Why CBT?

Most traders waste weeks writing boilerplate, debugging data pipelines, and manually tracking experiments. CBT Framework automates the boring parts so you can focus on what matters: your edge.

| Without CBT | With CBT | |---|---| | Write backtest engine from scratch | /cbt:build generates it | | Manually track experiments in spreadsheets | /cbt:compare does it automatically | | Guess at parameter optimization | /cbt:optimize runs walk-forward analysis | | Copy-paste code to go live | /cbt:live deploys to 4 exchanges | | Lose context between sessions | /cbt:clear saves everything | | Google library docs constantly | MCP servers give Claude real-time docs + market data |

Install in 30 Seconds

npx cbt-framework

That's it. This installs 21 commands, 4 AI agents, templates for 4 exchanges, and optionally sets up MCP servers for market data and macroeconomic research.

Requirements

• Claude Code CLI
• Node.js 16+
• Python 3.8+

The Full Workflow

/cbt:new my_strategy          Create strategy (pick YOLO mode + engine)
    |
/cbt:discover                  Define your edge through guided Q&A
    |
/cbt:research                  Validate with literature + GitHub code
    |
/cbt:eda                       Explore data with Seaborn visualizations
    |
/cbt:config + /cbt:plan        Configure params + create build plan
    |
/cbt:build                     Generate strategy code (follows the plan)
    |
/cbt:run                       Execute backtest
    |
/cbt:deep-analyze              Forensic analysis + statistical tests
/cbt:plot                      Signal visualization on candlestick charts
    |
/cbt:optimize                  Parameter optimization (sweep/grid/walk-forward)
    |
/cbt:iterate                   One-change-at-a-time improvement loop
    |
/cbt:report                    Auto-generated living report
    |
/cbt:live                      Deploy to Bybit, Binance, Kraken, Hyperliquid
/cbt:export                    Standalone package for sharing

Quick Start

# 1. Install
npx cbt-framework

# 2. Open Claude Code in your project folder
claude

# 3. Start building
/cbt:new btc_momentum
/cbt:discover
/cbt:research
/cbt:eda
/cbt:plan
/cbt:build
/cbt:run
/cbt:deep-analyze

Example Session

> /cbt:new btc_momentum
  Mode: YOLO | Engine: fast

> /cbt:discover
  Strategy defined. Type: momentum. Data: 5M rows.

> /cbt:eda
  12 Seaborn plots generated. Key finding: strong hourly seasonality.

> /cbt:build
  All steps complete. Baseline: Sharpe 1.45

> /cbt:deep-analyze
  Monte Carlo 95%: positive. Rolling Sharpe: stable.

> /cbt:optimize walkforward
  IS Sharpe: 1.8, OOS Sharpe: 1.5. Robust.

> /cbt:live setup
  Exchange: Bybit. Paper trading started.

All 21 Commands

Setup

| Command | What it does | |---------|-------------| | /cbt:new <name> | Create strategy (YOLO mode + engine choice) | | /cbt:status | Show state, mode, engine, progress | | /cbt:help | Show all commands | | /cbt:update | Update to latest version | | /cbt:clear | Save context + prepare for reset |

Build

| Command | What it does | |---------|-------------| | /cbt:discover | Strategy Q&A + data scale + project type | | /cbt:research | Literature, implementations, risk analysis | | /cbt:eda | Exploratory data analysis with Seaborn plots | | /cbt:config | Configure backtest parameters | | /cbt:plan | Create step-by-step build plan | | /cbt:build | Generate code (plan-aware, engine-aware) |

Run & Analyze

| Command | What it does | |---------|-------------| | /cbt:run | Execute backtest | | /cbt:analyze | Quick text-based analysis | | /cbt:deep-analyze | Forensic analysis with Seaborn + stats tests | | /cbt:plot | Signal/indicator/equity visualization | | /cbt:compare | Compare experiments side by side |

Optimize & Report

| Command | What it does | |---------|-------------| | /cbt:optimize | Parameter sweep, grid search, walk-forward | | /cbt:iterate | Guided one-change-at-a-time loop | | /cbt:observe | Save observations and hypotheses | | /cbt:report | Auto-generated living project report |

Deploy

| Command | What it does | |---------|-------------| | /cbt:live | Deploy to Bybit, Binance, Kraken, or Hyperliquid | | /cbt:export | Standalone package (zip, git, Docker) |

Dual Engine

Choose your engine when creating a strategy:

pandas (default)

Standard pandas + numpy. Best for datasets under 1M rows. Simple and debuggable.

Fast Engine (Polars + NumPy + Numba)

For large datasets (1M+ rows):

• Polars for data loading (lazy evaluation, zero-copy)
• NumPy arrays for feature engineering
• Numba @njit for compiled backtest loops
• No pandas in the hot path

# Optional: install fast engine dependencies
pip install polars numba numpy

MCP Servers (Data Superpowers)

CBT Framework can set up 3 free MCP servers during installation to give Claude access to external data:

| Server | What it does | API Key | |--------|-------------|---------| | Context7 | Up-to-date library docs (pandas, ccxt, polars...) | None needed | | Alpha Vantage | Stocks, forex, crypto + macro indicators (CPI, GDP, rates) | Free key | | FRED | 840,000+ economic time series from the Federal Reserve | Free key |

This means Claude can pull real market data and macroeconomic indicators while building and analyzing your strategies.

Live Trading

Supported Exchanges

• Bybit - USDT perpetuals, inverse, spot
• Binance - Spot, USDT-M, COIN-M futures
• Kraken - Spot, futures
• Hyperliquid - Decentralized perpetuals

Safety Features

• Paper trading mode by default
• Kill switch with configurable drawdown threshold
• Max position size limits
• API rate limiting
• Credentials in .env (never hardcoded)

Notifications

• Discord (webhook)
• Telegram (bot API)
• SMS (Twilio)
• Email (SMTP)

Project Structure

strategies/<name>/
├── Data/               # Datasets
├── IDEA.md             # Initial notes
├── DISCOVERY.md        # Strategy spec from /cbt:discover
├── RESEARCH.md         # Research findings
├── EDA.md              # Exploratory analysis
├── BUILD_PLAN.md       # Build steps from /cbt:plan
├── REPORT.md           # Living report
├── DEEP_ANALYSIS.md    # Forensic analysis
├── config.yaml         # Backtest config
├── src/                # Generated source code
├── strategy.py         # Main strategy
├── backtest.py         # Runner
├── experiments/        # All backtest runs
├── observations/       # Iteration notes
├── plots/              # Visualizations
│   ├── eda/            # EDA plots
│   └── deep_analyze/   # Analysis plots
├── trades/             # Trade logs
└── .cbt/
    ├── state.yaml      # Framework state
    └── handoff.md      # Session handoff

Best Practices

1. Lookahead Prevention - Always .shift(1) your indicators
2. One Change Per Iteration - Change only one thing at a time when optimizing
3. Paper Trade First - Always validate before going live
4. Use EDA - Let the data inform your strategy before building
5. Kill Bad Ideas Fast - Define kill criteria upfront, abandon if met

Contributing

Contributions are welcome! Feel free to open issues or submit PRs.

License

MIT License - see LICENSE for details.