Question 1

How is the overall sentiment score calculated?

Accepted Answer

Every morning the system pulls articles from four news APIs, deduplicates them, and scores each one using Claude Haiku on a scale from -1.0 to +1.0. Before scoring, each article gets a relevance rating from 4 to 10 based on how directly it covers AI sector topics. The final category score is a weighted average where high-relevance articles count for more. The overall score is then an average across whichever categories had enough articles to score that day. Articles that are clearly off-topic get discarded entirely before they reach the scoring stage.

Question 2

What does above normal or below normal mean for a category?

Accepted Answer

Each category has a 30-day rolling history of daily scores. Above normal means today score is high enough, relative to that history, that there is less than a 5% probability it got there by random variation. It is a statistical statement, not a qualitative judgment. A category can score above normal on a day when the absolute score is still only mildly positive, if that category has been consistently flat for weeks and today reading is unusually high by its own standards.

Question 3

What does the p-value on a category signal mean?

Accepted Answer

The p-value comes from Welch t-test comparing today score against the category 30-day baseline. A p-value of 0.004 means there is a 0.4% probability that today reading is just random noise relative to recent history. It does not mean the signal is large in absolute terms, and it does not mean anything is about to happen in the market. It means the day coverage for that category was statistically unusual by its own recent standards, which is worth paying attention to.

Question 4

What is the 30-day baseline and how is it calculated?

Accepted Answer

For each category, the system keeps a rolling record of the previous 30 days of daily scores. The baseline is the mean and standard deviation of that window. As each new day score comes in, the oldest day drops off. This means the baseline adapts over time, so a category that has been running hot for several weeks will have a higher baseline and will be harder to flag as above normal than a category that has been flat. The baseline is category-specific, which is why a score of +0.3 can be unremarkable for one category and statistically significant for another.

Question 5

Where do the articles come from and how is source quality controlled?

Accepted Answer

Four commercial news APIs: NewsAPI, NewsData, Finlight, and Finnhub. Between them they cover most of the major financial and technology publishers. After fetching, articles go through deduplication to remove cross-source repeats, then a relevance filter. Articles that are tangentially about AI, or clearly off-topic, get a low relevance rating and either discarded or down-weighted heavily before scoring. The relevance rating is also done by Claude, so the same model that scores sentiment is making the judgment about whether an article belongs in the dataset at all.

Question 6

What role does Claude play in the scoring process?

Accepted Answer

Two roles. First, every article gets a relevance score from 1 to 10 and a narrative frame label. Second, the articles that pass the relevance threshold get a sentiment score from -1.0 to +1.0 with a short rationale. The model used is Claude Haiku, chosen for speed and cost at the volume of articles the system processes each morning. The scores are not manually reviewed or adjusted after the fact. What you see on the dashboard is what the model produced, which means the system is only as good as the model judgment on any given day.

Question 7

What is the Semantic Volatility Index actually measuring?

Accepted Answer

It is measuring how different this week AI news looks from last week, across five different dimensions. Two of those dimensions are leading indicators: whether the vocabulary of article descriptions has shifted, and whether the distribution of story types has shifted. The other three are concurrent or lagging: how spread out today scores are, how much the overall score moved since yesterday, and how variable scores have been over the past seven days. The composite score from 0 to 100 is weighted toward the leading indicators, because the point of the SVI is to flag narrative change before it shows up in the headline sentiment numbers.

Question 8

What should I actually do with this information day to day?

Accepted Answer

Most days the answer is nothing. A routine day with all categories in normal range and a low SVI is useful information in itself. It means the narrative environment is stable. The days worth paying attention to are when a category that has been flat for weeks suddenly flags as statistically significant, when the SVI leading indicators start rising while the headline scores look calm, or when multiple categories move in the same direction simultaneously. Those configurations have historically preceded broader narrative shifts by a day or two. It is not a trading signal and is not designed to be one. It is an early warning system for people who care about where the AI sector narrative is heading before it becomes obvious.

Category	Score	95% Confidence Interval	vs History	Verdict	Coverage
🔬 AI Chips and Hardware	+0.416	+0.285 to +0.547 Moderate (n=60)	→ Normal range hist +0.488 p=0.296	POSITIVE ↑	62 articles
☁️ AI Platform Hyperscalers	+0.361	+0.138 to +0.584 Low (n=25)	→ Normal range hist +0.302 p=0.597	POSITIVE ↑	26 articles
💼 AI Enterprise Software	+0.035	-0.289 to +0.358 Low (n=20)	↓ Below normal hist +0.423 p=0.022	NEUTRAL →	21 articles
🏗️ Data Center Infrastructure	+0.644	+0.535 to +0.753 Moderate (n=12)	↑ Above normal hist +0.429 p=0.001	POSITIVE ↑	12 articles
⚡ AI GPU Cloud	-0.036	-0.704 to +0.632 Low (n=7)	→ Normal range hist +0.463 p=0.084	NEUTRAL →	7 articles
🤖 Pure Play AI	+0.602	-0.088 to +1.000 Low (n=3)	→ Normal range hist +0.308 p=0.203	POSITIVE ↑	3 articles

Category	05-06	05-07	05-08	05-09	05-10	05-11	05-12	05-13	05-14	05-15	05-16	05-17	05-18	05-19
🔬 Chips	+0.51	+0.63	+0.43	+0.56	+0.51	+0.43	+0.51	+0.38	+0.52	+0.61	+0.36	+0.70	+0.55	+0.42
☁️ Hyperscalers	+0.35	+0.19	+0.36	+0.33	+0.56	+0.16	+0.34	+0.19	+0.61	+0.15	+0.46	+0.28	+0.46	+0.36
💼 Enterprise SW	+0.59	+0.56	+0.54	+0.19	+0.25	+0.26	+0.55	+0.54	+0.53	+0.44	+0.20	+0.37	+0.22	+0.03
🏗️ Data Centres	+0.17	+0.45	+0.41	+0.59	+0.36	+0.49	+0.43	+0.41	+0.63	+0.56	+0.30	+0.37	+0.16	+0.64
⚡ GPU Cloud	+0.73	+0.59	+0.40	+0.38	-0.03	+0.03	+0.70	+0.56	+0.27	+0.70	+0.64	+0.05	+0.57	-0.04
🤖 Pure Play	+0.53	+0.17	+0.36	+0.19	+0.19	+0.19	+0.54	+0.58	+0.52	-0.16	+0.48	+0.42	-0.30	+0.60

Canary Dashboard

Overall Sentiment Trend

Category Breakdown — 2026-05-19

Category Sentiment Over Time

Category Trends — Individual View

Narrative Frame Distribution

Semantic Volatility Index