fastmatrix-profile 0.1.0

Pure-numpy Matrix Profile (motif and anomaly discovery in time series). No Numba required.

fastmatrix-profile

Pure-numpy Matrix Profile for time-series motif and anomaly detection. No Numba. No C extensions. No GPU. Just numpy + scipy.

Status: v0.1.0 — 26/26 tests passing, output matches STUMPY, NYC Taxi tutorial recovers 4/5 documented anomalies.

What it does

Given a 1D time series T and a window length m, computes the Matrix Profile P — a 1D summary where:

• argmin(P) points to a repeated pattern (motif)
• argmax(P) points to the most unusual subsequence (anomaly / discord)

One call, no tuning, no labels.

Install

pip install fastmatrix-profile

30-second example

import numpy as np
from fastmatrix_profile import matrix_profile

T = np.loadtxt("your_timeseries.csv")
P, I = matrix_profile(T, m=100)

anomaly_idx = int(P.argmax())
motif_idx   = int(P.argmin())

That's it. See examples/nyc_taxi_anomalies.ipynb for a real walk-through on the Numenta NAB NYC taxi dataset — recovers the NYC Marathon, Christmas, New Year and the Jan 2015 snowstorm without any parameter tuning beyond m.

Benchmarks vs STUMPY

Measured on macOS arm64, Python 3.12, scipy + Accelerate BLAS. See the interactive demo for the raw numbers and methodology.

Warm wall-time (median of 3 reps after warmup)

n	m	fastmatrix-profile (default)	+ dtype="float32"	STUMPY	winner
1,000	50	1.4 ms	1.0 ms	7.4 ms	ours, 5–7×
2,000	50	6.4 ms	4 ms	11.1 ms	ours, 1.7–2.5×
5,000	100	39 ms	19 ms	27 ms	f32 beats STUMPY (1.4×)
10,000	100	136 ms	72 ms	62 ms	STUMPY (2.2× / 1.15×)
20,000	100	590 ms	263 ms	175 ms	STUMPY (3.4× / 1.5×)

dtype="float32" is opt-in: ~2× faster at the cost of ~3e-5 absolute error in P (argmin/argmax positions unchanged in practice — verified on the NYC taxi dataset, both modes flag identical anomalies).

Cold start (fresh Python process, n=2000, m=50)

	Time
fastmatrix-profile	0.17 s
STUMPY	8.6 s (49× slower — Numba JIT compile)

When to use this vs STUMPY

Situation	Use
Long series (n > ~10k), Numba installs fine, many calls per process	STUMPY
Lambda / edge / Docker-slim / Pyodide / locked-down corporate Python	fastmatrix-profile
One-shot calls or interactive notebooks where cold start matters	fastmatrix-profile
GPU available, very large n	STUMPY-CUDA or SCAMP

This is not a STUMPY replacement at large n. It's the option for when you can't or don't want to install a JIT toolchain.

Algorithm

Uses a row-tiled BLAS-batched self-join: form the z-normalised window matrix Wz of shape (L, m), then sweep row blocks of Wz, multiply each block against Wz.T in a single gemm call, mask the trivial-match band, reduce to per-row min, and discard the tile. Peak memory: tile × L floats per tile (tens of MB) instead of the GBs a full (L, L) Gram matrix would need. Cost: O(L² m) flops at near-peak BLAS throughput.

Falls back to the STOMP recurrence (Zhu et al., ICDM 2016) when even a minimum-size tile would exceed max_mem_mb (default 512 MB) — typically n ≳ 1M. You can call matrix_profile(very_long_series, m) without blowing up RAM; it will just be slower past the dense regime.

dtype="float32" runs the inner GEMM in single precision: ~2× faster, ~half the peak memory, ~3e-5 absolute error in P (argmin/argmax positions unchanged in practice).

mass(Q, T) is exposed as a standalone primitive (Mueen's MASS distance profile via FFT).

Limitations

• 1D series only. No multi-dimensional matrix profile.
• No streaming / incremental updates.
• No GPU.
• NaN/Inf in the input are not handled — clean them first.

These are deliberate v0.1.0 scope decisions. Open an issue if you have a concrete use case.

Citation

@inproceedings{yeh2016matrix,
  title={Matrix profile I: all pairs similarity joins for time series},
  author={Yeh, Chin-Chia Michael and others},
  booktitle={2016 IEEE 16th International Conference on Data Mining (ICDM)},
  year={2016}
}

License

MIT