bunker-stats-rs 0.1.0

Ultra-fast Rust-powered statistics and time-series utilities for Python.

.

🚀 bunker-stats-rs

Ultra-fast Rust-powered statistics + time-series utilities for Python. Designed for data scientists, quants, researchers, analysts, and ML engineers who need NumPy-compatible accuracy with massive speedups on rolling statistics, covariance/correlation, outlier detection, ECDF, KDE, and more.

Goal: A lightweight, zero-dependency, high-performance alternative to many NumPy / Pandas / SciPy statistical operations — with predictable performance on large arrays.

📦 Installation pip install bunker-stats-rs

⚡️ Why bunker-stats?

Pure Rust kernels

No Python loops

No Pandas overhead

Predictable vectorized performance

Identical numerical results (within fp tolerance)

Minimal dependencies

Up to 1700× faster depending on the operation

Built for large 1D/2D NumPy arrays

🔥 Benchmark Summary

Benchmarks run on: Windows 10 • Intel i7 • Python 3.10 • NumPy 1.26 • Pandas 2.2 Dataset sizes: 1,000,000-element 1D arrays and 200,000×10 2D matrices

Below is a curated “top wins” summary:

Top Speedups (reference_time / bunker_time) Group Operation Ref Backend Ref Time (ms) Bunker (ms) Speedup Allclose Max Diff rolling rolling_zscore python_ref 33934.42 19.49 ×1741.47 True 4.12e-11 diff_cum_etc cummean python_ref 297.37 2.35 ×126.72 True 0.0 rolling ewma numpy_ref 376.98 4.85 ×77.79 True 0.0 diff_cum_etc sign_mask python_ref 14.62 0.60 ×24.34 True 0.0 cov_corr rolling_cov pandas 157.25 14.06 ×11.18 True 4.48e-14 rolling rolling_mean pandas 54.68 5.14 ×10.63 True 7.99e-15 cov_corr cov_pair numpy 15.27 4.08 ×3.74 True 3.03e-18 outliers zscore_outliers python_ref 16.03 4.60 ×3.48 True 0.0 diff_cum_etc quantile_bins_10 pandas 82.57 44.68 ×1.85 True 0.0 scipy_compare iqr_scipy scipy 35.82 15.90 ×2.25 True 0.0

Full benchmark results are available in /benchmarks.

🧩 Features Basic Stats

mean / std / var (ddof=1)

percentiles

IQR, MAD

min-max scaling

robust scaling (median/MAD)

winsorizing

Rolling Windows

rolling mean

rolling std

rolling zscore (z of last element)

EWMA (exponential smoothing)

Diff / Cumulative Operations

diff

pct_change

cumsum

cummean

ECDF

quantile binning

sign masks

demean with sign mask

Covariance & Correlation

covariance (pair)

correlation (pair)

covariance matrix

correlation matrix

rolling covariance

rolling correlation

KDE (Kernel Density Estimate)

Fast Gaussian KDE

📌 Examples import numpy as np import bunker_stats_rs as bs

x = np.random.randn(1_000_000)

Fast std

s = bs.std_np(x)

Rolling mean

r = bs.rolling_mean_np(x, window=50)

Covariance

cov = bs.cov_np(x, x * 2.0 + 1.0)

ECDF

vals, cdf = bs.ecdf_np(x)

🧱 Design Goals

Be a surgical, ultra-fast replacement for statistical hot paths in Python workflows

Work directly with NumPy arrays (input/output stays NumPy)

Zero hidden state, deterministic execution

Predictable performance across large inputs

Low-level but ergonomic API

⚠️ Limitations (v0.1.0)

float64 only

1D and 2D arrays only

No nan* functions yet (nanmean, nanstd, nanpercentile)

Rolling windows do not skip NaNs

Percentile + KDE slower than NumPy/SciPy on small arrays

Not a drop-in replacement for pandas — focuses on raw NumPy data

These will improve in future releases.

🗺 Roadmap v0.2 — NaN-Aware API

nanmean / nanstd / nanvar

nanpercentile

NaN-friendly rolling windows

v0.3 — 2D Rolling Stats

rolling mean/std/cov/corr for matrices

v0.4 — Parallelism

Optional Rayon parallel kernels for 50M+ elements

v0.5 — sklearn-like Transformers

Scaling transformers

Outlier detectors

Binning transformers

🧪 Running Benchmarks cd benchmarks python bench_all.py

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

🤝 Contributing

PRs welcome — especially for:

new statistical kernels

rolling ops

SciPy parity

tests + benchmarks

performance improvements

⭐️ Support

If this library speeds up your workflow, please ⭐ the repo!