Crypto Liquidation Feed

Browser-based dashboard for live liquidations, funding rates, clustering, replay, and anomaly tracking.

Live Website

What It Is

This project tracks live liquidation activity across exchanges and pairs it with funding, clustering, replay, and anomaly tools. Everything runs in the browser.

Features

• Live liquidation stream for perpetual markets
• Exchange selection across Binance, Bybit, and OKX
• Client-side market stress meters for cascade risk, chaos, and anomaly
• Funding rates and cross-exchange funding spread view
• Replay controls with play, pause, seek, speed, and compare-to-live support
• Replay dataset capture, import, export, and browser-side management
• Shared minute-vector pipeline for live scoring and offline model training
• Optional offline-trained TensorFlow.js autoencoder scoring in the browser
• Responsive dashboard with audio alerts and inline help for market jargon

Supported Exchanges

• Binance
• Bybit
• OKX

Installation

git clone https://github.com/Mohamed1756/Crypto-Liquidation-Feed/
cd Crypto-Liquidation-Feed
npm install
npm run dev

Production Build

npm run build

Replay And ML Workflow

The replay pipeline is built around full timestamps and canonical minute vectors.

• Capture raw liquidation events into JSONL:

npm run replay:capture

• Build a replay dataset with dated timestamps and minute vectors:

npm run replay:build -- --input=./data/captures/binance-2025-01-01.jsonl

For legacy CSV files with time-only rows, pass an assumed date:

npm run replay:build -- --input=./my-file.csv --assume-date=2025-01-01

• Train a small offline autoencoder and emit a browser-loadable manifest:

npm run replay:train -- --input=./public/replay/liquidation-replay-dataset.json

By default the trained model is written to public/ml/liquidation-autoencoder.json, which the app can load at runtime.

Inside the app, imported datasets can be replayed locally with:

• play and pause controls
• minute-by-minute seek
• adjustable playback speed
• compare-to-live matching
• replay-driven stream, cluster, and canvas views

Notes

• The browser anomaly meter works without a shipped model by using the rolling statistical baseline.
• Shipping an offline-trained model improves regime recognition, but the replay dataset quality matters more than model complexity.
• Live ingestion, replay, and analytics all run client-side.