Pancake vs vectorbt
vectorbt is a vectorized Python backtesting library that tests thousands of parameter combinations in seconds using NumPy and Numba. Pancake runs one strategy at a time and emits a verified, reproducible result at a permanent URL.
At a glance
| Capability | Pancake | vectorbt |
|---|---|---|
| Open-source engine | ✓ Apache-2.0 (batter, Python 3.12+) | ✓ open-source core (Python); vectorbt PRO is commercial |
| Vectorized mass parameter sweeps | ✗ one spec per run; run_sensitivity_analysis covers entry×sizing grids | ✓ thousands of combinations in one NumPy pass |
| Prediction-market native | ✓ Polymarket, Kalshi, binary outcomes | ✗ any OHLCV series; no binary-resolution semantics |
| Verification boundary doctrine | ✓ explicit 3-tuple in every result | ✗ no structured epistemic scope statement |
| Agent-callable MCP surface | ✓ 10-tool surface | ✗ Python library, no MCP integration |
| Shareable verified result URLs | ✓ /<handle>/<strategy_slug>/v/<version_n> with SHA-256 result hash | ✗ results live in the notebook or script that produced them |
| Paper trading | ✓ live paper deployments via MCP | ✗ research library only |
What's different
vectorbt treats backtesting as an array problem. Strategies, parameters, and assets become dimensions of NumPy arrays, and Numba-compiled kernels evaluate enormous parameter grids orders of magnitude faster than event-driven loops. It is the strongest open-source tool for exploring a parameter space: heatmaps of Sharpe across thousands of combinations are its native output.
That strength is exactly the overfitting hazard Pancake is designed to expose. A grid of 10,000 backtests will contain impressive Sharpe ratios by chance alone, and nothing in the output distinguishes selection effects from edge. Pancake runs one declared spec against one hashed evidence dataset and certifies that single run: the verification boundary names what was verified and what was agent-supplied, the permutation test puts a p-value on the Sharpe, and the result hash makes the run reproducible byte-for-byte.
The two tools sit at different stages. vectorbt is for the exploration phase — finding candidate parameter regions. Pancake is for the certification phase — committing to one spec and producing a result another party can audit. An agent workflow can use both: explore locally, then submit the chosen spec to Pancake for a verified result that travels with the strategy toward paper and live deployment.
Methodology overlap
Both compute Sharpe, drawdown, win rate, and trade statistics from P&L series and support custom fees and slippage. vectorbt exposes far more raw statistics across parameter grids; Pancake adds Wilson CI95, bootstrap CI, permutation testing, small-sample suppression, and byte-stable determinism (canonical JSON, PCG64, Python 3.12+).
When to use each
When to use Pancake
Use Pancake when you have one strategy spec to certify on prediction-market evidence and the result must be reproducible and citable by agents or readers who did not run it. Pancake is the right tool once exploration is over.
When to use vectorbt
Use vectorbt when you are sweeping large parameter spaces over OHLCV data and need raw research throughput in a notebook. vectorbt is the right tool for the exploration phase of quantitative research.
Citation
vectorbt is a Python library for vectorized backtesting built on NumPy, pandas, and Numba. vectorbt.dev. Pancake comparison: usepancake.com/compare/pancake-vs-vectorbt