01Each platform is scored on its own terms, and never pooled
A measured note in an earnings call and a one-line post written at midnight are both opinions about a company, and they are not the same kind of evidence. Most sentiment products average everything into a single number, which throws away the most useful thing in the data.
We score each source separately and keep them separate. Eleven sentiment platforms, each with its own reading for the same company on the same day.
When two platforms disagree about the same company, that disagreement is the finding. An average would have deleted it.
A company that reads strongly positive in retail conversation and flat in the financial press is in a different position from one where both agree, even when a blended score would put the two side by side. We show you which platform said what.
02The four labels, and the fifth answer
Scores run from −1 to +1. The label is not calculated from the rounded score you see, which is why two companies can show the same displayed number and carry different labels. The label comes from the full distribution underneath it.
Net positive across the window, with the balance of conversation leaning one way.
Net negative across the window, on the same basis.
Close enough to zero that calling a direction would be inventing one.
Often a company nobody is arguing about this week.
Plenty of conversation, genuinely pulling both ways. Strong opinions in both directions that cancel out.
Not the same as Neutral. A heavily argued company sits here, and the argument is the point.
Neutral and Balanced both render without an arrow, because neither points anywhere. The difference is how much conversation sits behind the reading.
The fifth answer is too few to call. Where a company has attracted too little conversation to measure, it gets no score. A number built on a handful of posts looks exactly as confident as one built on two thousand, so we would rather publish nothing than publish that.
03Coverage is part of the reading, not a footnote
Two readings can show the same value and mean different things, because the amount of conversation behind them is not constant. It moves with the week. Weekends run materially lighter than weekdays, since the market closes and the talking thins out even though it never stops.
Thin data moves easily. A few loud posts about a smaller company on a thin night can push its reading somewhere the same posts would not push a busy Wednesday. So every number we publish carries the window it covers and an indication of how well covered it is, and readings come with error bars rather than a single decimal pretending to precision it does not have.
The question worth asking of any sentiment tool
How much conversation is behind the number on screen, and over what window? If the tool will not tell you, the number is doing less work than it appears to.
The rolling seven-day readings on our homepage ribbon are dated for exactly this reason.
04Direction is not the whole story
Positive and negative is a thin description of how people talk about money. Alongside direction we measure five emotional registers, because two companies can sit at the same score for very different reasons.
A company at +0.2 carried by conviction is in a different place from one at +0.2 carried by hype. The direction matches. The texture does not.
05How a post becomes a reading
Collect
Thousands of posts, comments, articles and transcript passages every day, across eleven sentiment platforms including Reddit, Bluesky, StockTwits, YouTube, Hacker News, financial news and earnings calls (the full list is on Sources). Collection runs nightly and every source is checked for whether it actually returned anything, because a source that stops returning data looks identical to a slow night.
Work out who is being talked about
This is the hard part, not the scoring. People write "Netflix", not NFLX. They write "Google" for a company that trades under two tickers. A single word can be a company, a place, a brand somebody else owns, or an ordinary English word. Sentiment attached to the wrong company is worse than no sentiment at all.
Score, per source
Each item is read for direction and emotional register in the context of where it was written, then aggregated within its own source. Nothing is blended across sources at this stage, or any later one.
Aggregate with the coverage attached
Readings are rolled up over a stated window, with the volume behind them and a confidence interval. Where coverage is too thin, the output is "too few to call" rather than a number.
Lay it over market data
Sentiment on its own is a curiosity. Beside price, volume and a second dataset of structured filings, disclosures, ratings and market context drawn from ten further lenses, it becomes a third lens, next to fundamentals and technicals.
06What makes this hard, and what we do about it
We would rather tell you about these than let you discover them.
Names collide with ordinary language. Short company names are the worst offenders, because they were words before they were companies. A high match count is not proof of a problem and a low one is not proof of health, so the only reliable test is reading the text.
Tickers move and companies carry several of them. Symbols get retired and reissued, and a company with multiple share classes has its conversation split across them unless something puts it back together. Sentiment belongs to the company; a ticker is the address you look it up at.
Some bearish material is not news at all. A meaningful share of negative items in conventional news feeds is promotional or templated material that is one-directional by construction. Averaging cannot fix a bias that only ever points one way, so it has to be identified instead.
New listings arrive every week. Which means name matching is not a problem that gets solved once. Every new listing brings a fresh set of words that now mean a company, and some of them were already in use. We check the difference every week. We are not going to claim this is finished, because by design it never is.
07What we don't do
- ✕We don't predict prices. Investolytix describes what is being said and how strongly. It does not forecast what a price will do next, and no part of the product is built to.
- ✕We don't give advice or signals. There are no buy or sell recommendations, no scores to trade off, no portfolio suggestions.
- ✕We don't blend platforms into one number. Doing so would be easier to display and would destroy the most informative part of the data.
- ✕We don't publish a reading we can't stand behind. Thin coverage gets said out loud instead of being dressed up as a measurement.
- ✕We don't claim the data is clean. It is checked daily and it is better than it was. That is a different statement.
Investolytix isn't live yet. The waitlist is.
Founding members get early access and free Pro at launch. If anything on this page is wrong, unclear or too convenient, we would rather hear it now.
Join the waitlist