Scalp New
LIVEML-driven short-horizon crypto scalping. LightGBM + 54 features, 35+ strategies, multi-venue execution across Bybit, Binance, MEXC.
Key Numbers
At a Glance
35+
Strategies
54
ML Features
3
Exchanges
<5ms
Decision Latency
Overview
About This Project
A production-grade machine learning platform engineered for short-horizon crypto scalping across perpetual futures markets. The system fuses a LightGBM ensemble with 54 hand-crafted microstructure features -- order flow imbalance, volume profiles, regime indicators, and cross-asset correlations -- to generate sub-5ms trading decisions.
At its core, 35+ modular strategies span momentum, breakout, absorption, trapped traders, failed auction, and volume profile patterns. Each strategy is gated by an ML scoring layer and routed through a cost-aware engine that selects the optimal exchange in real-time based on fees, spread, and available liquidity.
The full pipeline follows a rigorous promotion path: offline walk-forward validation with purged cross-validation, shadow trading with simulated fills, canary deployment on a subset of symbols, and finally live execution with hard drawdown gates. A two-stage pretrain/finetune model architecture ensures the system adapts to shifting market regimes without catastrophic forgetting.
Features
What It Does
54-Feature Microstructure Engine
Real-time feature pipeline with Rust PyO3 extensions computes order flow imbalance, volume profiles, regime detection, and cross-asset correlations in under 3ms.
Regime-Aware ML Gate
Two-stage LightGBM model with pretrain/finetune architecture detects market regime shifts and dynamically adjusts strategy selection and position sizing.
Multi-Strategy Engine
35+ pluggable strategies -- momentum, breakout, absorption, failed auction, trapped traders, volume profile breakout -- each independently backtested and shadow-validated.
Cost-Aware Venue Router
Smart order routing across Bybit, Binance, and MEXC selects the optimal venue per trade based on real-time fees, spread, and depth-of-book liquidity.
Shadow Validation Pipeline
Governance-driven promotion: shadow trading with paper fills, canary deployment on 3-5 symbols, then full rollout with hard drawdown and expectancy gates.
Walk-Forward Validation
Purged time-series cross-validation with embargo periods prevents look-ahead bias while maximizing training data utilization across retraining cycles.
Architecture
How It Works
Challenges
What Made This Hard
Maintaining strict feature compatibility across model retrains while continuously adding new signals required an immutable schema contract. Handling venue-specific quirks -- MEXC denominating quantity in contracts versus coins on Bybit -- without leaking exchange semantics into the shared ML pipeline demanded a clean abstraction layer. Achieving sub-5ms end-to-end decision latency in a Python-dominant stack required surgically offloading hot paths to Rust while keeping the system debuggable.
Stack