How the engine reads, and where it stops.
Every exhibit carries its data date and its sample size. This page carries everything else.
How it reads
ORION matches windows of price history on two axes: the shape of the tape, scored with a 0–100 similarity index over a configurable lookback, and the macro regime behind it, scored across seven series including rates, curves, the dollar, and commodity ratios. Matches above the quality bar become analogs.
For each analog the engine reports what actually followed: forward returns at +5, +21, +63, +252 and +504 sessions, the maximum drawdown along the path, and the consensus, the pointwise median of all analog paths. The unconditional all-history baseline is printed beside every conditional figure, so a reader can always see whether the setup carries information or just repeats the market's default drift.
Where it stops
Samples are small by construction; five analogs are five, not five hundred. Regimes drift, and a pattern that held for two decades can stop holding without notice. Past distributions do not bind the future, and the engine will say "no honest read" rather than stretch a thin sample. Reads flagged THIN carry fewer than ten instances and should be read as anecdote, not evidence.
Data sources
Refresh cycle
The record is rebuilt after every close. The nightly scan runs at 16:00 ET plus settlement, checks roughly three hundred conditions, and regrades every tracked read. Each exhibit shows the date of the data it was drawn from.