StockTitan methodology
How Rhea-AI scores market news
Last reviewed
Rhea-AI is StockTitan's proprietary news analysis engine. It reads every press release and SEC filing published on StockTitan, roughly 70,000 documents a month, and assigns two independent scores within milliseconds of publication:
- Impact: the potential for short-term price volatility
- Sentiment: the tone of the language
Few analytics products publish how well their scores hold up against the market. This page does exactly that: definitions, live calibration, verified accuracy, and limitations.
The two scores
Sentiment measures tone. Impact measures magnitude.
Each score answers a different question, and they move independently. A press release can be neutral in tone yet carry very high impact potential, and enthusiastic language can accompany news that rarely moves a stock. Reading the two together is what makes the feed scannable.
Rhea-AI Sentiment
Is the language of this document positive or negative?
Multi-layered natural language analysis of the full text, combining word-level and contextual reading. The engine understands sector-specific terminology: clinical vocabulary in a biotech release is not scored the way the same words would be in a retail earnings report.
- 5Very Positive. Language dominated by positive terms: growth, breakthrough, record results.
- 4Positive. Mostly positive language with some neutral terms.
- 3Neutral. Balanced or predominantly factual language.
- 2Negative. Mostly negative language with some neutral terms.
- 1Very Negative. Language dominated by negative terms: crisis, default, failure.
Rhea-AI Impact
How likely is this news to move the stock, in either direction?
An estimate of the potential for short-term price volatility, based on the type and significance of the event described. Impact is direction-blind by design: a major clinical failure and a major acquisition can both score 5, because both tend to produce large moves.
- 5Very High. Major M&A, pivotal trial readouts, bankruptcy, transformative deals.
- 4High. Significant earnings results, major partnerships, regulatory actions.
- 3Moderate. Analyst rating changes, product launches, executive changes.
- 2Neutral. Minor updates, small contracts, routine announcements.
- 1Low. Regular disclosures and administrative filings.
For thinly traded stocks, the stock's own volatility can raise the score: the same announcement carries more movement potential on a stock that already trades violently.
Is the tone positive or negative?
Question answeredCould this move the price?
Directional: positive to negative
DirectionDirection-blind: magnitude only
The document itself
HorizonShort-term volatility potential, about one trading session
Overwhelmingly positive language
A score of 5 meansAn event class that historically produces the largest moves
Inputs
What goes in, and what stays out
Rhea-AI scores the document in front of it, nothing else. That constraint is deliberate: it keeps every score neutral, reproducible, and comparable across 70,000 documents a month, from mega-cap earnings to the small-cap filings no other analytics desk covers.
+ Considered by the scoring engine
- The full text of the specific news release or SEC filing
- The class of event described: financing, trial readout, M&A, earnings, corporate action
- Financial figures, quantities, and magnitudes stated in the document
- Sector-specific terminology and its context
- Structural signals of the document type
− Excluded by design
- Macroeconomic conditions and broader market moves
- The company's prior news history or reputation
- Social media chatter and analyst opinions
- Price action before or after publication
- Any judgment on whether the claims in the document are true
A score that mixed in market mood or company reputation would drift with hype and compound bias.
By scoring only the document, two identical announcements from two different companies receive the same treatment, and readers can compare scores across the entire feed.
Live calibration
What an impact score means in practice
We hold our scores to the only standard that matters: what the market actually did next. The data below covers every scored news article from January 1 to June 30, 2026: 78,539 articles, each matched to the close-to-close reaction of its primary ticker over the first full trading session after publication. Moves are absolute values, so direction is ignored.
| Impact level | Articles | Avg. absolute move | Moved > 10% | Moved > 20% | Moved > 30% |
|---|---|---|---|---|---|
| 1 · Low | 3,319 | 3.1% | 5% | 1.7% | 0.7% |
| 2 · Neutral | 32,003 | 3.3% | 5% | 1.4% | 0.7% |
| 3 · Moderate | 24,261 | 5.6% | 13% | 4.1% | 1.9% |
| 4 · High | 17,175 | 10.2% | 25% | 10.0% | 5.0% |
| 5 · Very High | 1,781 | 16.0% | 44% | 23.5% | 13.5% |
Calibration is an association, not a promise: 38% of Impact 5 articles in the same sample were followed by a move of 5% or less, and realized moves also depend on factors outside the news, such as liquidity and float.
