Signal Construction

Signal construction

Two failures dominate: using a library whose warm-up values are silently wrong, and using an indicator that cannot be computed causally. §2 and §4 cover them. Both were measured, not assumed.

1. What changed since ~2024

🚨 Treat pandas-ta as do-not-install. A package with no licence, wiped release history, a dead upstream repo and a dead homepage is a supply-chain risk regardless of intent.

Also dead: finta (archived 2022, LGPL-3.0), tulipy (2019, LGPL), bta-lib.

2. ⭐ Measured agreement with TA-Lib

Executed on 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 abs diff over the overlapping non-NaN region.

Ranking for "numerically identical to TA-Lib":

1. TA-Lib (reference)
2. talipp — closest, and the only one that also matches TA-Lib's warm-up start index on RSI
3. pandas-ta-classic (native, talib=False) — matches on the core set; ADX warm-up is the exception
4. ta (bukosabino) — matches on non-recursive indicators only

How long ta's RSI(14) takes to agree with TA-Lib 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.

🚨 ta's Wilder-smoothed indicators need ~15× the period of burn-in before they agree with TA-Lib. On short datasets, or when you restart the calculation per symbol on a short slice, ta and TA-Lib produce materially different signals.

Universal rule: discard the first ~10–15 × period bars of any recursive indicator (EMA, RSI, ATR, ADX, MACD), whatever the library. ta, pandas-ta-classic and talipp all emit values before TA-Lib does, computed on partial windows — silently wrong, not NaN.

✅ talipp's incremental .add() is bit-identical to batch construction — so backtest and live produce the same numbers. That is the whole reason the library exists, and it is the correct choice when a strategy must run streaming.

3. 🚨 Verified look-ahead bugs in ta 0.11.0

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

1. IchimokuIndicator(..., visual=True).ichimoku_a() — CONFIRMED LOOK-AHEAD. Source: spana.shift(self._window2, fill_value=spana.mean()) — the whole-sample mean fills the first 26 rows. Editing only bars 200+ changed 26 early bars, max |Δ| = 35.34. Happens even with fillna=False, because fill_value sits inside .shift(), outside the guard. visual=False is clean.
2. DPOIndicator(..., fillna=True).dpo() — CONFIRMED LOOK-AHEAD. self._close.shift(int(0.5*window)+1, fill_value=self._close.mean()). With fillna=True the rolling min_periods drops to 0 so those whole-sample-mean rows survive. Editing bars 200+ changed 11 early bars, max |Δ| = 34.43. Clean with fillna=False.
3. ichimoku_b() hardcodes min_periods=0, ignoring fillna. Span B's first_valid_index() is 0 while Span A's is 25 — the first 51 Senkou Span B values are computed on partial windows and are silently wrong.
4. _check_fillna() does ffill().bfill() — the bfill() propagates future values backwards into leading NaNs. Verified affected: Ichimoku conv/base/spanA/spanB, PSAR up/down (which are NaN by construction on the opposite trend, so their leading NaNs are backfilled from the future), Keltner, Donchian, cumulative return.
5. ta.utils.dropna() silently deletes every row where any numeric column equals 0.0 — it will quietly destroy zero-volume or zero-return rows.

Verdict: ta is fine for a quick pure-Python feature set with fillna=False; never use visual=True on Ichimoku for modelling; discard warm-up generously.

4. 🚨 Indicator classes that inherently leak

5. Detect it automatically

The generic test is six lines and catches almost everything — scripts/assert_causal.py:

def assert_causal(fn, df, k):
    """Perturb only rows >= k; assert nothing before k moved."""
    a = fn(df)
    df2 = df.copy(); df2.iloc[k:] *= 2.0
    b = fn(df2)
    bad = (a.iloc[:k].fillna(-9e9) - b.iloc[:k].fillna(-9e9)).abs() > 1e-9
    assert not bad.any().any(), f"LOOK-AHEAD: {int(bad.sum().sum())} pre-{k} cells depend on post-{k} data"

Off-the-shelf: freqtrade lookahead-analysis (reruns on truncated data, flags indicators whose history changes — only checks triggered signals) and freqtrade recursive-analysis (varies startup_candle_count ∈ {199, 499, 999, 1999} and reports % variance — the only off-the-shelf tool that measures the unstable-period problem directly).

6. Choosing

7. Reference files

references/<library>.md carries the full per-indicator comparison tables, exact signatures and gotchas. references/_repainting.md is the blacklist above with worked detections.

Per-library deep dives

The optional fin-libraries plugin carries a dedicated skill for each library below. Load one only after this skill has told you which library you want:

• lib-talib — talib

Below the bar

Bar construction (tick, volume, dollar), trade classification, effective and realized spreads, and order-flow measures - ../intraday-microstructure/SKILL.md. The bar-timestamp leak above is the first trap there as well.