News moves Markets,
so we Measured it.
We classified 72,036 historical events spanning ninety years. Analyse how asset classes reacted under stress, policy shocks, and structural shifts. No predictions. Just cold, empirical evidence.
"Iraq invades Kuwait."
This has happened before.
When news drops, VEILLE matches it against 72,036 classified events and pulls its closest historical analogues. Below, four live pairings: the modern event's path laid directly over its precedent's, session by session, with the correlation measured — not asserted.
Nothing here forecasts. Both paths are what actually printed; correlation is Pearson r on the closing path across the sessions after each event — the same measurement the terminal runs when it overlays any two events in the archive.
Every pairing above is one query in the terminal. When the next headline breaks, pull its analogues, overlay the paths, and test the strategy against exactly this history — sizing, stops, holding period — before committing capital.
Backtest against historyAnalyse live news against ninety years of archives.
When a headline lands, the engine narrows the archive to prior events of the same kind and ranks them on the state of the market each one landed in. Two archive queries below, each joined to the precedent the engine ranked first — and the shape similarity measured on the paths that followed.
"FOMC policy statement: March 15-16, 2022"
"U.S. to Restore Sanctions on Iran, Deepening Divide With Europe"
Shape similarity is Pearson r on the two post-event closing paths (n = 21 closes, t₀ to +20 sessions). It says how alike the two drawn lines are — not a probability that the move repeats. Across real pairs it is centred near zero — most matches do not rhyme this closely, which is what makes these two worth showing. Every figure is the engine's own output, unedited.
What moved the price today — and how far the shock travels.
Four instruments read a session out of the record: which kind of shock moved the price, how wide the band around it has to be to hold, which past days the market resembles, and which markets a shock has measurably reached. Every figure below is pulled from the live database, with its n.
The Daily Decomposition
What moved the price on a given day, split into supply, demand, risk and what the model cannot place. Read off a model, not off the tape.
Supply carried the largest share of the +2.72% move on 7 Jul 2026 — 24% of it, +0.64 percentage points. 16% stays unclassified, and the panel prints that rather than spreading it over the other three.
Model-based decomposition — a sign-restricted SVAR, not ground truth. Bar width is the posterior median share of the day's move; the figure on the right is that shock's median contribution in percentage points. Medians are taken shock by shock, so they do not add up to the printed return. Model 3da9bf62, run 11, fitted 7 Sep 2026.
Risk Bands
A range with a stated hit rate, checked against what actually happened. Never a point prediction.
of outcomes landed inside the band that advertises 80%. Checked on 241,405 realised outcomes. That is 4.1 points short of what it advertises, so the bands are currently too narrow — printed here rather than tuned away.
The instrument prints an interval, never a single number, and the audit above is the only claim made for it: how often the truth fell inside. Model 944abd49, run 3, computed 7 Sep 2026.
The Analog Engine
Which past days this market most resembles, and whether those days carry any information at all.
Each day is matched to its 8 nearest past days. Tested on 169 query days against 2,000 shuffled draws: at one week ahead the retrieved days carry information; at one day ahead they do not, and the panel says so.
A resemblance is not a forecast. The verdict rows are the honest part: the retrieved days are compared with shuffled ones, and a horizon that fails the test is printed as failing. Metric analog-v1-b7dc9e15fe564236, run 6, run at 7 Sep 2026.
Cascade
Which markets a shock has measurably reached. Every link is measured from the record, or reported as not measurable — never asserted.
A link earns its place by being measured across the episodes, or it is printed as unmeasured. The panel counts links; it makes no claim about how fast one travels. Fit d917e299, run 5, fitted 7 Sep 2026.
Four instruments are live in the terminal; the psychology exhibits are still in build. Every figure reports its n, or says the record does not cover the question yet.
Sign up for beta testEight cases, as the engine ranked them.
Each card is one archive record: the day the news landed, its classification, the benchmark it is measured on, and the closes that followed — drawn, not described. Four are recent archive queries; four are the rank-one precedents the engine returned for them.
“Over 20 sessions” is the close twenty sessions after t₀ against the close of the session the news landed in; “next session” is the first close after it. Both are recomputed from our own daily candles and printed as they come — including the flat ones. Nothing here forecasts.
The Classified Event Archive
Veille catalogues occurrences according to strict structural profiles, ensuring metadata is comparable over ninety years.
Review historical evidence before committing capital.
Stop trading based on qualitative narratives. Gain institutional access to ninety years of parsed macro precedent data joined with session-by-session pricing paths.