensible of the two ideas. Nobody is selling correlated-reaction testing of LLM agents to broker risk desks or exchange surveillance teams, and growing regulatory attention to AI in trading could create compliance demand. The main risk is that the buyers are few and slow, and the work could collapse into consulting or be absorbed by surveillance vendors like Nasdaq SMARTS or by agent-based simulation firms.

• AutoRedTrader (arXiv 2605.09185)

  This is an academic paper, not a product. It red-teams LLM trading agents by injecting synthetic misinformation and reports a 26.67% attack success rate. It is the closest match to the herding stress test idea, but it tests a single agent on Bitcoin data and does not measure how correlated many agents' trades become.

• Dataminr

  A long-established real-time event detection vendor sold to finance clients. In the 2013 hacked AP tweet case it said it flagged the hoax two minutes early. It competes on speed and significance filtering. It does not offer cryptographic provenance or per-headline confirmation scores built for agents.

• RavenPack

  An incumbent machine-readable news analytics vendor for quant funds. It scores relevance, novelty and sentiment for each story. That makes it the natural owner of a provenance-score field if it chose to add one, but none of the sources I found say it judges whether a story is true.

• IPTC Verified News Publisher / C2PA

  These are open standards for signing publisher content. Reuters is piloting them with Canon and Sony, but so far only for photos and video. Nothing I found applies them to text headlines or financial wire feeds. This is the signing layer a provenance feed would build on, not a competitor.

• Webcite

  A general-purpose API that checks claims for AI agents and scores how credible each source is before the agent acts. It is not built for finance and does not run at the latency that market-moving headlines need.

• Parallel

  Web search and extraction infrastructure for agents that attaches citations and calibrated confidence scores to each fact. It overlaps with cross-confirmation scoring but is not a news feed for traders.

• Provenance AI

  A small startup founded in 2025 that sells fact-checking with blockchain attestations. It targets social media claims and deepfakes, not financial feeds, and its marketing reads as promotional.

• Lopez-Lira LLM market simulator (arXiv 2504.10789)

  An open-source order-book simulator populated by LLM trader agents. Like MarketSim and the ECB synthetic-trader study, it could be the base for a herding stress test. The ECB paper found strongly correlated forecasts even with naive prompting, but none of these projects is sold as a service to surveillance desks.

Checked 2026-10-20