A research platform that measures the social side of markets. We collect the public conversation about more than 14,000 stocks, ETFs, and crypto from across the internet, filter out the noise, and turn what's left into attention and sentiment metrics you can screen, chart, and compare against 13 years of history — alongside price and fundamentals. The long version is on our about page.
No. If it did, we would be running a hedge fund, not selling subscriptions. Attention and sentiment tell you what investors are talking about, how strongly they feel about it, and whether today's conversation is unusual against years of history. That is evidence for a thesis, not a trading signal. Use it alongside fundamentals, not instead of them.
No. There are no buy ratings, no price targets, and no proprietary score designed to drive a trade. We publish measurements and document how they're computed. What you do with them is up to you.
Most of the site is free with an account: trends, heatmaps, asset overviews, daily summaries, financials, basic screener filters, and 90 days of daily history. Paid plans add the full screener filter set, 13 years of hourly history, portfolios, CSV exports, and real API volume. Details on the pricing page.
Public posts from across the internet — the popular social networks as well as smaller forums and communities. Everything is collected continuously and processed through the same pipeline, whatever the source.
The data is public, but the people aren't the product. We use it only to build aggregate analytics and to train models for our own use. We do not sell, display, or redistribute anyone's posts without permission — what you see on this site are measurements, not other people's content.
Coverage starts in July 2012 — about 13 years and counting. Individual assets start later if they listed later: Ethereum from July 2015, newer listings and tokens from launch.
~7,500 stocks (NYSE, NASDAQ, NYSE American), ~6,700 ETFs, and ~190 cryptocurrencies. More than 14,000 in total.
Billions collected; far fewer kept. Most of what's posted about any asset is spam, bots, or off-topic noise, and we discard it before it reaches the dataset. The number that matters isn't how much we collect — it's how little we keep.
Collection runs continuously. Posts are classified and aggregated as they arrive. Timeseries data rolls up hourly. There is no end-of-day batch.
Every post passes through two classifiers before entering the dataset. The first is a spam filter. The second is an asset-relevance filter — a post that mentions "$AAPL" in passing isn't the same as a post analyzing Apple's earnings. We train separate filters for each source and asset class.
Ticker collisions are constant — Apple the fruit, tickers that double as ordinary words, crypto tokens named after memes. For ambiguous assets we train a dedicated disambiguation classifier that decides whether a post is about the asset at all before anything else runs. Posts that can't be attributed confidently are discarded.
Each post is classified as Bullish, Bearish, or Neutral. We train separate models for each combination of source and asset class because the language and norms differ substantially from one community to the next.
Each post is tagged with the topics it discusses. A separate binary classifier runs for each topic — a post is either about that topic or it isn't. Topics are defined per asset class:
Daily and monthly narrative summaries per asset. They cover sentiment shifts, active topics, and what's driving discussion.
Three big ones. Private communities — invite-only servers, group chats, anything behind a login — are invisible to us; we only read public posts. Classifiers are imperfect — they misread sarcasm and slang some fraction of the time, which is part of why we weight by engagement rather than counting every post equally. And engagement is shaped by where it happens — a share, an upvote, and a video view are not the same act, so comparisons across sources are approximations. Social data is a noisy signal. We'd rather you know that.
Engagement is the total interaction count on a post — likes, replies, shares, upvotes, video views, whatever the source offers — summed into a single number. Video views count because watching a video is something a person chose to do. Impressions do not: we never count how many feeds a post happened to scroll past in. We aggregate engagement rather than post count, so a viral post with 10,000 shares is weighted more heavily than 100 posts nobody touched.
Attention is an asset's share of total engagement across all assets in its class, over a rolling window. If NVDA generates 3% of all stock engagement in a 24-hour window, its daily attention is 3%. It answers the question that actually matters: is this asset getting an unusual amount of the conversation right now? Computed for day, week, month, and year windows, with 50-day and 200-day moving averages, year high/low, and all-time high.
The change in percentage points versus the prior window. "6.3% +0.4" means the asset holds 6.3% of attention now, up from 5.9% — a 0.4-point gain, the same convention as a stock quote. "new" means the asset had no measurable attention in the prior window. We use point deltas instead of relative percent change because relative change misleads at small bases: going from 0.1% to 0.2% of the conversation is "+100%" but barely a ripple. Price changes are the exception — those are ordinary percent moves.
Net bullish = max(0, bullish engagement - bearish engagement), normalized against the total net-bullish conviction across all assets in the class — not total engagement. Assets with evenly split sentiment contribute nothing to that pool, so it's small and concentrated: a single asset can legitimately hold most of a day's net-bearish conviction. Net bearish is the inverse, and an asset is either net bullish or net bearish in a given window, never both. This is engagement-weighted — it reflects the sentiment of posts people actually interact with, not just post volume.
Any metric in the dataset — attention, engagement, sentiment, topic activity, plus price and fundamentals — can be a column, a filter, or a sort, for each asset class. Screeners are free with an account; free accounts get a basic filter set, and paid plans unlock all of them plus CSV export.
Twelve Data. Daily and hourly prices for stocks, ETFs, and crypto.
Yes. Each stock's financials page covers earnings (actuals against analyst estimates), revenue, full financial statements, key statistics and ratios, and dividends. Fundamental data comes from Twelve Data, the same source as our price data.
Yes. Save assets into portfolios to follow them together. Personal plans include 5 portfolios; Pro includes 30.
Yes. REST endpoints for asset snapshots, historical timeseries, and summary metrics, authenticated with an X-API-Key header. Your key is generated when you register, and rate limits scale with plan — the free tier includes a demo budget to try it. Full reference, including an OpenAPI spec, is in the documentation.
Yes. We run a Model Context Protocol (MCP) server at https://rumor.io/api/mcp — point Claude, Cursor, or any MCP-aware client at it using the same API key as the REST API. Four tools cover the surface: a single-asset snapshot, historical timeseries, summary metrics, and market-wide sector trends. Setup instructions are in the documentation.