TrendLib

325 technical-analysis indicators and patterns. Rust core, Python API.

Every indicator runs two ways: over a whole array, or one bar at a time. The two agree bitwise, so what you backtest is what you trade.

---

import numpy as np
import trendlib as tl

close = np.array([...])

tl.rsi(close, period=14)                  # one array out
macd, signal, hist = tl.macd(close)       # three
tl.cdl_engulfing(open_, high, low, close) # int32: +100, -100 or 0

Install

pip install trendlib

The first release ships a source distribution, so this compiles the Rust core and needs a toolchain (1.95 or later). Nothing else is required: no system TA-Lib, no C dependencies.

Pre-built wheels are on the way. The extension is abi3, so one wheel per platform will cover Python 3.11 and every version after it, and the install becomes a download with no toolchain at all.

What's in it

Group	Count	Examples
Chart patterns	63	head and shoulders, double top, wedges, channels, flags, cup with handle
Candlestick patterns	61	engulfing, harami, morning star, three black crows, hikkake
Momentum	51	RSI, MACD, stochastic, ADX, CCI, Williams %R, Connors RSI, TSI
Overlap studies	34	SMA, EMA, WMA, HMA, KAMA, T3, Bollinger, Keltner, Ichimoku, SuperTrend
Statistics	23	linear regression, correlation, beta, standard deviation, percentile
Bar patterns	19	inside day, outside day, key reversal, pipe top, narrow range
Volume	16	OBV, A/D, Chaikin, MFI, VWAP, VWMA, Twiggs money flow
Math transforms	15	log, exp, trigonometric and hyperbolic functions
Harmonic patterns	12	Gartley, bat, butterfly, crab, ABCD, Wolfe wave
Volatility	9	ATR, NATR, true range, Chaikin volatility, GAPO
Levels	7	traditional, Camarilla, Fibonacci, Woodie and DeMark pivots, CPR
Cycle	6	Hilbert transform: dominant cycle period, phase, trend mode
Math operators	5	add, subtract, multiply, divide, cumulative sum
Price transforms	4	typical, weighted, median and average price

201 of them carry a TA-Lib-style uppercase alias, so tl.RSI(close, timeperiod=14) works beside tl.rsi(close, period=14) and takes TA-Lib's parameter names.

How it works

Three layers, each doing one job.

A Rust core holds the arithmetic. It has no runtime dependencies at all (cargo tree prints one line) and is #![forbid(unsafe_code)].

One kernel per indicator, written once as a step that takes a bar and returns the next value. The batch function folds that step over an array and the stream calls it per bar, so they cannot drift apart: the bitwise agreement is structural, not something a test happens to confirm.

A thin Python layer does the conversion: NumPy in, NumPy out, with pandas and polars recognised and their index handed back.

Streaming

Streams are values, not handles. Each one is independent, nothing is shared, and there are no global settings anywhere.

live = tl.stream.rsi(history)   # seed with what you have
live.update(bar)                # advance, and get the new value
live.peek(bar)                  # what it would read, without committing
live.copy()                     # fork the state to explore a branch
live.value, live.bars_seen      # where it stands

tl.stream.<name> exists for all 325.

Frames

Pass a DataFrame and the bar columns are matched by name, case-insensitively:

tl.atr(frame)                      # finds high, low, close
tl.rsi(frame["close"])             # a Series comes back indexed like it went in

Warm-up

Every output starts where the indicator is first defined. Earlier rows are NaN, or 0 for the integer columns patterns return, and the arrays always come back the same length as the input.

tl.lookback("rsi", period=14)      # 14 - the first row that carries a value

Errors

tl.TrendLibError          # the base
tl.InvalidInput           # wrong shape, wrong dtype, NaN mid-series
tl.InsufficientHistory    # fewer bars than the lookback needs

Nothing is silently clamped or filled. A parameter outside its range raises rather than snapping to the nearest legal value.

Design notes

Measurements, not advice. No function, output or parameter is named for a trading action: there is no buy, entry or target anywhere. Pattern functions return +100, -100 or 0, meaning this shape is present and points up, points down, or is absent; what to do about it is yours.

Where a published formula and a widely used implementation disagree, the division and ordering are chosen to keep precision rather than to match an implementation quirk. Several indicators here are measurably closer to exact arithmetic than the reference libraries they are compared against.

Platforms

Python	3.11 and later (abi3)
Rust	1.95, edition 2024
OS	Linux (glibc, musl), macOS, Windows
Arch	x86-64, aarch64

License

Apache License 2.0. It grants patent rights explicitly, which a permissive licence that is silent on patents does not.

Built by GAIIN Technologies.