A database that learns
Every night EquationDB re-tests what it believes on the data that arrived since, keeps every result and writes down what changed. Confidence over time is evidence too.
Articles from the EquationDB team: how the data is built, how to use the language, and what we learn running it.
Every night EquationDB re-tests what it believes on the data that arrived since, keeps every result and writes down what changed. Confidence over time is evidence too.
A database answers when it is asked; a market does not wait. MarketEye puts the whole market on one screen, live, and every picture on it is a query you can open.
We tested 259 beliefs traders act on with one rule. 74 held. Breakouts lag, fear has paid, misses drift and beats no longer do.
Write what you believe about markets; EquationDB writes the test, runs it on decades of history and grades it with one fixed rule. The model writes the test, it never grades it.
Daily weather for 399 US stations and 55 world cities, back to 1869, on the same engine and in the same language as markets: degree days, anomalies, ten weather events scored against each station's own season, and what markets did after them.
Prices, statement lines, earnings surprises, insider buying, analyst actions, events, theme indexes and market stats now share one name and one shape, so any of them can be compared with any other: on one grid, from the date each value was known.
A new statement, SIMULATE, sets a behaviour for one or more instruments and reads how everything else responded when that happened before: distributions with sample sizes, the dates to check, model-implied numbers kept apart, and seeded paths. What it does, how analogues are matched, and where it refuses to answer.
WHY is a new statement that takes a price move and returns the explanations the data can support, each with its share of the move and the query that reproduces it, and then says how much is left unexplained.
A query answers one question once. An app is that query saved with a slot for the symbol and typed inputs such as dropdowns and sliders; a dashboard puts several on one screen and links them by colour, so one change of symbol moves every linked panel. Why we built them, how inputs stay safe, and what they mean for agents.
EquationDB ships with its own demonstration: a Quick tour and a full Demo sheet where every command has Copy, Open in Query and Run, with what you should see and how long it should take. We use the same sheet to test our releases. Eight statements across eight capabilities, explained.
A graph-derived universe, events scored for rarity, outcomes without hindsight, fundamentals and filings next to prices, cross-asset conditions and models, combined in single statements. Seven of the lead queries from the Demo, with the business question, how to read the result and why hindsight matters.
A new crypto market: every actively quoted token, about 2,000, a daily bar for every calendar day, 5-minute quotes around the clock, sub-cent tokens quoted per 1,000 or 1 million tokens, and Bitcoin next to the Nasdaq in one query.
A new fx market: the seven majors, their crosses and the exotics, with full daily history, 5-minute quotes from the Sunday open to the Friday close, and the majors in the live tier. Dollar indexes, volatility regimes and cross-asset studies are one query.
A taxonomy up to 8 levels deep (5,947 nodes), 244 themes in a hierarchy, 69,778 companies classified with typed AI answers, supply chains from descriptions and earnings calls, and an index series for every node and theme.
Write what you want in plain English, a query, or both. EquationDB plans it, checks every query, asks only when the answer changes the result, runs it under your entitlements and reports back — and you can steer it from the comments.
Every night EquationDB writes down what is unusual in its data, makes calls where history supports them, and scores every call afterwards — hits and misses alike, with no hindsight.
General-purpose databases make multi-symbol market questions slow: screens, cross-sectional ranks, event studies and point-in-time joins take seconds to minutes. We wanted a database that answers them in milliseconds, and one that can think. Here is what we built and why.
EquationDB speaks SQL on the same engine as its market language, and any indicator or fundamental is a column. A tour of finance aggregates, SAMPLE BY candles, named windows, QUALIFY, and CTEs and joins across symbols.
Every night EquationDB builds a graph of how the 500 largest US stocks and the reference ETFs move together: correlations, betas and intraday lead-lag. GRAPH finds true peers, leaders, clusters and the spread of a hypothetical shock, and flags relationships that break.
SEARCH lets EquationDB look for trading edges on its own, with guardrails that are always on: a train, validate and test split, Benjamini-Hochberg false-discovery control over every candidate tried, and stability checks across years and symbols.
EVENTS tells you what just happened, with rarity against each stock's own history. OUTCOMES tells you what usually happened next, against a same-period baseline, with a t-statistic. Here is how to run a proper event study in two statements, and the traps they are built to avoid.
We rebuilt every price in EquationDB from a single market data provider: about 70,000 symbols and 253 million daily bars, full history per symbol, checked against what we had before. Here is how, and what keeps it correct every night.
A screen is one line in EquationDB. This walk-through builds one step by step, from a simple filter to a ranked, sector-aware list, and then checks whether the setup has historically worked.
EquationDB never rewrites stored history when a split or dividend happens. Splits are applied as factors when bars are read; dividends are kept as per-bar factors and applied only when you ask for total return. Here is why.