Every move is split into the part its peer group explains and the part left over. The leftover — the residual — is scored against that stock’s own history, not a fixed percentage. So a 2% move on a day the sector ran 2% says nothing, and a 1.2% rise while its peers fell 2.6% is the one you get told about.
The demo needs no signup: a real 12-stock NSE watchlist, read-only, with the decomposition and the signals live.
Of the 3,877 alerts the engine drops, 2,140 go because the whole market was moving, 1,094 because the residual never cleared the floor, and 643 because the stock routinely moves that much anyway. It is not a quieter version of the same list — it is a different one.
Inspect the noise-reduction evidenceEach stock is measured against a behavioural peer cluster where one exists, and a beta-adjusted Nifty where it does not. What survives the peer move is the residual, scored in units of that stock's own residual volatility — so the bar is set by how quietly it usually trades, not by a number someone picked.
EXCESS_MOVE, VOLUME_CONFIRMED, GAP, LEVEL_BREAK, SINCE_SEEN_MOVE. Each carries the numbers it fired on: a gap clears 2%, volume confirmation needs 1.5× the stock's normal, a move since you last looked has to clear 1.5% and twice the drift expected over that gap. You read why, not just that.
Describe a trigger — "drops more than 2% against its peers on 3x volume" — and it compiles to a bounded JSON condition you read and confirm before it saves. You can open the compiled rule. The model writes it; the engine, never the model, evaluates it.
The evidence page replays all three alert rules over the same stored bars and breaks the gap down: market-wide, below the floor, or inside the stock's own noise. It also lists the six days a percentage rule misses — raw moves under 1.8% whose peer-adjusted residual was over 2.8%.