Artale/quant-finance-arb-signal-v1

Quant Signal Classifier v0.1.0

Real-time arbitrage signal quality scorer. Trained on 30,743 labeled trades from a live 27-service quant pipeline running on 2 VPS servers across 4 crypto venues.

How It Works

The model takes 6 features from each arbitrage signal and predicts whether it's worth trading:

Architecture: Random Forest classifier (200 estimators, max depth 6). Chosen for interpretability and fast CPU inference (<1ms per signal).

Quick Start

import json
import numpy as np

# Load model (safe JSON format)
with open("quant_model_safe.json") as f:
    model = json.load(f)

# Score a signal
signal = {
    "spread_bps": 50,
    "confidence": 0.7,
    "type": "funding_arb"
}

features = np.array([[
    signal["spread_bps"] / 100,
    signal["confidence"],
    0.3,  # irr
    1.0 if signal["type"] == "funding_arb" else 0,
    0.0,  # is_spread_arb
    0.0,  # is_btp_basis
]])

# Feature importances tell you which signals matter most
print("Feature importances:", model["feature_importances"])

Training Data

• —Source: 168K+ trade records from live HyperLiquid/Binance/Bybit pipeline
• —Labels: 30,743 closed trades with realized PnL
• —Coverage: 231 coins, 1,475 funding rates, 4 venues
• —Period: June-July 2026

Live Deployment

This model runs as a systemd service in production, scoring signals every 60 seconds:

HMM Regime Detector → Markov Bias → Quant Scorer → Auto-Trader → Paper Executor
      (port 4990)       (port 4989)   (port 4992)   (port 4994)   (port 4000)

27 services, fully automated, CI/CD retrains daily.

Files

Performance

License

MIT