Reproducible Python research environment with locked dependencies
Researchers who rerun last quarter's notebook and get a different Sharpe ratio have usually been bitten by an unpinned NumPy, pandas or SciPy upgrade, not by the market. Locking every dependency, including the BLAS build underneath NumPy, means the same commit gives the same numbers on any machine and a new joiner is productive on day one.
First tasks
- Export the packages and versions currently installed on your research machine
- Choose one environment manager and write the reason in the repository readme
- Generate and commit a lockfile that pins the BLAS backend as well
- Rebuild on a clean container and compare a reference notebook's output byte for byte