A search engine for the market's past, read like case law: bring a moment and read the precedents, each with what was known at the time and what happened next.

Three ways to ask

• Moments like this one. The past moments whose preceding stretch looked most like the one you picked: now, a date, or an event.
• An asset after a kind of event. Every halving, every rate cut, every CPI print, each with its own before and after, listed one by one.
• Last time. The moments that met a condition — a move of a given size, a level crossed — most recent first. Asked through the API, which is where a question in words belongs: a field that invites a sentence has promised to read sentences, and this one reads a handful of shapes.

Before and after

Before is measured on the bars that had closed by the moment, and names only the events already public then. Each event is stamped with the time it became public, from its source; one known only by its date counts from the end of that day. After starts at the moment itself. Moments are compared by their before alone, and a past moment becomes a candidate only once its 32 hours after have closed.

How matching works

Moments are compared by the shape of the price over the preceding stretch, as the log path of the closes. The window is always 144 bars. What a bar means depends on the source, so the stretch does too, and the chips over each chart say which window is being drawn.

• A source that gives intraday bars — the two crypto pairs — is compared on four windows at once: 12 hours of 5-minute bars, 36 hours of 15-minute, 3 days of 30-minute and 6 days of hourly. Each is normalised on its own, so each answers for a quarter of the comparison and the longest does not swamp the rest.
• A source that gives one value a day has one window: 144 of its days. That is about seven months of calendar rather than 144 days, because a weekend carries no number — so a precedent for the ten-year is a seven-month shape, where a precedent for bitcoin is a six-day one. The reading beside the chips is measured off the series itself, not multiplied.

A window must also cover the stretch it claims. Where the market was shut inside it, or where the moment itself falls in a shut stretch, there is no answer and the page says which of the two it was rather than calling it a shortage of history. The sentence above each result compares how widely the matches' 32 hours after ranged against random past moments: a width, never a direction. Every API result names the model and its version.

What we don't do

No forecasts, no directions or probabilities, no buy or sell signals, no picks. Results are listed one by one and never averaged. Past similarity does not guarantee future outcomes.

Data and sources

Every series below is named with the identifier it is fetched by, so an answer can be checked against its source rather than taken on trust.

• Crypto — 5-minute bars from Bitstamp: BTC/USD from 2013 and ETH/USD from 2018, closed bars only. Bars where nothing traded are carried forward by the exchange and the charts draw those stretches faintly.
• Rates, the dollar, energy — one value a business day, from the agency that publishes it, through FRED: the 2-year and 10-year Treasury (DGS2, DGS10), 10-year breakeven inflation (T10YIE), the broad dollar index (DTWEXBGS), WTI crude at Cushing (DCOILWTICO, EIA) and Henry Hub natural gas (DHHNGSP, EIA). The St. Louis Fed financial stress index (STLFSI4) is weekly, so it is read beside a moment and never searched as one.
• Events — FOMC statements and decisions from the Federal Reserve, scheduled releases from the agency that issues them, and the crypto calendar from contract specifications. Each event carries its own source link on its page, and a revised figure keeps the time and the value of the first print.
• News — press releases from the Federal Reserve, the SEC, the CFTC, the White House and the Justice Department, stamped with the time each was first seen, with the earlier years filled from each body's own archive.

What is not here

Equities, gold, silver, the volatility index and credit spreads are not searchable, and the reason is licensing rather than difficulty. An exchange index and a volatility index are licensed by their owners; a vendor's terms reach derived values as well as delivered ones, which a percent path is. Rather than serve them on a reading nobody has confirmed, they are held back until a licence is in hand. The same test is applied to everything above: a source whose terms do not permit redistribution is not served, whatever is on disk.