Discourse is what public companies are saying, read line by line. Every earnings call we can find — the S&P 500, five years deep — and the earnings release when the call isn't in yet, scored for tone, confidence, guidance, and the words that keep coming up.
What gets scored
Each document is read once by Numen and stored as a set of measures:
- **Sentiment** — management tone, from −1 to +1. Above +0.65 is bullish; below 0 is bearish.
- **Confidence** — how certain management sounded, 0 to 1. When confidence spreads apart across the universe, that dispersion is itself a signal.
- **Guidance** — for revenue, margin, and capital spending: raised, held, or lowered, and by how many percentage points when a number was given.
- **Language** — how often the document mentions AI, recession, layoffs, tariffs, and buybacks.
- **Themes** — what the call was about, with a salience and tone per theme. These drive the emerging and fading lists.
A call transcript always wins the quarter over a release. Fiscal labels are the company's own; everything aggregates on the calendar quarter of the call date.
Public sentiment
The panel at the top reads the whole universe: sentiment with its change against the prior quarter, confidence and its spread, the bull/neutral/bear split, guidance posture, and sentiment by sector. The trend chart runs at one week, thirty days, sixty days, one year, or five years — the shorter scales show only the documents scored in that window, and the counts above the chart say how many companies and reportings are in view. Drag across the chart to read the value on any date.
The corpus
Every tracked company, one row each, read from its latest scored document. Click any column header to sort; click again to reverse. Rows with no value sort to the bottom either way.
**Columns** opens the picker. Twenty-one are available, in five groups:
- **Tone** — Sentiment · Δ (vs prior quarter) · Tone (the score as a word) · Confidence · Δ Conf
- **Guidance** — Rev guide · Margin guide · Capex guide, each shown as ▲ raised, — held, ▼ lowered, with the move in percentage points when the extractor found one
- **Language** — AI · Δ AI · Recession · Layoffs · Tariffs · Buybacks, as mention counts on the latest document
- **Market** — Price (latest close) · Δ since call (the stock's return from the first close on or after the latest document) · Talk vs Tape (see below) · 1M %
- **Document** — Sector · Source (call or release) · Latest call
Your column set and sort are remembered in this browser. **Reset** returns to the default view — sector, sentiment, Δ, revenue guidance, AI, recession, price, Talk vs Tape, latest call.
Companies with no scored document yet sit dimmed beneath the table. The call lane syncs daily; the release lane hourly.
Talk vs. Tape
What management said, weighed against what the stock did afterward. For every company with a scored document and a price, Discourse takes the sentiment on its latest document and the stock's return from the first close on or after that document to the latest close. Both are z-scored across the corpus; divergence is the sentiment score minus the return score — talk minus tape.
- **SPIN** — tone well above the corpus (z above +0.4) while the stock trails it (z below −0.4). Talked up, stock down.
- **UNDERRATED** — the reverse. Talked down, stock up.
- **ALIGNED** — everything else.
Two guards keep the read honest. Returns are capped at the 2nd and 98th percentile before scoring, so one event-driven move can't flatten everyone else's read; the return shown is still the real one. And a company whose latest document is more than about five months old is left out — a return that long no longer reads as a reaction to what was said.
The strip above the corpus gives the universe read in one line — whether optimism is running ahead of the tape, the tape is ahead of the talk, or the two roughly agree — with the average sentiment, the average return since the call, and the three largest SPIN and UNDERRATED names. Each company's page carries the same read for that company, with its latest close, one-day and one-month change, and a price line over the same window as its sentiment history.
Horizons differ by company — a call last week and a call two months ago sit in the same cross-section. Read it as a comparison across the corpus, not an event study. Prices are end-of-day closes, refreshed each trading evening.
What the Fed is saying
`/discourse/fed` reads the Federal Reserve with the same scorer. Every Beige Book since 2021 is split into its national summary and twelve district sections, each scored on its own; FOMC statements and minutes and the Board's speeches are scored as they publish. The page shows the Board's sentiment and hedging rate over time, a twelve-district table with each district's tone and its change since the prior book, and the latest statement's shift against the one before.
Hedging is counted, not inferred — the rate of words like *may*, *could*, *uncertain*, *approximately* per thousand words. It is computed for every document in Discourse, Fed and corporate alike, so the two can be read against each other.
The Fed lanes never enter the corporate aggregates. Public sentiment, the sector table, and the corpus below stay S&P 500 only.
What the filings are saying
For every company in the universe, the latest 10-K risk-factor section and 10-Q management discussion are scored as well. Risk-factor language is read for how often it names AI, tariffs, cyber, litigation, labor, and climate, how hedged it is, and how long it has grown. These feed the `filings_discourse.*` signals and sit beside the call-based reads on Signal Watch; they are not mixed into the earnings-call averages.
Where it flows
Discourse feeds the lattice as the `discourse.*` signal family (with `fed_discourse.*` and `filings_discourse.*` beside it) — sentiment, confidence, dispersion, guidance posture, revision pressure, mention rates, and sentiment per sector — and those feed The Discourse Index, The Hesitation Index, Capital Deployment Posture, and AI Margin Compression. Macro View's *What companies are saying* section is the same read, condensed.