EquationDB

← Back to all articles

A database that notices — and keeps score

2 October 2026 · EquationDB team

Most databases wait to be asked. EquationDB now also looks through its own data every night and writes down what is unusual — and then, crucially, checks whether it was right.

What a finding is

After the close, six detectors read every market's new session. Each one looks for a different kind of "that's not normal":

  • an event with an edge — a crossover, gap or breakout fired on a stock whose own history (or the top500's) shows a strong, statistically significant move afterwards;
  • a multi-year extreme — volume, realized volatility or the distance from a 52-week high or low beyond the 99th (or below the 1st) percentile of the symbol's last three years;
  • a relationship break — a stock that tracked its sector ETF or closest peer for two years suddenly decouples;
  • a regime shift — breadth or volatility entering an extreme band of its own history;
  • a group move — a market, exchange, country, sector or industry moving more than four standard deviations as a group;
  • a new setup — the nightly SEARCH finds a condition that survives false-discovery control.

The candidates are ranked by signal size, rarity, sample size, participation and liquidity, merged per symbol, capped, and written up. Every finding carries the numbers behind it and a query that reproduces it — press Run and you are looking at the evidence yourself.

Calls are recorded before the outcome

Some findings make a call: this stock has tended to beat its normal 5-day return after this event. The call is written down with its direction, horizon, baseline and entry price before anyone knows the outcome, and it can never be edited. When the horizon passes, it is scored: total return against SPY (or the equal-weight exchange outside the US). Hits and misses both stay on the record. A stock that stopped trading is scored on its last bar; a call that cannot be scored is counted, not hidden.

The track record

The Track record page shows, per detector: how many calls, the hit rate with a 95% confidence interval, the average excess return in the call's direction, and whether the last 90 days look like the past. A detector whose recent record breaks down is demoted automatically — it sinks in the ranking and is flagged "stopped working" — and comes back when it recovers. The weights and the reason for each are on display.

No hindsight

The hard part of any track record is not cheating. A finding about a given day may only use data that existed at that day's close: percentiles from the three years before it, event outcomes only where the forward window had already closed, correlations ending that day. We replayed the detectors over the last three years the same way — strictly as of each past session — and show that replay separately from live calls, so nobody mistakes a backtest for a forecast. Automated tests inject a future price spike and check that past findings do not move.

Try it

SHOW FINDINGS
SHOW TRACK RECORD

Or open Findings in the app. The full rules of every detector are in the docs under Findings & track record.