What sentiment analysis is, and why a score is not a signal
Sentiment analysis is software reading a piece of text and scoring how positive or negative its wording is. It measures LANGUAGE, not a business. Everything else here follows from that.
Two well-known models disagree often enough to be worth comparing. VADER is a general social-media lexicon: it scores tone from the words themselves. FinBERT is trained on financial text, so it recognises vocabulary VADER cannot.
When two models disagree, that disagreement is itself information — and averaging them into one confident-looking number throws it away. A single score hides the fact that two readers of the same sentence reached opposite conclusions.
A real example from testing the two models during development: “Earnings beat estimates but shares slip on cautious outlook.” VADER scored it +0.51 — it heard “beat” and “estimates.” FinBERT scored it −0.94 — it understood “shares slip on cautious outlook.”
Where it fails: VADER is deaf to financial phrasing, and both models are reading tone, not judging a business. A sentence can be worded gloomily about a healthy company and cheerfully about a failing one.
And this is why Aureus does not sell a sentiment score as a signal. A number that measures wording cannot tell you whether something is worth owning, and presenting it as though it could would be the clearest possible version of the thing this product does not do. The technique is genuinely useful for understanding how a story is being told. It is not a forecast, and no arrangement of it becomes one.
How much each source counts
Not every outlet gets an equal vote. Before the tone scores are averaged, each headline is multiplied by a weight for the publication it came from — a wire service that reports a number is weighted a little above a site that reports an opinion about the number. The weights are modest on purpose: the widest spread in the table is Reuters at 1.20 against Seeking Alpha at 0.92, so weighting nudges a reading, it never decides one. Anything not listed counts as 1.00 — plain, unadjusted.
This is the complete table. There's no longer version, and it's the same table the scoring code holds. Weighting applies when headlines arrive; with none, there is nothing to weight.
| Source | Weight |
|---|---|
| reuters | 1.20 |
| associated press | 1.15 |
| ap | 1.15 |
| bloomberg | 1.15 |
| the wall street journal | 1.12 |
| wall street journal | 1.12 |
| financial times | 1.10 |
| marketwatch | 1.05 |
| cnbc | 1.02 |
| investing.com | 0.98 |
| yahoo finance | 0.98 |
| benzinga | 0.95 |
| motley fool | 0.92 |
| seeking alpha | 0.92 |
Fourteen rows, twelve publications: Associated Press and AP, and The Wall Street Journal and Wall Street Journal, are alternate spellings carrying the same weight — news vendors are not consistent about either name, so both are matched.
Worked example
Two headlines, both scoring +0.50 on tone before weighting. One came from Reuters, one from Motley Fool. (The scores are the example; no headline is being invented and attributed to either newsroom.)
- Reuters: +0.50 × 1.20 = +0.60
- Motley Fool: +0.50 × 0.92 = +0.46
- Unweighted average: +0.50. Weighted average: +0.53.
So the weighting moved this reading by 0.03 — the size of thing it does. A headline can never be pushed past the ends of the scale: after multiplying, each score is clamped to the −1.00 to +1.00 range the models already use.
When it doesn't run at all
Weighting compares sources against each other, so it needs more than one. If every headline in a reading came from the same publication, weighting is switched off and the scores are used raw — otherwise one outlet's weight would be applied to every row alike, which shifts the whole reading without telling you anything about the sources. That happens often, and each reading records which way it went rather than this page claiming a general rule.
One more place it does not apply: when a coin has only one or two headlines, each is shown with its own tone and no average is produced. Those per-headline numbers are always unweighted — a single headline has nothing to be weighed against.
The two models are never compared through this table. Whether tone and financial language agree, and how far apart they are, is computed from unweighted scores on both sides — comparing a weighted number against an unweighted one would measure the publications as well as the language.
That is also why the reading changes character once both models have run: the score you then see is the plain average of the two raw readings, with no source weighting in it at all. The card says so, and when you open the two readings side by side, the two numbers shown are exactly the two the gap between them is measured from — so you can do the subtraction yourself and get the same answer.
This explains what things mean. It isn't investment advice, and nothing here says whether an asset is worth buying.