Models · NLP / Investing
Sentiment from Filings
Extracts forward-looking sentiment, risk language, and management tone from 10-K, 10-Q, 8-K, and equivalent EU/UK filings.
- Risk: low
- GA
- US
- UK
- EU
- F1 (binary tone)
- 0.91
- Backtest IR (2018-2025)
- 0.78
- Median p99 latency
- 120 ms
- Calls served
- 12.4 M
What it does
sentiment-from-filings ingests the full text of a regulatory filing and
returns:
- A scalar sentiment score (
-1.0to+1.0) for the document and per section. - A list of detected risk-language phrases with their location offsets.
- A management-tone vector (confidence, hedging, novelty) for use as a feature in downstream models.
Where it shines
Long-only equity funds running event-driven strategies on earnings releases and 8-Ks. Backtests on 2018–2025 US large-cap show a 0.78 information ratio when used as a single signal in a market-neutral overlay; full methodology in our research notes.
Where to be careful
Sentiment models are reflexive. Crowded usage erodes the signal. Don’t size positions purely on this output — combine with at least one orthogonal signal.
How it’s trained
Trained on 25 years of EDGAR filings + Companies House filings, fine-tuned on a hand-labelled tone corpus. Re-trained quarterly; all training-data provenance hashes are published in the audit log.
Pricing
Included in Starter and above for up to plan-tier inference calls; metered overage at $0.001 / call.
Audit & jurisdiction
- US — reviewed and routed to local counsel.
- UK — reviewed and routed to local counsel.
- EU — reviewed and routed to local counsel.
Every run of this model is appended to a hash-chained audit log. Anyone with a run id can fetch a Merkle inclusion proof — see the Compliance & Trust page.
Reported metrics are self-attested by the model author and verifiable against the audit log.