Not an indicator. Five engines read the recorded order flow, the book, and open interest, and each states what it sees and the evidence behind it — so you can disagree with the reading rather than trust it.
Every input the engines read, with its current value and where that value sits in this symbol's own recorded history. A spread at the 97th percentile is wide for this symbol, which is the only comparison that means anything across coins that differ by orders of magnitude.
Rules over the recorded stream. Each detector's thresholds are stored on every row, so an episode found today stays interpretable after the definition changes.
Among episodes of the same kind, comparing the pre-episode state of those followed by a large move against those that went nowhere. Split within each symbol, so a difference cannot come from one coin being more liquid than another; Mann-Whitney on ranks, Benjamini-Hochberg across the whole scan, and a stability check on both halves. Level variables like price and dollar depth are excluded — inside a short sample they are proxies for when, not what.
Probability of a future event, never a price. Walk-forward folds with an embargo between train and test, so no fold scores itself on overlapping labels. Nothing is published unless it beats both the base rate and a naive rule on every fold.
Bybit taker fee both ways plus half the recorded spread per side. When stop and target fall in the same second the stop is assumed to hit first — a one-second bar cannot order them. Random entries at the same rate are the null to beat.
The research collector publishes every few seconds. If this persists, the collector or the feature builder is not running.