Installation¶

Requirements¶

Python 3.11+
numpy >= 1.26
scipy >= 1.11
pandas >= 2.1
arch >= 6.2 (univariate GARCH base)
statsmodels >= 0.14 (VAR, cointegration)

PyPI (pure Python)¶

pip install mfe

With Cython extensions (recommended for production)¶

pip install mfe[cython]

This compiles the realized._core and multivariate._core Cython extensions. Binary wheels are provided for Linux x86-64, macOS, and Windows (Python 3.11-3.13).

From source¶

git clone https://github.com/your-org/mfe-toolbox
cd mfe-toolbox
uv sync --group dev
uv run python setup_cython.py build_ext --inplace   # optional Cython

With uv (recommended)¶

uv add mfe
# or for the latest development version:
uv add git+https://github.com/your-org/mfe-toolbox

Verifying Cython compilation¶

from mfe.realized._core import _autocovariance_sum
print("Cython compiled: OK")

If this raises ImportError, the pure-Python fallback is used automatically. All functions work correctly without Cython — it is a performance optimisation only.

Performance without Cython¶

Operation	numpy	Cython
Realized kernel (T=500K)	34 ms	7 ms
MedRV (T=500K)	62 ms	1 ms
Hayashi-Yoshida (N=10K)	730 ms	< 1 ms
DCC Q-recursion (T=2000, K=5)	8.7 ms	0.06 ms
BEKK scalar (T=2000, K=2)	32 ms	0.04 ms