Past calibration does not guarantee future results.
Reading the tone
What a sentiment score means in practice
Sentiment answers one question: how positive or negative is the language of this document? Every word carries weight in the reading. A release built around debt, defaults, and bankruptcy reads low; one built around record results and successful launches reads high. The score describes the text itself, never the stock.
Same engine, different words · illustrative excerpts
“…announced a restructuring of its outstanding debt, citing substantial doubt about its ability to continue as a going concern and the possibility of bankruptcy protection…”
“…reported record quarterly revenue, driven by the successful launch of its flagship platform and accelerating growth across every segment…”
Highlighted terms illustrate how wording pulls a score down or up. The engine weighs the full text, and the same word can weigh differently depending on sector and context; the exact vocabulary and weights are proprietary and are not published.
+ What the score tells you
- How positive or negative the language of the document is, on the 1–5 scale
- Which way the wording leans: growth, breakthrough, and record results pull it up; debt, default, and bankruptcy pull it down
- Tone read in context: sector-specific terms are weighed by what they mean in that industry
- A reading computed the same way for every document, so tone is comparable across the entire feed
− What it does not tell you
- Anything about stock performance: sentiment does not predict the direction or size of any price move
- How the market will react: the score is computed from the text alone, before and independently of any reaction
- Whether the claims are true: upbeat language scores as upbeat language regardless of substance
- How significant the event is: that is Impact's job, and the two scores move independently
Press releases are written by the companies they describe, and framing matters: most issuers present even routine news in the best possible light.
Expect the sentiment distribution to lean toward the positive half of the scale, and read scores against that baseline: a Neutral on a company-authored release is often closer to lukewarm, and a Negative one means the language could not be framed upward.
Engine accuracy
The 2026 impact engine, measured before you see it
The newest generation of the impact engine was verified on a full month of market events it had never seen, in strict chronological order, before being cleared for production. These are its measured results on that held-out month.
Verification sample: 2,766 market events from a held-out month, excluded from all training and tuning, evaluated in July 2026. A separate materiality signal, flagging news likely to move a stock at least twice its normal volatility, tested at 2.9 times chance on the same sample.
Validated cases
Impact 5 in the wild
Three articles from 2026, scored at publication, with the market reaction measured afterward by StockTitan's own tracking. Each links to the original article, where the score and the measured reaction are still visible.
KORE announces agreement to be acquired by Searchlight Capital Partners and Abry Partners
Innodata reports record first quarter 2026 results
Quince Therapeutics announces topline results from pivotal Phase 3 NEAT clinical trial
The QNCX case shows why impact is direction-blind: a pivotal trial readout is a maximum-impact event whichever way it resolves. Impact told readers this document mattered before the market decided which way it mattered.
Validation
How we test a scoring model before you see it
Every generation of the scoring engine goes through the same discipline before it is allowed to score live news. No exceptions, including the current one.
Historical grounding
Models are developed against StockTitan's own archive, which now holds over a million scored news items. The current engine was built from 953,080 news events paired with the market reaction that followed them, going back to April 2020 and spanning NYSE, Nasdaq, and OTC listings.
Out-of-sample testing
Candidates are evaluated in strict chronological order on events they have never seen, so results cannot benefit from hindsight.
Shadow deployment
Before promotion, a candidate scores live news in parallel with the production engine. Only the incumbent's scores are shown to readers during this period.
Promotion and monitoring
A candidate replaces the incumbent only if it performs better on data neither has seen. After promotion, calibration is monitored continuously.
The exact features, weights, and architecture of the scoring engine are proprietary and are not published.
What we publish is everything needed to judge the scores from the outside: their definitions, their inputs and exclusions, their measured calibration against real market reactions, and their limitations.
Material changes to the engine are announced in the version history.
Limitations
What Rhea-AI cannot tell you
Reading the scores well means knowing their boundaries. These are the ones that matter.
- Not investment advice. Scores are statistical indicators built to speed up research, not recommendations to buy or sell.
- No claim verification. Rhea-AI scores what a document says, not whether it is true. A confident press release from an unreliable issuer can still score well.
- Content-only scope. The engine does not know the macro backdrop, the company's history, or what the market already expected.
