numbatalib 0.1.0

TA-Lib reimplemented in pure Python + Numba.

numbatalib

Reimplement TA-Lib (technical analysis indicators) in pure Python + Numba, aiming for:

• algorithmic parity with the local installed talib (ground-truth reference for tests/benchmarks)
• near-C performance for large vectors (Numba JIT)
• no compiled extension at install time (no libta-lib dependency)

Status

• Implemented TA-Lib Core functions: 161 / 161
• Parity tests: generated/pytest_results.txt
• Parity + benchmark results: generated/parity_results.csv, generated/bench_results.csv
• Status tracker: port_checklist.csv

Note: the local talib wheel in this environment does not expose ACCBANDS, AVGDEV, IMI; those are validated against upstream TA-Lib C (see generated/parity_results_upstream_c.csv).

Install

After publishing to PyPI:

pip install numbatalib

Usage

numbatalib exposes TA-Lib function names dynamically:

import numpy as np
import numbatalib as ta

x = np.random.default_rng(0).normal(size=1000).cumsum()

sma = ta.SMA(x, timeperiod=20)
rsi = ta.RSI(x, timeperiod=14)

TA-Lib compatible API (minimal habit cost)

If you want talib-like APIs + error messages, use the compatibility shim:

import numpy as np
import numbatalib.talib as talib  # or: import numbatalib.compat.talib as talib

x = np.random.default_rng(0).normal(size=1000).cumsum()

talib.SMA(x, timeperiod=20)
talib.stream.SMA(x, timeperiod=20)        # streaming scalar
from numbatalib.talib import abstract
abstract.Function("SMA")(x, timeperiod=20)

Notes:

• set_compatibility/get_compatibility and set_unstable_period/get_unstable_period are supported (matching TA-Lib behavior for EMA/RSI/CMO and unstable-period masking).

Dev

• Run parity tests vs installed talib: pytest -q
• Regenerate parity + speed CSVs and update checklist: python tools/compare_vs_talib.py --bench --write-checklist