Lib Talib

TA-Lib

The C reference every other Python indicator library is measured against — and the reason two libraries computing "the same" RSI hand you two different backtests.

The trap that costs you money

🚨 Every pure-Python port emits values before TA-Lib does, computed on partial windows — silently wrong, not NaN. TA-Lib's documented "unstable period" behaviour is exactly why its warm-up start indices differ from every port. Swap libraries, or restart the calculation per symbol on a short slice, and the same strategy becomes two different backtests.

Measured — TA-Lib python 0.7.1 / C core 0.7.1, pandas 3.0.5, numpy 2.4.6, 500-bar synthetic GBM, seed 7; maxabs = max absolute difference over the overlapping non-NaN region:

How long ta's RSI(14) takes to converge on the same input: |diff| > 1.0 through bar 36 · > 0.1 through bar 67 · > 0.01 through bar 97 · > 1e-6 through bar 217.

🚨 Universal rule: discard the first ~10–15 × period bars of any recursive indicator (EMA, RSI, ATR, ADX, MACD), whatever the library.

The install folklore is obsolete

"Needs the C library compiled by hand; no Windows wheels" was true and is not any more. ta-lib-python ships prebuilt wheels bundling the TA-Lib C library since v0.6.5. pip install TA-Lib just works on Windows. If you are still writing build instructions into a README, that guidance is stale.

Calling it correctly

It takes numpy float64 arrays, not Series. Pass .values and reattach the index yourself, or use the abstract API, which accepts dict / pandas / polars inputs.

import talib
rsi = talib.RSI(close, timeperiod=14)          # close must be float64 ndarray

from talib import abstract
out = abstract.MACD(df, fastperiod=12, slowperiod=26, signalperiod=9)

Which library for which job

✅ talipp's incremental .add() produces bit-identical output to batch construction. That is the whole reason it exists, and it is the correct choice when a strategy runs streaming — it removes the backtest-vs-live warm-up divergence.

🚨 pandas-ta — supply-chain caution

• github.com/twopirllc/pandas-ta → HTTP 404 (removed ~June 2025).
• PyPI now lists only two releases, 0.4.67b0 and 0.4.71b0. All prior history, including the widely used 0.3.14b, was deleted.
• No license field and no license classifier; requires_python >=3.12; the Repository URL still points at the dead 404 repo.
• pandas-ta.dev — DNS does not resolve, and archive.org has no snapshots.

A package with no licence, wiped release history, a dead upstream repo and a dead homepage is a supply-chain risk regardless of intent. Pin to pandas-ta-classic, ta, or TA-Lib.

🚨 ta's verified bugs, if you must use it

Isolated by perturbing only future bars and checking whether early bars changed:

1. IchimokuIndicator(..., visual=True).ichimoku_a() — CONFIRMED LOOK-AHEAD, 26 early bars changed, max |Δ| 35.34, even with fillna=False.
2. DPOIndicator(..., fillna=True).dpo() — CONFIRMED LOOK-AHEAD, 11 early bars, max |Δ| 34.43.
3. _check_fillna() does ffill().bfill() — the bfill() pulls future values backwards into leading NaNs; PSAR up/down are genuinely affected.
4. ichimoku_b() hardcodes min_periods=0 — the first 51 Span B values are computed on partial windows and are silently wrong.
5. ta.utils.dropna() silently deletes every row where any numeric column equals 0.0 — it will destroy zero-volume and zero-return rows.

See also

• ../../../fin-core/skills/signal-construction/SKILL.md — leak-free signal construction, §2 measured agreement
• ../../../fin-core/skills/signal-construction/references/indicator-libraries.md — the source card
• ../../../fin-core/skills/signal-construction/references/_repainting.md — the full repainting detail

Where this sits

This file is the deep dive on one library and assumes the choice is already made. For which library to pick, how it compares with the alternatives, and the traps that span several of them, the entry point is the domain skill signal-construction (../../../fin-core/skills/signal-construction/SKILL.md).