HFT Bybit+Binance
SHADOWHigh-frequency liquidity detection on Bybit with ML ranker for persistence prediction. 15 real-time collectors, 42 v2 feature columns, 200ms confirmation window. Execution via MEXC bridge.
Key Numbers
At a Glance
15
Collectors
42
Feature Columns
200ms
Confirmation
Cross-Exchange
Execution
Overview
About This Project
A high-frequency liquidation detection and execution system that identifies forced selling events on Bybit and routes execution through MEXC for optimal fee structure. The system runs 15 parallel data collectors streaming real-time market data, computing 42 v2 feature columns that capture the microstructure signature of liquidation cascades.
The ML persistence ranker goes beyond simple detection: it predicts how long a liquidation-driven price move will persist, enabling the system to size positions and set exit timing based on the expected duration and magnitude of the cascade. A 200ms confirmation window balances speed against false positive rate.
Cross-exchange execution via the MEXC bridge allows the system to detect events on Bybit where liquidation data is richer, while executing on MEXC where fee structures are more favorable for the trading strategy.
Features
What It Does
Liquidation Event Detection
Real-time identification of forced liquidation events using trade tape analysis, order book microstructure, and volume anomaly detection across 15 parallel data streams.
ML Persistence Ranker
Predicts duration and magnitude of liquidation-driven price moves, enabling intelligent position sizing and exit timing beyond simple event detection.
Feature Schema v2
42-column feature schema capturing temporal, volume, orderbook, and cross-symbol signals specifically engineered for liquidation cascade dynamics.
MEXC Bridge Execution
Cross-exchange execution architecture: detect on Bybit where liquidation data is richest, execute on MEXC where fee structure is most favorable.
Architecture
How It Works
Challenges
What Made This Hard
The 200ms confirmation window creates a fundamental speed-accuracy tradeoff: too fast and false positives dominate, too slow and the profitable window has passed. Engineering a feature set that achieves reliable classification within this constraint required extensive feature selection and temporal analysis. Cross-exchange execution adds latency that must be accounted for in the persistence prediction -- the model must predict moves that will still be available after the bridge routing delay.
Stack