Methodology

What the percentage means

Narrative sentiment is the bullish share of an asset’s eligible articles that take a side. Bearish is the rest: 100 minus the bullish share. An article can argue different directions for different assets, and articles that take no side count toward neither.

A reading is the share of directional coverage that argued the asset higher, over the seven publication days behind the edition.

The unit counted is an article–asset pair: each article is read once for each asset it takes a view on, so a piece arguing Bitcoin higher and the S&P 500 lower is one bullish pair and one bearish pair. A pair that takes no side is held out of the reading and reported separately — the denominator is the pairs that took a side, bullish plus bearish, and nothing else. A reading of 41% means that of the coverage that took a side on that asset, 41% argued it higher.

The reading counts the coverage an asset is a subject of. Where an article takes a view on an asset only in passing, that pair is held out of the reading and reported separately, so the share describes coverage about the asset rather than coverage that mentions it.

Below twenty directional pairs in the window, the asset has no reading. The page says so, in words, instead of printing a number that four articles could move ten points. The bond is read as a price, not a yield, so “argued the 10-year higher” means its price higher and its yield lower.

Nothing about the reading is smoothed, calibrated or weighted by recency. The seven days are summed flat and the share is taken once. A share needs no calibration and cannot pin at 0 or 100 by construction.

An edition dated the 22nd covers articles published on the 15th through the 21st. The pipeline runs once nightly and reads the prior day’s coverage, so the edition date and the publication window are one day apart by construction, and every page states the date its data describes.

Each article is counted once per asset, using the newest scoring of it, and the share is rounded to a whole percentage.

The time window

A current reading uses seven publication days. An edition dated on a given day reads the articles published on the seven days before it (UTC), scored the following morning.

The edition date, the data cutoff and the article window are separate things: the edition is the day the reading is published; its articles stop at the end of the day before. On a chart, “History shown” changes how many daily readings are drawn, not each reading’s own seven-day window.

A narrative’s share of coverage uses a shorter, three-day window, weighted toward the most recent day, because it measures attention rather than direction.

From articles to narratives

Coverage is collected, read for each asset it discusses, and grouped by recurring explanation into narratives. Summaries describe that evidence. Automated reading can miss context, and a repeated report is not independent confirmation.

1. Collection Coverage is gathered daily from a defined set of public news and market sources, and each article is filed under every asset it discusses. Sources are counted by article, not by outlet: one publisher contributing five pieces contributes five.
2. Reading Each article–asset pair is read by a large language model (Anthropic Claude) for the case the article makes about that asset — higher, lower, or no side — with the sentence it rests on kept as evidence.
3. Clustering Pairs making the same argument are grouped by text embedding into a narrative, so the same thesis in different words is one argument tracked over time rather than many.
4. Writing A nightly language-model pass writes the plain-English summaries, grounded in the extracted evidence. Where the writer states something the site does not publish — a count, a raw score — the passage is dropped rather than rewritten.

Lifecycle measures attention

Emerging, Accelerating, Saturated, Fading and Stable describe the attention a narrative receives. They do not say whether an asset is bullish or bearish, and a narrative does not have to pass through them in order.

Each narrative is staged against its own asset’s coverage. Its share is that asset’s narrative coverage over three days, weighted toward the most recent; its momentum is the change in that share against the three days before. If a narrative has fewer than two articles in the window, it is Stable, whatever its share or its move. Otherwise the first rule that fits applies, in this order: Fading when its share fell by more than 3 share points; Emerging when its share is under 3% and rose by more than 3 points, or when it is new (first seen less than fourteen days ago, with fewer than ten articles) and rising at all; Accelerating when its share is 3% or more and rose by more than 3 points; Saturated when its share is 3% or more otherwise; and Stable for everything else.

What this can tell you

The readings describe published coverage. They are not a poll of investors or a price forecast. Coverage can be incomplete, concentrated in a few publishers, or wrong.

When a reading is unavailable

Too little eligible coverage produces an unavailable reading, not 0% or a neutral score. Below twenty directional pairs in the seven-day window, the asset has no reading, and the page says so in words.

A page whose data could not be loaded says so as a loading failure, which is different again: it is not a reading of zero and not thin coverage.

Validation and corrections

Published validation should identify its method, sample and limitations. What has been tested so far, and how, is below.