Installation

edgekit is a source install: a small numpy + pandas core that always imports, plus opt-in extras for the heavy dependencies. Install the lean core in seconds, or pull everything with "[all]".

edgekit is not on PyPI yet. You install it editable (-e) from a local checkout at ~/Documents/edgekit, so edits to the source are picked up without a reinstall. It requires Python 3.10 or newer.

Editable install#

From the repository root, pick the extras you need. The -e flag makes it an editable (development) install; the bracketed name after the dot selects optional-dependency groups.

cd ~/Documents/edgekit

pip install -e ".[all]"      # everything: io + viz + ml + dev
pip install -e .             # lean core only: numpy + pandas

The lean install is enough to load bars, run a causal backtest, compute trade stats, and run the permutation test — the whole prove-or-kill loop lives in the numpy + pandas core. You only need extras for parquet I/O, charts/HTML reports, or the ML layer.

pip install -e ".[viz]"          # + matplotlib (charts + HTML reports)
pip install -e ".[ml]"           # + scikit-learn / xgboost / lightgbm
pip install -e ".[io]"           # + pyarrow (parquet caches + splits)
pip install -e ".[viz,ml]"       # combine groups with a comma

The extras#

Each extra is a group in pyproject.toml. [all] is simply the union of the four. Install only what a given piece of work touches — a research script that never draws a chart does not need matplotlib.

Extra	Pulls in	Unlocks
`[io]`	pyarrow ≥ 12	Parquet: `data.load_bars` on `.parquet` splits and `data.hashed_parquet_cache`.
`[viz]`	matplotlib ≥ 3.7	The `viz` charts (equity, drawdown, monthly heatmap, MC fan) and every `report` HTML page.
`[ml]`	scikit-learn ≥ 1.3, xgboost ≥ 2.0, lightgbm ≥ 4.0	The `ml` layer: triple-barrier labels, purged walk-forward, models, meta-labeling, cloud-safe tree export.
`[dev]`	pytest ≥ 7.4, pytest-cov ≥ 4.1, ruff ≥ 0.4	The test suite and linter — for working on edgekit itself.
`[all]`	all four groups above	Everything. The one to install if you are unsure.

→Heavy deps load lazily

import edgekit only ever needs numpy + pandas — it never imports matplotlib, scikit-learn, xgboost or pyarrow at import time. Those are imported inside the functions that use them, so a missing extra surfaces as a clear error (pip install edgekit[viz]) only when you call the code that needs it — not when you import the package. You can do the entire load → backtest → prove loop on the lean core.

Verify the install#

Confirm the package imports and reports its version. This works on the lean core with no extras:

python -c "import edgekit as ek; print(ek.__version__)"
# 0.1.0

If you installed [dev] (or [all]), run the suite from the repo root. All 99 tests should pass — the load-bearing ones are the causality property tests (perturb a future bar, assert the past doesn't move), the statistical-validity checks (a no-edge strategy must not score significant), and the tree-export round-trip (cloud-safe inference matches scikit-learn to 1e-6).

pytest            # 99 passed

✦Use it from another project

A separate project can install edgekit the same way, from its own checkout: pip install -e ../edgekit. New research is written against edgekit; the R-multiple is the shared currency and causality is a tested property on both sides.

Next#

Quickstart — your first validated backtest, step by step.
The pipeline — load → backtest → prove → size → ship, stage by stage.
The validation gauntlet — why a good backtest is where skepticism starts.
API reference — every public function, module by module.