Data / Narratives

Narrative data

What the market is talking about this hour, clustered into discrete topics, ranked by how many independent sources are covering each one, and labelled with a category.

What this data is

Every hour, the news and social flow of the preceding period is clustered into topics. Each snapshot carries the 10 strongest clusters, and each cluster carries the number of distinct sources that contributed to it. That count is the whole point: it separates an event that eighteen outlets independently thought worth covering from one that two did.

Because snapshots are retained, the series answers a question a single snapshot cannot: is this topic growing? A cluster that appears at count 4 and reaches count 18 over six hours is a narrative forming. The same cluster flat at 4 for a day is not.

Narrative detection

Input flow

deduplicated news itemssocial topic activity

Clustering

group by entities and eventmerge near-duplicatescount contributing sources

Labelling

headline for the clustercategory assignmentsummary

Snapshot

10 topicsshared timestampretained

Momentum

compare snapshot to snapshotrising / falling / new

Update frequency

Hourly. All topics within one snapshot share a timestamp to the second, which makes snapshot boundaries unambiguous when comparing consecutive hours.

Available fields

idintegerIdentifier of the topic within the snapshot.

topicstringA short, human-readable headline for the cluster. Written to name the event, not to be clicked.

countintegerNumber of distinct source items that fell into the cluster. This is the ranking signal.

categorystringClassification of the topic. The set is dynamic — see below.

summarystringA paragraph-length account of what happened and why it is being discussed.

createdDatetimestampUTC timestamp of the snapshot. All topics in one snapshot share it.

one topic from an hourly snapshot

[
  {
    "id": 62136,
    "topic": "Cronos Tectonic Exploit Drains Roughly $75 Million",
    "count": 18,
    "category": "Hack",
    "summary": "A rapid exploit on the Cronos network targeted the Tectonic lending protocol …",
    "createdDate": "2026-08-31T09: 00: 15.502Z"
  }
]

Limitations

Coverage volume is not importance. count measures how many sources covered a topic, which correlates with newsworthiness and with how well-resourced the parties involved are at getting covered.
Top-N truncation. Only the strongest 10 clusters per hour are retained. A narrative below that threshold is invisible until it crosses it, so the series cannot be used to study the long tail of emerging topics.
Categories can shift. The label set is dynamic. A dashboard that hard-codes today's categories will break, and should treat category as an open string.
Clusters are not assets. A topic mentions assets but is not keyed to one. Narrative-to-asset attribution is an analysis step with the same ambiguity described under news mapping.
English-language bias. Inherited from the underlying news coverage.

Which applications use it

Narrative snapshots power the trending topics surface on moonboard.ai, including the rising and new markers that come from comparing consecutive snapshots.

News data — the deduplicated flow that clustering runs over.
Social data — per-asset attention, the other half of narrative growth.
Historical data — retained narrative snapshots.