Pancake vs Lumibot
Lumibot is an open-source Python framework for backtesting and live/paper trading through broker APIs — self-hosted, you run the process. Pancake is hosted verification infrastructure for AI-built prediction-market strategies.
At a glance
| Capability | Pancake | Lumibot |
|---|---|---|
| Open-source | ✓ Apache-2.0 engine (batter, Python 3.12+) | ✓ open-source Python framework |
| Prediction-market native | ✓ Polymarket, Kalshi, binary outcomes | ✗ equities, options, crypto via broker APIs |
| Strategy hosting (no user infrastructure) | ✓ platform runs the deployed strategy version | ✗ self-hosted — you run and supervise the bot process |
| Verification boundary doctrine | ✓ explicit 3-tuple in every result | ✗ backtest output without epistemic scope statement |
| Receipt URLs with byte-stable hashes | ✓ /<handle>/<strategy_slug>/v/<version_n> — public shareable URLs | ✗ results live in local files and logs |
| Agent-callable MCP surface | ✓ 10-tool surface | ✗ Python API; no MCP integration |
| Paper trading | ✓ simulated fills on live Polymarket data, append-only public ledger | ✓ paper mode via supported brokers |
| Live broker execution | ✗ paper trading live; real-money execution on the roadmap | ✓ broker integrations (Alpaca, Interactive Brokers, and others) |
What's different
Lumibot is a framework, and a productive one: Python classes for strategies, a backtester, and broker adapters that let the same strategy object run in paper or live mode. The operative word is framework — you write Python, you run the process, and the operational stack (uptime, data subscriptions, crash recovery, logs) is yours. For a Python-fluent trader working US equities or options through a broker API, that is a reasonable trade.
Pancake inverts the ownership: the platform owns the runtime and the strategy is data. A declarative spec — authored by an MCP-capable agent rather than written as Python classes — is validated against content-hashed evidence by the deterministic batter engine, versioned immutably, and deployed server-side as a paper deployment on live Polymarket data. There is no process to babysit, and the strategy's record is public by construction: an append-only event ledger, not a log file on the operator's machine. Live venue execution is a v2-roadmap capability.
The audience split mirrors the architecture split. Lumibot assumes the strategy author is a person who writes Python and wants broker connectivity. Pancake assumes the strategy author is an AI agent whose claims need independent verification — reproducible result hashes, an explicit verification boundary, and permanent URLs a third party (or another model) can check without trusting the author.
Methodology overlap
Both run rule-based strategies against historical data with per-trade fee and slippage assumptions, and both support a paper mode against live data. Pancake adds formal statistical guards as first-class output — Wilson CI95 on win rate, bootstrap CI on returns, permutation test p-values, and small-sample suppression below 10 trades — computed by a byte-stable engine so the numbers are reproducible across machines.
When to use each
When to use Pancake
Use Pancake when the strategy trades prediction markets, is built by an AI agent over MCP, and needs hosted execution plus verifiable, citable results rather than local logs. Pancake is the right tool when nobody should have to trust the operator's machine.
When to use Lumibot
Use Lumibot when you write Python, trade equities, options, or crypto through a supported broker, and want one codebase for backtest, paper, and live execution that you host yourself. It is the right tool for hands-on Python traders who want broker connectivity today.
Citation
Lumibot is an open-source Python trading framework maintained by Lumiwealth. lumibot.lumiwealth.com. Pancake comparison: usepancake.com/compare/pancake-vs-lumibot