Skip to main content

Documentation 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.

Pump Anomaly is a TypeScript library that turns a stream of Telegram trading recommendations into statistically certified, ready-to-execute trade signals. It separates real capital inflow (independent authors hitting the same ticker in sync) from single-actor manipulation, detects stop-hunting traps via liquidation-cascade analysis, and produces a complete exit plan trained on your actual production exit — not close-to-close approximations.

Quickstart

Train a model, load it in production, and execute your first signals in minutes.

How It Works

Understand the five-layer detector pipeline and how signals are validated.

PumpMatrix API

Full reference for fit, load, signals, plan, backtest, and all model properties.

Per-Asset Grids

Pre-tuned TrainGrid configurations for 15 cryptocurrencies from BTC to Fartcoin.

​
What it solves

Pump Anomaly addresses three distinct problems that naive close-to-close signal detection misses:
1

Separate real inflow from single-actor manipulation

A single actor running 10 anonymous Telegram channels looks like synchronized activity. Pump Anomaly’s author-cluster deduplication (union-find over cross-correlation) collapses sibling channels into one author before counting independent votes — so a true multi-author burst is required for a matrix-mode signal.
2

Separate pumps from stop-hunting traps

A wick into leveraged positions triggers a liquidation cascade. Close-to-close labeling misses this — the candle closes positive but the position was stopped out on the wick. Path-aware exit replay on 1-minute OHLC data labels every training sample using the same trailing-take / hard-stop / life-cap logic your production code runs.
3

Produce statistically certified, executable plans

Grid search over thousands of configurations is guaranteed to find a “best” config even on pure noise. Pump Anomaly applies five independent statistical barriers (Deflated Sharpe, PBO, SPA, minTRL, nested-CV OOS) after training to distinguish real edges from brute-force artifacts — and refuses to trade when the evidence doesn’t hold up.

​
Key features

Author-Cluster Deduplication

Union-find clustering over lag cross-correlation collapses sibling channels. Only independent author clusters vote on matrix-mode signals.

Path-Aware Exit Replay

Training labels come from simulating your exact production exit (trailing take, hard stop, life-cap) on 1m candles — not two price points.

Liquidation Cascade Detection

Symmetric volZ + squeezePressure analysis detects stop-hunting on both long and short positions. Four configurable policies: none, tighten, veto, invert.

Statistical Certification

Five-barrier gate (DSR, PBO, SPA, minTRL, nested OOS) built from López de Prado and White’s methodology. certified: false means don’t trade.

Exit Tensor with Fallback

Exit parameters trained separately per [mode][channel][symbol][direction][volRegime] cell with hierarchical fallback — no magic constants.

Meta-Overfitting Guard

MetaLedger tracks every fit attempt. Family-wise DSR correction and a cadence gate prevent “click your way to a certificate” failure modes.

​
Quick example

import { PumpMatrix } from "pump-anomaly";
import * as fs from "fs";

// Train once on historical Telegram signals
const model = await PumpMatrix.fit(history, getCandles);

// Only certified models should trade
if (!model.certification.certified) {
  console.warn("Edge not proven:", model.certification.reasons);
  process.exit(1);
}

// Serialize for production
fs.writeFileSync("model.json", model.save());

// In production — no retraining needed
const model = PumpMatrix.load(fs.readFileSync("model.json", "utf8"));
const trades = await model.plan(liveItems, getCandles);

for (const s of trades) {
  openPosition(s.symbol, s.direction, { from: s.entryFromPrice, to: s.entryToPrice }, s.exit);
}

Training Guide

Learn grid search, K-fold CV, and the one-standard-error winner selection rule.

Certification

Understand all five statistical barriers and when a model is safe to trade.

Live Trading

Use plan() for look-ahead-free live decisions with cascade detection.

Backtesting

Replay signals over closed history to measure realized PnL with backtest().

Build docs developers (and LLMs) love

Get started for freeTalk to us