Where every number comes from
Every signal is scored on two independent questions, shown separately and never combined into one number:
Validity — is the event real? (corroboration across independent sources)
Direction confidence — which way might it move, and how sure are we? (a blend of the four dials below)
The dominant directional input. It answers: “Historically, when a politician made this type of statement, how did the stock move — and how often was it right?” Each claim is mapped to an archetype (contract, endorsement, tariff, antitrust, subsidy, …) that carries an expected direction, magnitude, and historical hit-rate.
Current status: the per-archetype numbers are hand-seeded estimates, not yet measured from real history. A real event study (mining past statements + actual price reactions, with a recalibration job) is the next data milestone. Until then these figures are reasonable placeholders — which is why this dial is tagged ESTIMATE and the others are tagged REAL.
STOCK Act disclosures: when a member of Congress buys or sells the same ticker, that's a strong secondary signal. Buys lean bullish, sells bearish; confidence grows with the number of disclosures.
Current status: the free community datasets (senate/house-stock-watcher) went offline, and the official House Clerk data is ZIP/XML/PDF. So this dial currently uses a small, clearly-labeled sample set. Real House Clerk ingestion is on the roadmap.
Computed from real daily OHLCV pulled live from Yahoo Finance: SMA (20/50), RSI (14, Wilder's smoothing), MACD, ATR, and a trend read. Standard textbook formulas, self-tested and cross-checked against an independent recomputation. If prices can't be fetched, the dial simply doesn't fire — no value is invented.
Note: the Yahoo endpoint is unofficial (their ToS is gray for programmatic use); the numbers are genuine market data, the access method is being hardened for production.
Pulled from official SEC EDGAR filings (XBRL company-facts): latest annual revenue, year-over-year growth, and profit margin, detected by the actual reporting period (~365-day span) so we always use the genuine most-recent fiscal year. Missing figures are left blank, never guessed.
For each claim we gather independent coverage (GDELT + Google News) and have a language model judge every article: corroborates, mentions, or contradicts. Merely naming the company is not corroboration. Validity blends the count of distinct corroborating publishers, their source quality, and agreement — shown in full on each signal's proof panel.
The four dials are combined with regime-aware weights (event-study leads, congress second, technical and fundamental lighter). Only the dials that actually fired are counted, and their weights are renormalized — so a missing input never dilutes the result. The final number is a confidence, not a certainty.