tslib

Fast time-series analytics library with a C++/AVX2 core and Python bindings via pybind11.

Rolling stats (mean, variance, z-score, autocorrelation), RSI, MACD. Lazy API, chunked streaming for large arrays, Polars DataFrame integration.

install

uv pip install -e . --no-build-isolation

Requires Python 3.12, CMake, a compiler with AVX2 support (falls back to scalar on ARM).

usage

import numpy as np
from tslib import TimeSeries

ts = TimeSeries(prices)
rw = ts.rolling(20)

rw.mean()
rw.variance()
rw.zscore()
rw.autocorr(lag=1)

# technical indicators
rsi = ts.rsi(14)
macd = ts.macd(12, 26, 9)  # returns dict with 'macd', 'signal', 'histogram'

# chunked processing for large arrays
for chunk in ts.chunked(5_000_000, overlap=19):
    m = TimeSeries(chunk).rolling(20).mean()

# polars integration
from tslib import from_polars, apply_rolling

df = apply_rolling(df, col="close", window=20, ops=["mean", "variance", "zscore"])
ts = from_polars(df, "volume")

benchmarks

uv run benchmarks/bench_python.py

Rolling mean, window=20, N=10,000,000 (4 threads):

	time
numpy convolve	207.8 ms
pandas rolling	243.4 ms
tslib	20.3 ms

10.3x faster than numpy, 12.0x faster than pandas

Rolling variance: 14.8x faster than pandas rolling var.