Every execution method inDocumentation Index
Fetch the complete documentation index at: https://mintlify.com/tripolskypetr/pump-anomaly/llms.txt
Use this file to discover all available pages before exploring further.
PumpMatrix produces a TradeSignal — a single, flat, immediately actionable object. The design principle is that production code reads the flat executable fields (symbol, direction, exit) and calls openPosition without branching. Provenance, audit information, and detector metadata live in the nested origin object, which is for logging and tracing, not for conditional logic.
backtest() and planForAt() extend this contract with a result field containing the realized PnL from a forward replay of the exit plan over closed historical candles.
plan() and signals() return TradeSignal — no result field, because the position has not closed yet. backtest() and planForAt() return BacktestSignal, which extends TradeSignal with result: BacktestResult.TradeSignal
The unified executable output ofsignals() and plan(). One signal equals one decision. The direction field is always the final executable direction — inversion has already been applied when action === "invert".
Trading pair ticker, e.g.
"SOLUSDT", "BTCUSDT". Passed through from the source ParserItem.Final executable direction. If the cascade detector fired
squeezePolicy: "invert", this field already contains the flipped direction — the original channel direction is preserved in origin.invertedFrom. Production code always executes direction without checking for inversion.The single discriminator for this signal’s character. See the Action Values section for full semantics.
Signal timestamp in Unix milliseconds. For a matrix burst, this is the timestamp of the burst itself; for single mode, it is the post publication timestamp.
Lower bound of the entry zone from the source
ParserItem. If undefined, enter at market (open of the first candle). Used by backtest() and planForAt() as the lower bound for the entry-zone scan.Upper bound of the entry zone from the source
ParserItem. If undefined (or equal to entryFromPrice), the zone collapses to a single price point. Pairs with entryFromPrice to define the live order range.Flat, ready-to-use exit plan. All fields are absolute values suitable for passing directly to
openPosition. See ExitPlan below.Provenance object for audit and tracing. Not intended for production branching — all branching decisions are already reflected in
direction, action, and exit. See SignalOrigin below.ExitPlan
The flat executable exit plan embedded in everyTradeSignal. All five fields are required and immediately usable by your position management code. When action === "tighten", trailingTake is already reduced by the trained tightenFactor (default 0.5).
Trailing take percentage — the maximum pullback from the peak PnL that triggers an exit. Measured from peak once
currentProfit >= 0. If action === "tighten", this value is already multiplied by tightenFactor (typically halved) to exit before the reversal. Unit: percent (e.g. 1.5 = 1.5%).Hard stop percentage — a fixed loss threshold from the entry price that immediately closes the position. For a long, this fires when price falls
hardStop% below entry; for a short, when it rises hardStop% above. Unit: percent. The realized PnL on a hard-stop exit is the honest -hardStop% loss — the library never rolls back to the last positive peak on a stop-out.Maximum position lifetime in minutes — the empirical impact horizon of the channel signal. The position is closed at the life-cap regardless of current PnL. Corresponds to
staleMinutes in the training grid; tuned by cross-validation, not hardcoded. Exits at the close of the last candle in the window.Peak-staleness profit threshold. The staleness timer is only armed once the peak PnL has reached this percentage. Below this threshold the position is not considered stale regardless of how long it has been open. Unit: percent (e.g.
1.0 = 1%).Peak-staleness window in minutes. Once the peak has exceeded
stalenessSinceProfit, if the position goes this many minutes without registering a new peak (price stagnates below the previous high), it exits — even if the hard stop and trailing take have not fired. Captures slow-decay situations where the momentum has died but price never quite reverses hard enough to hit the trailing take.SignalOrigin
The provenance record embedded in everyTradeSignal. Fields here describe where the signal came from and how the exit was resolved. This object is for auditing, tracing signals back to source posts, and debugging — not for branching in production code.
Which detection mode produced this signal.
"matrix" means the signal came from a synchronized burst across ≥ 2 independent author clusters. "single" means the signal came from a single channel post (single-channel fallback mode).Source channel name for single-mode signals.
null for matrix-mode signals, which are cross-channel by definition (no single owner).If this signal was inverted (the cascade detector fired
squeezePolicy: "invert"), this field holds the original channel direction. The signal’s direction is already the flipped direction. null means no inversion occurred. This is not a duplicate of direction — direction is what to execute, invertedFrom is what the channel said.The level of the exit tensor from which the
exit plan was resolved. The tensor uses hierarchical fallback:"cell"— exact match for[mode][channel][symbol][direction][volRegime]"symbol-dir"— collapsed acrossvolRegime;[mode][symbol][direction]"mode"— collapsed across channel, symbol, and direction"global"— root fallback, trained across all data
Volume regime detected at entry.
