xau-ai-trading-bot/CLAUDE.md

CLAUDE.md — XAUBot AI

Project Overview

XAUBot AI is an automated XAUUSD (Gold) trading bot that combines Machine Learning (XGBoost), Smart Money Concepts (SMC), and Hidden Markov Model (HMM) regime detection. It runs on MetaTrader 5 via an async Python loop, executing trades on M15 candles.

Directory Structure

.
├── main_live.py              # Main async trading orchestrator
├── train_models.py           # Model training script
├── src/                      # Core modules
│   ├── config.py             # Trading configuration & capital modes
│   ├── mt5_connector.py      # MetaTrader 5 connection layer
│   ├── smc_polars.py         # Smart Money Concepts (Polars-based)
│   ├── ml_model.py           # XGBoost trading model
│   ├── feature_eng.py        # Feature engineering (37 features)
│   ├── regime_detector.py    # HMM market regime detection
│   ├── risk_engine.py        # Risk calculations & validation
│   ├── smart_risk_manager.py # Dynamic risk management
│   ├── session_filter.py     # Trading session filter (Sydney/London/NY)
│   ├── position_manager.py   # Open position management
│   ├── dynamic_confidence.py # Adaptive confidence thresholds
│   ├── auto_trainer.py       # Auto-retraining pipeline
│   ├── news_agent.py         # Economic news filtering
│   ├── telegram_notifier.py  # Telegram alerts
│   ├── trade_logger.py       # Trade logging to DB
│   ├── utils.py              # Utility functions
│   └── db/                   # Database schemas
├── backtests/                # Backtesting scripts
│   ├── backtest_live_sync.py # Main backtest (synced with live logic)
│   └── archive/              # Old backtest versions
├── scripts/                  # Utility scripts
│   ├── check_market.py       # Quick SMC market analysis
│   ├── check_positions.py    # View open positions
│   ├── check_status.py       # Account status check
│   ├── close_positions.py    # Close all positions
│   ├── modify_tp.py          # Modify take-profit levels
│   └── get_trade_history.py  # Pull trade history from MT5
├── tests/                    # Test scripts
│   ├── test_modules.py       # Module integration tests
│   ├── test_mt5_connection.py# MT5 connection test
│   └── test_risk_settings.py # Risk settings test
├── models/                   # Trained models (.pkl)
├── data/                     # Market data & trade logs
├── docs/                     # Documentation
│   ├── arsitektur-ai/        # Architecture docs (23 components)
│   └── research/             # Research & analysis files
├── web-dashboard/            # Next.js monitoring dashboard
├── docker/                   # Docker configuration
└── logs/                     # Runtime logs

Key Commands

# Run the live trading bot
python main_live.py

# Train/retrain ML models
python train_models.py

# Run backtest with threshold tuning
python backtests/backtest_live_sync.py --tune

# Run backtest with specific threshold
python backtests/backtest_live_sync.py --threshold 0.50 --save

# Run module tests
python tests/test_modules.py

# Check market status
python scripts/check_market.py

Architecture

The bot runs an async candle-based loop on M15 timeframe:

1. Data Fetch — Pull OHLCV from MT5, convert to Polars DataFrame
2. Feature Engineering — Calculate 37 technical features (RSI, ATR, MACD, Bollinger, etc.)
3. SMC Analysis — Detect Order Blocks, Fair Value Gaps, BOS, CHoCH
4. Regime Detection — HMM classifies market as trending/ranging/volatile
5. ML Prediction — XGBoost outputs BUY/SELL/HOLD with confidence score
6. Entry Filtering — 11 entry filters must pass (session, regime, spread, cooldown, etc.)
7. Risk Sizing — ATR-based SL, dynamic position sizing, Kelly criterion
8. Trade Execution — Send order to MT5 with broker-level SL/TP
9. Position Management — 10 exit conditions (trailing SL, time exit, regime change, etc.)
10. Logging — Trade logged to PostgreSQL + Telegram notification

Tech Stack

• Python 3.11+ — Main runtime
• Polars — Data engine (not Pandas)
• XGBoost — ML model for signal prediction
• hmmlearn — Hidden Markov Model for regime detection
• MetaTrader5 — Broker connection
• asyncio + aiohttp — Async execution & HTTP
• loguru — Structured logging
• PostgreSQL — Trade database
• Next.js — Web dashboard (optional)

Configuration

All secrets in .env:

• MT5_LOGIN, MT5_PASSWORD, MT5_SERVER, MT5_PATH — Broker credentials
• TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID — Notifications
• CAPITAL — Trading capital
• SYMBOL — Default: XAUUSD

Capital modes auto-configure risk parameters:

• MICRO (<$500): 2% risk/trade
• SMALL ($500-$10k): 1.5% risk/trade
• MEDIUM ($10k-$100k): 0.5% risk/trade
• LARGE (>$100k): 0.25% risk/trade

Important Notes

• All data processing uses Polars, not Pandas
• The bot targets < 50ms per loop iteration
• Models are stored as .pkl files in models/
• Backtest logic is synced with live (backtest_live_sync.py mirrors main_live.py)
• Scripts in scripts/ and tests/ include sys.path fix so they work from any directory