How the pipeline works

Headlines are collected around the clock, scored by a language model, aggregated into an index per asset, and published here as research and data. Dots travel along the arrows as headlines flow through.

System architecture · liveCollecting since 4 Oct 2024
01 · Ingest02 · Quantify03 · Index04 · Publish
News media
Global crypto press, wires and aggregators, polled every 15 minutes.
Official feeds
Exchange listings, filings, protocol and regulator announcements.
On-chain & social
Flows, governance votes and social chatter as corroborating context.
Scoring model
Every headline scored and typed at ingestion time.
score−1.0 … +1.0
typeproactive / reactive
assetstagged per headline
News index
A weighted calculation over the scored headlines: one continuous series per asset, set against price.
Research
Event studies and audits, with the data behind each figure.
Data & API
The index and the scored feed as JSON; history as CSV.
Production trading
The index is traded live: 60–150% a year on the production configurations.
end-to-end automated · no human latency
01 · IngestNews mediaGlobal crypto press, wires and aggregators, polled every 15 minutes.
02 · QuantifyScoring modelEvery headline scored and typed at ingestion time.
03 · IndexNews indexA weighted calculation over the scored headlines: one continuous series per asset, set against price.
04 · PublishResearchEvent studies and audits, with the data behind each figure.

Collecting since 4 October 2024. 243,760 headlines scored as of 28 October 2026. The dataset is described in note 4.

About

katelabs is a one-person research project by Alex Senkevich, a software engineer. It started in 2024 as a question: once the headlines that merely describe a price move are set aside, does the rest of crypto news tell you anything about what comes next?

Everything published here follows the same rule: the number comes with the data behind it and the things it does not show. The index is tested out of sample (note 1, note 2), live trades are reconciled against the simulation one by one (note 3), and the dataset is described together with its defects (note 4).

What this site is not: a signal service, a price-prediction tool, or investment advice. The trading system that uses this data is run privately; its parameters are not published here.

ContactEmail · [email protected]Telegram · @l3xxxyLinkedIn · axsenkevichX · @axsenkevichInstagram · @axsenkevichTelegram channel · Quant Lambda, publicAbout the creator →

Support the research

The site has no sponsor. News APIs, servers and model inference are paid out of pocket. If the work is useful to you, a transfer to any of these addresses keeps the headlines flowing and the data public.

Monthly, on PatreonIf you would rather support the work monthly than by transfer, there is a Patreon page. The tiers are symbolic: the research stays public at every level.
Open Patreon →
BTCbc1qagjep0uckyc4mzkfs8p0wrtm3pavt4j803wvt3
USDT · TRC-20THAnphFabG9T3wLuqWwyFq7VsCij4tRX7G
ETH0xd7b9634194480F26E0a778FaBb920f6617165ebe
XRPrJLD7TEpDHpADR3U9KaacH3949srUXKTU1
SolanaCjdt9Hnhcs4AzQFfXQjZ6fo9YN1waAowLpX2LaHsubRk

Nothing is owed in return and nothing is unlocked: the research stays public either way.

What the money goes toResearch time: the next notes and the data behind themBolder ideas at the crossroads of mathematics, crypto markets and economicsKeeping the data and the findings publicThe machinery that makes it run: feeds, servers, models