"anomalous" means the entry candle’s volume z-score exceeded the trained volZThreshold — the crowd entered on leverage (fuel accumulated). "calm" means normal volume. null when called via signals() (no candles provided) or when candle data was unavailable.Burst sharpness score from the detector, in
[0, 1]. For matrix mode: (clusters / channels) × (saturation relative to minClusters × 2). For single mode: a fixed baseline value. Higher means more independent sources converged clearly.Number of independent author clusters that contributed to this signal. Always
1 in single mode. In matrix mode, this is the count of deduplicated clusters (echo channels from the same author collapse to one cluster).The model’s training-time
confidence score (0..1) — carried from model.confidence. Reflects how much data was available and how stable and significant the edge was during training.The model’s training-time
reliable flag — carried from model.reliable. false means the model is operating with insufficient training data and its signals should be treated cautiously.The
id of the anchor ParserItem that created this signal — normalized to a string even if the original was a number. Useful for tracing a live signal back to the exact source post. undefined if the ParserItem had no id.All
ParserItem ids that were folded into this signal burst. In matrix mode a burst may include multiple posts; ids contains all of them. In single mode, typically a one-element array matching id.BacktestSignal
BacktestSignal is the output type of backtest() and planForAt(). It extends TradeSignal with a result field containing the realized outcome of replaying the exit plan forward over closed historical candles.
exit plan, and the entry zone from the source ParserItem (entryFromPrice / entryToPrice), falling back to the first candle’s open if no zone is specified. The replay is path-aware: it sees the full OHLC path, not just open and close, so a wick into a stop-hunt trap produces a negative PnL even if the candle ultimately closes positive.
BacktestResult
The realized result of replaying the exit plan forward. Only present onBacktestSignal — plan() and signals() do not carry this field because the position is still open.
Whether the replay found an entry.
false if the entry zone was never touched on the candle window (the price never reached entryFromPrice..entryToPrice). When false, all price and PnL fields are 0.Realized PnL as a fraction — e.g.
0.05 = +5%, -0.02 = −2%. For a hard-stop exit, this is the honest -hardStop% loss. The peak is tracked separately; the PnL is the actual realized value, never rolled back to the last positive peak on a stop-out.Peak PnL over the position’s lifetime as a fraction. Always ≥ 0 when entered. Useful for computing trailing-take efficiency — how close the actual exit was to the theoretical maximum.
Why the position was closed. Values from the simulation:
"hard-stop"— price hit the hard stop threshold"trailing-take"— pullback from peak exceededtrailingTake"peak-staleness"— position went stale after peak profit threshold was reached"life-cap"— impact horizon expired"cascade-veto"— the cascade detector firedsqueezePolicy: "veto"and the entry was blocked (entered: false)"no-entry"— no candle in the window touched the entry zone (entered: false)"no-candles"— no candle data was available at all (entered: false)
Minutes from entry to exit.
0 if not entered.Actual entry price. For entry within a zone, this is the candle
close if it falls in [entryFrom, entryTo], otherwise the clamped midpoint of the entry zone to [low, high]. 0 if not entered.Price at which the position was closed (producing the realized
pnl). 0 if not entered.true if there were not enough candles after entry to cover the full impactHorizonMinutes. A truncated measurement is kept in the output (unlike during training, where truncated entered trades are dropped to avoid comparing a 24h horizon against a 1h one). Check this flag if reason is "life-cap" and heldMinutes seems shorter than expected.Action Values
Theaction field is the only discriminator needed to understand a signal’s character. Production code typically branches on action only to apply routing logic — the direction and exit are always already correct.
"enter"
A standard entry in the direction the channel posted. The cascade detector either was not active (no candles), found no anomaly, or found an anomaly with squeezePolicy: "none" or "ignore". Open a position in direction with exit.
"invert"
The cascade detector fired and squeezePolicy: "invert" was selected. The crowd entered in one direction, accumulated leverage, and is now vulnerable to a liquidation cascade. The signal flips the direction to trade against the channel post — enter in direction (already flipped), which is the opposite of what the channel said. origin.invertedFrom records the channel’s original direction for audit purposes.
"tighten"
The cascade detector fired and squeezePolicy: "tighten" was selected. Enter in the same direction as the channel posted, but with a tighter trailing take — exit.trailingTake is already reduced by the trained tightenFactor (default 0.5×). The tighter trail exits the position earlier, before the expected reversal from the cascade consumes the gains.