The model that judges every headline.
One pass. Actionability and sentiment, scored together. Trained on a million-plus headlines labelled with their realized price impact — not generic sentiment, not crowd-sourced tags.
[01]The output
One verdict per event. A score you can route, threshold, or feed straight into a strategy. No post-processing, no second model.
Actionability + sentiment, single pass.
Most NLP scoring pipelines run two models: sentiment first, then a heuristic for whether the sentiment matters. We collapse the two. The Engine emits one continuous score that says how much this should move the underlying, not how positive the wording sounds.
- Verdict score (-1.0 → +1.0), calibrated to realized 1-hour return.
- Confidence band, so you can drop low-conviction events early.
- Per-event provenance: source, timestamp, model version.
NVIDIA beats Q3 estimates, raises Q4 guidance on data-center demand.
$NVDA · Globe Newswire · 2026-05-26 16:01:03 UTC
"verdict": +0.83,
"confidence": 0.91,
"ticker": "NVDA",
"horizon": "1h",
"latency_ms": 28
}
[02]The labels
Labelled by the market, not by humans. Every training headline carries its own ground truth: the realized price move on the underlying in the hour after it dropped.
Headlines in the training set, growing weekly.
Realized return window used as label, matched per-ticker.
No crowd-sourced sentiment. No analyst tags. Just price moves.
Why not generic sentiment?
Generic sentiment models score language, not market reaction. A neutral-sounding 8-K can move a stock 12%. A glowing press release can flatline. We label on outcomes, so the model learns what the market actually cares about.
Why retrain weekly?
Markets re-price what matters. A guidance beat in 2022 meant something different than in 2026. Continuous retraining on rolling windows keeps the Engine calibrated to the current regime.
[03]Performance
Verdicts at wire speed. Headline to scored verdict in under 50 ms, end-to-end. The model doesn't sit between you and the market — it runs alongside it.
P95 end-to-end latency, source to subscriber.
Model inference, per verdict.
Verdicts per second, sustained on a single node.
Inference availability, trailing 90 days.
[04]Integrations
However you consume it. Stream it, query it, or hand it to an agent. Same verdict, three surfaces.
WebSocket feed
Subscribe to verdicts by ticker, sector, or score threshold. Reconnect-safe, ordered, replayable.
REST + SDKs
Score arbitrary text on demand. Python, Node, and a typed CLI. Bring-your-own corpus.
MCP server
Drop the Engine into any AI agent stack. The model answers questions like the user just asked it.
Start scoring headlines today.
Free tier includes 50 realtime feeds a day. No credit card required.