Notes — Investolytix

These started as posts written twice a week while building Investolytix. They are collected here because a finding about how investor conversation actually behaves is worth more than a feed item, and because the reasoning behind a measurement belongs next to the measurement. The method itself is set out on How we measure.

October 2026

wk 19

Fourteen of thirty have no direction

Not missing data. No direction. A week earlier the same board had two. Balanced means there is plenty of conversation about a company and it is genuinely pulling both ways, with strong opinions in both directions that cancel out. Most tools report a number near zero and leave you to assume nothing is happening, but “nobody is talking about this” and “everybody is arguing about this” are opposite findings and a score of 0.01 describes both. The board went from 23 Bullish and no Bearish to nine and four, with most of the rest moving into Balanced rather than crossing over. The crowd did not change its mind. It stopped agreeing with itself.

wk 19

This week one of them was the word “verified”

New companies list every Sunday, and each one brings a name to match. This week one of those names was an ordinary word. It turned up 26 times in seven nights and almost none of those sentences were about a company. Blocking it outright is the tempting rule and the wrong one: the same logic would have blocked “fargo”, which matched 204 times and was literally Wells Fargo in 203 of them. A high count is not proof of a trap, only proof that someone has to read the text. This week's refresh added fourteen names and dropped eleven, so the total moved by three while twenty-five things changed underneath it.

wk 18

The night it detected itself

One night we took in 22,279 items against a normal night of about ten thousand, because capacity had been added to a source. Nothing had happened in the market. Any measure of “unusual attention” built on raw mention counts would have flagged every company at once, reporting a change in our own pipeline as news about the world. The measure is now share of the corpus rather than raw count, because a metric that counts your own activity eventually measures you instead of the thing you were watching.

wk 18

MD is Maryland, and also a ticker

So are TX, NC, SD, MI, MT, AZ, NE and DC. “Weak demand in TX” is a sentence about Texas, and counting it as a company mention attaches real sentiment to the wrong company. Blocking every two-letter match fails too: WY is Wyoming, but in investor conversation it is overwhelmingly Weyerhaeuser, and blocking it would delete genuine coverage to avoid a problem that barely happens. Nine are now blocked and five kept, decided one at a time against the actual text. No rule settles this.

wk 17

A Sunday reading is not a Wednesday reading

Two sentiment readings can show the same value and mean different things, because the amount of conversation behind them moves with the week. Across one week the weekend averaged 6,389 items a night against 9,432 on weekdays, about a third less. Thin data moves easily, and a bare score looks equally confident either way. Which is why every reading we publish carries its window, thin names are marked as thin, and some get no score at all.

wk 17

Alphabet has two tickers

One company, two share classes, and the question of where a mention of “Google” belongs. In our own data Alphabet's mentions land on GOOGL about 80% of the time and on GOOG about 20%, and a large share name no ticker at all. Keep the tickers fully separate and each reads as half as discussed as the company really is; pool them and you tell a GOOG holder something measured about a share class they do not own. The rule we settled on: sentiment belongs to the company, and a ticker is the address you look it up at.

wk 16

Most bad news isn't news

A majority of the bearish items arriving from one conventional news source turned out to be securities class-action law-firm advertising. For dozens of companies, that material on its own moved the displayed sentiment more than a typical week's real movement did. A bias that only ever points one direction cannot be fixed by averaging over more of it. It has to be identified.

September 2026

wk 15

100 days, no missed nights

A hundred consecutive nights of collection, and the number worth stating carefully is the distinct one rather than the raw one, because the same article reaches you several times. The more useful detail: when a source went dark mid-September it was flagged the first night, because a source returning nothing looks exactly like a slow night unless something checks.

wk 15

Tickers move

BNY Mellon now trades as BNY rather than BK. We found 187 symbols marked retired that were in fact live companies trading under new tickers, and a news source still filing BNY stories under the old symbol 189 times in a fortnight. Symbol history is not trivia; a company whose ticker changed is a company whose conversation gets split in two unless somebody notices.

wk 14

Know what moves your metric

A coverage metric fell several points over four nights with nothing broken. The cause was the earnings calendar rather than the collection pipeline. The line that stuck: nobody audits good news. A metric that rises is almost never investigated, which is exactly why a rising one can hide the same fault.

wk 14

Choosing not to fix it

We found a name-collision problem, measured it, and deliberately left it in place, because the fix would have deleted more real coverage than junk. Measuring a problem and declining to act on it is a legitimate outcome, and a more honest one than a fix that quietly makes the data worse.

wk 13

Balanced is not the same as thin

Four of the thirty most-discussed names had no defensible direction that week. That is not an absence of data. Plenty of conversation genuinely pulling both ways is a different finding from not enough conversation to measure, and a product that reports them with the same word is hiding the more interesting one.

wk 13

The Boeing bug

Four days spent deleting bad ticker matches, and then the biggest improvement came from adding data back. Companies colliding with their own preferred shares meant several large names could not be matched by name at all. A precision problem turned out to be blocked behind a recall problem.

wk 12

First look at the new interface

The redesigned interface, built around a platform-divergence radar. The reading it opens on is not a single score but a spread: ten platforms, and a gap of 0.44 between the most bullish and the most bearish of them. The disagreement is the headline, which is the whole argument for not pooling sources into one number.