- High score, quiet stock. 38% of Impact 5 articles in H1 2026 moved 5% or less. Low scores do not rule out surprises either.
- Size effects. Small-cap and low-float stocks move more on comparable news. Calibration averages blend all listings.
- AI-generated text can err. Summaries and takes are generated automatically and can contain mistakes. The original source document is always one click away.
Accountability
Who is behind Rhea-AI
Rhea-AI is built and operated entirely in-house by the StockTitan engineering and editorial team. The scoring engine was developed with guidance from financial industry professionals, and its output is reviewed continuously against professional standards.
All AI-generated content on StockTitan is labeled as such, kept visually separate from source documents, and accompanied by its limitations. When the engine changes in a way that affects what readers see, the change is recorded in the version history on this page.
- About StockTitan · who we are and how the platform works
- Editorial policy · standards for accuracy, review, and corrections
- Version history · every material change to Rhea-AI since 2020
FAQ
Common questions
What is the difference between Rhea-AI Sentiment and Impact?
Sentiment measures the tone of the language in a document, from Very Negative to Very Positive.
Impact estimates the potential of the described event to move the stock price in the short term, regardless of direction.
They are independent: a neutral-tone document can carry Very High impact.
Does the Rhea-AI Sentiment score predict stock performance?
No. Sentiment measures only the tone of the language in a document: wording about debt, defaults, or bankruptcy reads low, wording about record results and successful launches reads high.
It is computed from the text alone, before and independently of any market reaction, and says nothing about how the stock will perform.
Because press releases are written by the companies themselves, tone tends to lean optimistic, so scores should be read against that baseline.
What does a Rhea-AI Impact score of 5 mean in practice?
In the first half of 2026, news rated Impact 5 was followed by an average absolute move of 16.0% in the primary ticker over the next trading session. Nearly one in four Impact 5 articles moved more than 20%, and more than one in eight moved over 30%.
It is a strong historical association, not a guarantee: 38% of Impact 5 articles moved 5% or less.
How accurate is the Rhea-AI impact engine?
On a held-out month of 2,766 market events never used in training, the current engine matched the realized volatility outcome exactly in 57.4% of cases across five levels and was within one level 92.4% of the time.
Its extreme-move alert identified news preceding moves of 3.5 times a stock's normal volatility 5.4 times better than chance.
What data does Rhea-AI use to produce scores and summaries?
Scores and summaries are based solely on the text of each individual news article or SEC filing.
To remain neutral and unbiased, the scoring does not consider external factors such as macroeconomic data, broader market moves, company history, or other context.
Does a high score guarantee accuracy or predict stock movement?
No. A high score does not validate claim accuracy, assess company credibility, or guarantee that a stock will move.
Rhea-AI outputs are algorithmic, statistical indicators derived from document content, calibrated against historical market reactions.
How is Rhea-AI validated?
Every model generation is tested chronologically on historical events it has never seen, then runs in shadow alongside the production engine on live news. It is promoted only if it outperforms on unseen data.
After promotion, calibration against measured market reactions is monitored continuously and published on this page.
Should I make decisions based solely on Rhea-AI?
No. Rhea-AI is a research aid, not financial advice.
Always consider company fundamentals, broader market conditions, and your personal risk tolerance, and consult a licensed advisor where appropriate, before trading or investing.
Update cadence
Version history
Every material change to the engine since launch. Continuous monitoring runs between releases.
- Complete rebuild of the impact engine, verified at 57.4% exact and 92.4% within-one-level accuracy on a held-out month before promotion
- New extreme-move alert and materiality signal, detecting outsized movers at 5.4x and 2.9x chance respectively
- Scoring latency reduced to milliseconds per article at publication
- Overhaul of the sentiment engine with context-aware scoring of industry-specific language
- Greatly improved accuracy for biotech, pharmaceutical, and healthcare news
- Smarter recognition of corporate actions such as offerings, reverse splits, and regulatory decisions
See the scores on live market news
Every article on StockTitan carries its Rhea-AI scores the moment it is published, next to the original source document.
Open the live news feed About StockTitanRhea-AI outputs are automated, statistical indicators provided for informational purposes only and do not constitute investment advice, an offer, or a solicitation. Scores are computed from document content alone and do not verify the accuracy of any claim. Market data on this page reflects StockTitan's own measurements for the periods stated; past results do not guarantee future outcomes.