BacktestingMCP

GPU-accelerated cryptocurrency backtesting platform with CLI, MCP server, and AI strategy generation. Optimized for NVIDIA RTX 4090 (CUDA 12.x) using CuPy + Numba (~1,145 optimization tests/sec).

---

Features

• 6 built-in strategy templates — MA Crossover, RSI Mean Reversion, EMA+RSI, Bollinger Bands, ATR Breakout, Momentum
• DCA strategies — fixed monthly or signal-based Dollar-Cost Averaging
• Breakout scanners — scan symbols for unusual-volume and breakout setups
• GPU optimization — CuPy + Numba JIT for parameter sweeps, automatic CPU fallback
• Data management — Binance/CCXT historical data download, SQLite storage, incremental updates
• Risk management — ATR/volatility/Kelly position sizing, daily/weekly loss limits, correlation tracking
• AI strategy generation — natural language → Python strategy code (OpenAI, Anthropic, Ollama)
• MCP server — 6 tools for AI assistants (VS Code Copilot, Claude Desktop, etc.)
• Full CLI — data, backtest, strategy, results commands

---

Quick Start

1. Set up the environment

# Windows
python setup_venv.py
.\venv\Scripts\Activate.ps1

# macOS / Linux
python setup_venv.py
source venv/bin/activate

2. Launch the interactive menu

python run.py

==================================================
  BacktestingMCP
==================================================
  1  Download market data
  2  List available data
  3  List strategies
  4  Run a backtest
  5  GPU optimization
  6  Generate AI strategy
  7  Start MCP server
  8  Check GPU status
  9  Inspect database
  a  Run market scanner
  b  Compare breakout strategies
  t  Run tutorial walkthrough
  q  Quit
==================================================

3. Or use the CLI directly

# Download BTC/USDT hourly data for 2024
python -m src.cli.main data download --symbol BTC/USDT --timeframe 1h --start 2024-01-01 --end 2024-12-31

# Run a backtest
python -m src.cli.main backtest run \
  --strategy moving_average_crossover \
  --symbol BTCUSDT --timeframe 1h \
  --start 2024-01-01 --end 2024-06-01 \
  --cash 10000

# Optimize parameters (target Sharpe ratio)
python -m src.cli.main backtest optimize \
  --strategy rsi_mean_reversion --symbol BTCUSDT --timeframe 1h \
  --start 2020-01-01 --end 2022-12-31 \
  --objective sharpe_ratio \
  --param-grid '{"rsi_period":[10,14,20],"rsi_oversold":[25,30,35]}'

# Walk-forward validation (70% train / 30% test)
python -m src.cli.main backtest walk-forward \
  --strategy rsi_mean_reversion --symbol BTCUSDT --timeframe 1h \
  --start 2020-01-01 --end 2025-12-31 \
  --train-ratio 0.7

# Multi-symbol comparison
python -m src.cli.main backtest multi-symbol \
  --strategy rsi_mean_reversion \
  --symbols BTCUSDT ETHUSDT BNBUSDT \
  --timeframe 1h --start 2024-01-01 --end 2024-06-01

# Scan symbols for breakout candidates
python -m src.cli.main scan run \
  --scan all \
  --symbols BTCUSDT ETHUSDT SOLUSDT \
  --timeframe 1h --start 2024-01-01 --end 2024-06-01

# Compare all built-in breakout strategies on one symbol
python -m src.cli.main backtest compare-breakouts \
  --symbol BTCUSDT --timeframe 1h \
  --start 2024-01-01 --end 2024-06-01 \
  --sort-by sharpe_ratio

---

CLI Reference

python -m src.cli.main <group> <command> [options]

Group	Command	What it does
data	download	Download data for a symbol
data	download-multiple	Download multiple symbols
data	update	Incrementally refresh all data
data	list-data	Show available data (symbol / timeframe / date range)
strategy	list-strategies	Show all available strategies
strategy	show-parameters	Show parameters for a strategy
strategy	create	Generate a strategy from natural language (AI)
backtest	run	Run a single backtest
backtest	optimize	Sweep a parameter grid, target Sharpe / SQN / Profit Factor / etc.
backtest	walk-forward	Train/test split — detect overfitting
backtest	multi-symbol	Backtest across multiple symbols
backtest	compare-breakouts	Run and rank all built-in breakout strategies on one symbol
scan	run	Run one or all breakout scans on latest candle(s) for selected symbols
results	list-results	View recent backtest results

Scan command quick notes

• Available scan types: unusual_volume_breakout, new_local_high_breakout, resistance_breakout, ascending_triangle_breakout, all
• Use --symbols for specific symbols or --all-major to scan the built-in major list
• Optional --parameters accepts JSON to override scan thresholds

Scanner profile guidance:

• Apply predefined parameter bundles for scanner thresholds and breakout strictness.
• Aggressive finds more signals with looser filters.
• Conservative finds fewer signals with stricter confirmation.
• Normal sits between the two.

