🔮 Crypto Oracle MCP — Quantitative Spot Trading Intelligence

A production-grade quantitative trading system using machine learning to generate edge in crypto spot/perp markets. Built on XGBoost with a 130-feature stack combining microstructure signals, technical analysis, and on-chain metrics.

---

⚡ Quick Start — Train V6 Models on Google Colab

Everything runs in one command. No local setup needed.

Step 1 — Upload training data to Google Drive

Create this folder in your Google Drive:

MyDrive/crypto_oracle/processed/

Upload your 20 *_training_data.parquet files there.

Step 2 — Open Colab with GPU

→ Open Google Colab · Runtime → Change runtime type → T4 GPU

Step 3 — Clone the repo and run

Paste this into a Colab cell and run it:

!git clone https://github.com/haitham-kh/crypto-oracle-mcp.git
%cd crypto-oracle-mcp
!python colab_v6_train.py

That's it. The script automatically:

• ✅ Installs all dependencies
• ✅ Mounts your Google Drive
• ✅ Downloads 24 months of Binance 1m klines for all 20 coins (free, no API key)
• ✅ Computes 36 V6 technical features per coin
• ✅ Trains V6_full (130 features) + V6_micro (36 features) on GPU
• ✅ Saves trained models to MyDrive/crypto_oracle/models_v6/

Step 4 — Copy models to your machine

Download MyDrive/crypto_oracle/models_v6/ from Drive and copy all files into:

crypto-oracle-mcp/data/

Step 5 — Run the live engine

python live_demo_engine.py

On startup you'll see: [V6] models loaded — blend: V5×30% + V6×70%

---

🏗 Architecture

V6 Model Stack (130 features)

Group	Features	Tier	What it captures
Donchian Channels	6	S+	Structural range, breakouts
Anchored VWAP	3	S	Institutional fair value (day/week/month)
Volume Profile VAH/VAL	4	S	Auction-market value area (70% of volume)
Liquidity Sweeps	5	S	Stop-hunt detection via wick rejection
Order Flow (OFI/CVD)	4	A	Large-trade delta, taker pressure
SMA Stack	6	A-	20/50/200 trend alignment
Fibonacci Levels	5	B	38.2 / 50 / 61.8 / 78.6 confluence
Bollinger Extras	3	B	Squeeze / expansion regime
V5 Base	94	—	Hurst, Perp microstructure, VPVR, cross-sectional

Inference Blending

p_final = 0.30 × p_v5  +  0.70 × p_v6_full

Trade fires only if both p_final ≥ threshold AND p_v6_micro ≥ 0.50 (confirmation gate).

Training Pipeline

Colab: colab_v6_train.py
  ↓ downloads 1m klines (Binance Futures API, free)
  ↓ computes 36 V6 features inline
  ↓ joins with pre-computed V5 features (your parquets)
  ↓ trains XGBoost on GPU with recency + V6-strength weighting
  ↓ calibrates probabilities (Isotonic Regression)
  ↓ auto-selects threshold via EV maximisation
  ↓ saves to Google Drive

---

📁 File Structure

crypto-oracle-mcp/
│
├── colab_v6_train.py          ← THE MAIN COLAB SCRIPT (start here)
│
├── features_v6.py             ← 36 V6 feature definitions
├── features_v5.py             ← V5 base features (94)
├── features_v4.py             ← V4 base features
├── train_v6.py                ← Local training (if you prefer CPU)
├── train_v5.py                ← V5 training pipeline
│
├── live_demo_engine.py        ← Live trading engine (paper/live)
├── blind_backtest.py          ← Out-of-sample backtest
│
├── trading_config.py          ← Risk parameters, thresholds
├── regime_filter.py           ← Market regime classification
├── labels_multi_horizon.py    ← Triple-barrier labelling
│
├── fetchers/
│   ├── binance.py
│   ├── mexc.py
│   └── coingecko.py
│
├── data/                      ← Model weights (gitignored, from Colab)
│   ├── v6_full_clf_*.json
│   ├── v6_micro_clf_*.json
│   ├── v6_*_calib_*.pkl
│   ├── v6_meta.json
│   └── v5_clf_*.json          ← V5 models (fallback)
│
└── .env                       ← API keys (gitignored, never committed)

---

⚙️ Local Setup (optional — for live engine only)

git clone https://github.com/haitham-kh/crypto-oracle-mcp.git
cd crypto-oracle-mcp
pip install -r requirements.txt
cp .env.example .env
# Fill in your API keys in .env

Required API keys (.env)

BINANCE_API_KEY=your_key
BINANCE_API_SECRET=your_secret
MEXC_API_KEY=your_key
MEXC_API_SECRET=your_secret
CMC_API_KEY=your_key
COINGECKO_API_KEY=your_key

---

📊 Expected Performance (V6 target)

Metric	V5 Baseline	V6 Target
Test EV per trade	~+0.05%	>+0.15%
Win rate	~51%	>53%
Profit factor	~1.15	>1.40
Trades/day	~8–15	~6–12 (higher selectivity)

---

🔄 Workflow

Train (Colab)          Live Engine              Backtest
─────────────    →    ──────────────────    →   ─────────────────
colab_v6_train.py     live_demo_engine.py       blind_backtest.py
  GPU, ~25 min          V5×30% + V6×70%           OOS 6 months
  saves to Drive        auto-detects V6            prints PF, EV

---

🛡 Risk Controls

• LOOSEN_MODE = False — strict signal quality required
• Position sizing — Kelly + volatility-scaled, max 10% capital
• Stop loss — ATR-based, capped at 2.5%
• Regime filter — blocks trading in CHAOS / LOW_LIQUIDITY regimes
• Micro-gate — V6_micro must confirm before taking any trade

---

Built with Python, XGBoost, Polars, and Binance/MEXC APIs.