August 2026

wk 11

Why crowd mood matters now

The structural case for the category. Zero-commission retail participation, information flowing sideways between investors rather than downward from institutions, and attention that coordinates. Decades of infrastructure exist for fundamentals and for technicals. For the third force there is still almost nothing systematic.

wk 11

The market doesn't take holidays

Written from a holiday. The engine collected and scored every night regardless, which is the point of building it as infrastructure rather than as something a person runs. Also a re-introduction of the premise for anyone who had arrived recently.

wk 10

Sentiment analysis is not word counting

Sarcasm, emoji, market dialect and understatement are not edge cases in investor language, they are the register. An off-the-shelf polarity tool reads “great, another dip” as positive. Scoring investor text requires knowing what the words mean in the room they were written in.

wk 10

Frequency is the wrong filter

The tempting way to remove noise from name matching is to drop whatever appears too often. It deletes your best data, because Bitcoin, Google, Apple and Tesla are frequent for the right reason. The usable rule is an intersection of conditions, never frequency on its own.

wk 9

Too few to call

Every reading carries a confidence interval, and the intervals are what make the problem visible. A company with three mentions scored +0.43 with an error bar of ±0.49, against Apple at +0.01 ±0.02 on 1,858. At the time, 91.7% of instruments were too thinly covered to call. Saying so is the honest output.

wk 9

Counting the right number

A published figure was counting observations rather than distinct items, 369,060 against 283,383, because the same article arrives through several routes. We corrected it publicly. The same pass fixed a chart that made a partial month look like a 78% collapse.

wk 8

Facts versus opinions

Opinions cost money to score and facts are free. Filings, disclosures, ratings changes and corporate events are facts, and they are never sentiment-scored, because scoring a fact for mood is a category error that makes the output look richer and means less.

wk 8

The first screenshot

First public look at the product: the sentiment gauge, the per-platform breakdown, the mood chips showing divergence, and sentiment laid over price. Showing it was overdue.

July 2026

wk 7

Insiders versus the crowd

People sell for many reasons and buy for one. The crowd is loud and fast; insiders are quiet and deliberate, and the two measure different things. Reading them as one signal collapses the only interesting distinction between them.

wk 7

Honest about lag

We shipped a congressional-trading lens and then deliberately de-hyped it, because the disclosures behind it arrive around 44 days late. It is not a live feed and presenting it as one would be the easy, dishonest version. A lens is worth having when you know what it can and cannot see.

wk 6

Never pool your sources

Each platform has its own baseline. One community runs euphoric, another is structurally sceptical, the news reads flat. Averaging them gives a number that is arithmetically correct and practically meaningless. Calibrate each against its own normal, and treat the disagreement between them as a signal in its own right.

wk 6

Opinion and fact, side by side

Alongside what people are saying, a second dataset of what actually happened: earnings calls, filings and insider transactions. Two kinds of evidence, kept separate on purpose.

wk 5

The photo and the film

A sentiment reading on a single day is a photograph. The useful object is the film: how the mood arrived at today's value, and how quickly.

wk 5

Saying and doing

What a market says about a company and what participants actually do with their money are different measurements, and the gap between them is worth watching in its own right.

wk 4

The whole room

Listening to one platform tells you what one kind of investor thinks. The point of breadth is not volume, it is being able to tell when one room disagrees with another.

wk 4

Solid ground

An upgrade to the underlying market data. Sentiment laid over bad price data is worse than no sentiment at all, because it looks like analysis.

June 2026

wk 3

Sentiment versus price

When mood and price point in different directions, that divergence is a prompt to look closer. It is never a signal to act on, and we will not present it as one.

wk 3

More data, and cleaner data

Scale and quality are not the same project, and the second one is the one nobody posts about.

wk 2

The emotions underneath

Direction alone is a thin description. Hype, conviction, fear and doubt separate two companies sitting at the same score for entirely different reasons.

wk 2

Purpose-built beats generic

A sentiment engine built for investor language outperforms a general-purpose one on investor language. The finding is unsurprising. The size of the gap was not.

wk 1

A third lens on GameStop

A worked example of the premise: a company where fundamentals and technicals explained very little of what was happening, and where the crowd was the story.

wk 1

Where the data comes from

The first build-in-public note on sources. The principle has not changed since: breadth exists so that platforms can be compared against each other, not so the totals look bigger.

Investolytix isn't live yet. The waitlist is.

Founding members get early access and free Pro at launch. New notes land here most weeks.

Join the waitlist