Example with parameter overrides:

python -m src.cli.main scan run \
  --scan unusual_volume_breakout \
  --symbols BTCUSDT \
  --timeframe 1h --start 2024-01-01 --end 2024-06-01 \
  --parameters '{"volume_multiplier": 1.8, "breakout_lookback": 30}'

---

Examples

See examples/README.md for full descriptions.

Script	Purpose
examples/00_tutorial.py	Walkthrough of all features
examples/01_basic_backtest.py	Simple MA Crossover on real BTC data
examples/02_gpu_optimization.py	Recommended — GPU DCA sweep (~1,145 tests/sec)
examples/03_ema_crossover.py	EMA+RSI optimization on 4H timeframe
examples/04_ai_strategy_gen.py	Generate strategy from natural language

Run any example from the project root:

python examples/01_basic_backtest.py

---

Python API

from src import backtest_engine, database, list_available_strategies
from src import GPUOptimizer, check_gpu_status, StrategyGenerator

# List strategies
print(list_available_strategies())

# Run a backtest
result = backtest_engine.run_backtest(
    strategy_class=MyStrategy,
    symbol="BTCUSDT",
    timeframe="1h",
    start_date="2024-01-01",
    end_date="2024-06-01",
)

# Check GPU
check_gpu_status()

---

Project Structure

BacktestingMCP/
├── run.py                   # Interactive launcher (start here)
├── src/
│   ├── __init__.py          # Public API (from src import ...)
│   ├── core/                # Backtesting engine (backtesting.py)
│   ├── data/                # Downloader + SQLite database
│   ├── strategies/          # 6 templates + DCA strategies + AI-generated
│   ├── optimization/        # GPU optimizer (CuPy + Numba)
│   ├── risk/                # Position sizing, limits, correlation
│   ├── ai/                  # AI strategy generator
│   ├── mcp/                 # MCP server
│   └── cli/                 # CLI (main.py)
├── examples/                # Numbered example scripts (00–04)
│   └── archive/             # Older / superseded scripts
├── scripts/                 # DCA comparison utilities
├── tools/                   # Developer utilities
│   ├── check_gpu.py         # Verify GPU/CUDA setup
│   ├── check_db.py          # Inspect database
│   ├── param_sensitivity.py # Parameter sensitivity analysis
│   └── signal_dca_test.py   # DCA signal testing
├── config/
│   └── settings.py          # All configuration (timeframes, risk, MCP, etc.)
├── docs/
│   ├── GPU_GUIDE.md              # GPU setup and performance guide
│   ├── STRATEGY_WORKFLOW_FAQ.md  # Strategy creation → optimization → validation FAQ
│   ├── DCA_STRATEGIES_GUIDE.md
│   ├── PARAMETER_FORMAT_GUIDE.md
│   └── OLLAMA_RTX4090_SETUP.md
└── data/
    └── crypto.db            # SQLite (created on first data download)

---

MCP Server

Exposes 6 tools for AI assistants (Claude Desktop, VS Code Copilot, etc.):

python -m src.mcp.server   # starts on localhost:8000

Tools: download_crypto_data, list_available_data, list_strategies, get_strategy_info, run_backtest, create_strategy.

---

GPU Setup

See docs/GPU_GUIDE.md for full details.

# Install CuPy for CUDA 12.x
pip install cupy-cuda12x

# Verify
python tools/check_gpu.py

All GPU scripts fall back to NumPy/CPU automatically when CuPy is unavailable.

---

Configuration

Edit config/settings.py to adjust:

• RiskConfig — account risk %, daily/weekly/monthly loss limits, position limits
• TradingConfig — direction (long/short/both), commission, slippage, day/hour filters
• DataConfig — database path, exchange, default symbols
• MCPConfig — server host/port
• OptimizationConfig — method (optuna/grid/random), trial count

---

Requirements

• Python 3.10+
• NVIDIA GPU with CUDA 12.x (optional, for GPU acceleration)
• See requirements.txt for full dependency list

Key packages: backtesting, vectorbt, cupy-cuda12x, numba==0.56.4, numpy==1.23.5, pandas, ta, ccxt