Benchmarks

Numbers below are from scripts/compare_benchmarks.py, run head-to-head against FilterPy, pykalman, and simdkalman on the same task, with the same F/Q/H/R and the same noisy sine-wave measurements. This page is reproducible — run the script yourself:

uv sync --group benchmark   # or: pip install filterpy pykalman simdkalman
uv run python scripts/compare_benchmarks.py

Correctness

All three linear-KF implementations produce bit-for-bit identical position RMSE (0.6380) on the shared task — a solid cross-check that kalbee's KalmanFilter is implementing the standard equations correctly, not just "close enough."

Single-series speed

Library	Time / run (ms)	Position RMSE
filterpy	6.9	0.6380
kalbee	12.5	0.6380
simdkalman	25.5	0.6745
pykalman	34.1	0.6380

For a single bare predict/update loop, FilterPy's minimal, pedagogical implementation has less per-call overhead than kalbee. kalbee is faster than both pykalman and simdkalman's single-series path here — but if all you need is one filter stepping through one series, FilterPy is hard to beat on raw speed (it just does far less: no batching, no 17 other filter types, no tracking/smoothing/learning modules).

Vectorized: many independent series at once

This is where it matters in practice — tracking hundreds or thousands of targets, or backtesting a filter over many time series. simdkalman's whole pitch is speed here; kalbee has a purpose-built filter for exactly this, VectorizedKalmanFilter:

Approach	1,000 series (ms)
kalbee, naive per-series loop	13,442
simdkalman, batched	728
kalbee, VectorizedKalmanFilter	184

kalbee's vectorized filter is ~4x faster than simdkalman and ~73x faster than looping over the same 1,000 series — because it batches every filter through the same F @ x/F @ P @ F.T NumPy call instead of iterating in Python, the same idea as simdkalman but built into the standard BaseFilter interface (same predict/update/filter_sequence API as every other kalbee filter, so it's a drop-in swap, not a separate library).

Feature breadth

	kalbee	FilterPy	pykalman	simdkalman	Stone Soup
Filter implementations	18	~10	2	1	~6
Multi-object tracking (SORT/JPDA/PMBM)	✅	❌	❌	❌	✅ (heavier framework)
Smoothers (RTS/EKF/UKF/fixed-lag)	✅	RTS only	RTS only	❌	✅
Parameter learning (EM, online EM, NIS auto-tune)	✅	❌	EM only	❌	Partial
Neural hybrid filter (KalmanNet)	✅	❌	❌	❌	❌
Vectorized/batched filtering	✅	❌	❌	✅	❌
pandas / Polars DataFrame integration	✅	❌	❌	❌	❌
scikit-learn fit/transform API	✅	❌	❌	❌	❌
Actively maintained (2026)	✅	Mostly dormant	Mostly dormant	Mostly dormant	✅ (defence-oriented)

Stone Soup is the closest thing to a "does more than kalbee" comparison, but it's a full defence-grade tracking framework — heavier to learn, optimized for algorithm research rather than drop-in speed. If you want the breadth of a research framework with the ergonomics of a normal Python library, that's the gap kalbee is aimed at.