pandas-ta-classic/CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file. Format: Keep a Changelog.

[Unreleased]

Added

• backtesting.py integration (PR #124 by @Adhyansinghgupta): Bridge function (ta_bridge), SMACrossover example strategy (examples/backtesting_py_strategy.py), and integration tutorial (docs/tutorials/backtesting_py.md). Added backtesting to integration optional dependencies.
• backtrader integration (docs/tutorials/backtrader.md, examples/backtrader_strategy.py): Precompute-then-feed pattern using dynamic PandasData subclass (make_feed()); covers single-output, multi-output (MACD), and OHLCV-dependent (ATR) indicators. Added backtrader to integration optional dependencies.
• vectorbt integration tutorial (docs/tutorials/vectorbt.md): Documented the df.ta.tsignals() → vbt.Portfolio.from_signals() integration pattern, including multi-output indicator handling and benchmark comparison.
• Type stubs for AnalysisIndicators (pandas_ta_classic/core.pyi): IDE autocomplete and mypy now see full method signatures for every indicator.
• ichimoku() single-DataFrame return (as_dataframe=True): opt in to a single DataFrame of the visible period instead of the legacy (visible, span) tuple. Pass append_span=True to also append the future-dated projected span rows. append_span is now a real parameter on the underlying ichimoku() function (previously accessor-only) and is forwarded by df.ta.ichimoku().

Changed

• td_seq ~3500× faster (numba): the Tom DeMark Sequential consecutive-run count replaced rolling(13).apply(python_callback) with a numba @njit loop (~1912 ms → ~0.5 ms per 5000-bar call, the single largest indicator bottleneck). Bit-for-bit identical; falls back to a plain Python loop without numba. Reproduce with python tools/bench_indicators.py.
• hilo ~250× faster (numba): the Gann HiLo per-bar .iloc[] state loop moved into a numba @njit kernel over numpy arrays (~130 ms → ~0.5 ms per 5000-bar call). Output is identical for all mamodes; falls back to a plain Python loop without numba.
• jma ~130× faster (numba): the Jurik Moving Average per-bar recursion (volatility bands + 3-stage filter) moved into a numba @njit loop with scalar coefficients precomputed by the caller (~50 ms → ~0.4 ms per 5000-bar call). The trailing-window average uses np.mean in place of the original np.average; for an unweighted call numpy computes these bit-for-bit identically, so the numba-off fallback is bit-identical to the previous output. Under numba (JIT on), np.mean compiles to numba's own reduction, whose summation order differs from numpy's at the ULP level (~4e-16 relative, full float64 precision); compounded through the recursive jma[i] = jma[i-1] + det1 update the JIT path lands within ~1 ULP of the original but not bit-for-bit identical. Inconsequential for any practical use; jma is the only optimised indicator here whose JIT and numba-off paths are not bit-identical to each other (every other one is).
• ha ~50× faster (numba): the recursive Heikin-Ashi open (df.iat[] scalar loop) moved into a numba @njit scan over the precomputed HA_close array (~75 ms → ~1.5 ms per 5000-bar call). Output is identical; falls back to a plain Python loop without numba.
• stc ~40× faster (numba): the two Schaff Trend Cycle stochastic-smoothing loops moved into a shared numba @njit kernel, parametrised by the gate series (pass 1 gates on the rolling minimum, pass 2 on the rolling range) (~82 ms → ~2 ms per 5000-bar call). Output is bit-for-bit identical (the cumulative 8-decimal rounding is preserved inside the kernel); falls back to a plain Python loop without numba.
• wad ~15× faster: Williams A/D replaced Series.combine(prev_close, max/min) (a per-element Python callback) with vectorised np.fmax/np.fmin (~16 ms → ~1 ms per 5000-bar call). Output is identical; pure numpy, no numba needed.
• wilder_smooth / dm ~5× faster (numba): Wilder's cumulative smoothing (utils/_wilder.py, used by dm) moved its per-bar recursion into a numba @njit kernel. The seed (np.nansum(raw[1:length])) is still computed with numpy in the caller and passed in, so the cumulative sum matches bit-for-bit — the kernel runs only the deterministic scalar recursion. dm ~10.5 ms → ~2.0 ms per 5000-bar call. Bit-for-bit identical; falls back to a plain Python loop without numba.
• pmax ~4.6× faster (numba): the PMAX adaptive-band / trend-direction per-bar loop is now @njit-compiled with its Python-list accumulators replaced by numpy arrays (~5.4 ms → ~1.2 ms per 5000-bar call; the residual is the ATR/MA prep). The loop is pure comparisons + min/max + assignment, so output is bit-for-bit identical; falls back to a plain Python loop without numba.
• npround ~400× faster (vectorised): replaced the per-element close.apply(np.round) Python callback with a single vectorised np.round(close) (~6 ms → ~0.02 ms per 5000-bar call). Output is bit-for-bit identical.
• aroon, maxindex, minindex, minmaxindex ~9–17× faster (vectorised): the rolling argmax/argmin arg-position loops (Series.rolling(length).apply(np.argmax/argmin, raw=True)) now use a shared _sliding_argextreme() helper built on numpy.lib.stride_tricks.sliding_window_view (mirrors the existing _sliding_weighted_ma). aroon ~12 ms → ~0.7 ms, minmaxindex ~8 ms → ~0.8 ms, maxindex/minindex ~4 ms → ~0.4 ms per 5000-bar call. Integer arg-positions → bit-for-bit identical output; pure numpy, no numba needed.
• ebsw ~850× faster (numba): the Even Better SineWave per-bar .iloc[] HighPass + SuperSmoother recursion moved into a numba @njit kernel over a numpy array, with the (bar-invariant) filter coefficients hoisted out of the loop and the 2-element FilterHist list replaced by two scalars (~34 ms → ~0.04 ms per 5000-bar call). Output is bit-for-bit identical; falls back to a plain Python loop without numba.
• kama ~90× faster (numba): Kaufman's Adaptive MA per-bar smoothing recursion moved into a numba @njit loop over numpy arrays (the vectorised efficiency-ratio/smoothing-constant prep stays in pandas) (~30 ms → ~0.3 ms per 5000-bar call). Output is bit-for-bit identical; falls back to a plain Python loop without numba.
• mama ~55× faster (numba): the MESA Adaptive MA / FAMA inline Hilbert-Transform loop is now @njit-compiled (~16 ms → ~0.3 ms per 5000-bar call). Output is bit-for-bit identical (verified against the pure-Python .py_func); falls back to a plain Python loop without numba.
• Candlestick patterns ~32× faster (numba): all 60 native cdl_* pattern per-bar detection loops (the TA-Lib incremental running-window algorithm) moved into numba @njit kernels over numpy arrays, with the average-factor constants and range arrays hoisted by the caller. cdl_pattern(name="all") drops from ~288 ms to ~9 ms per 5000-bar call. Output is bit-for-bit identical for every pattern; falls back to a plain Python loop without numba.
• Faster cold import (~3–4×): import pandas_ta_classic takes ~280 ms (was ~930 ms) because indicator functions load on demand. Public API unchanged.
• Math operators individually importable: Each math/trig operator (add, sub, mult, div, rolling_max/min/sum, acos, cos, exp, ln, …) now lives in its own submodule, e.g. from pandas_ta_classic.math.add import add.
• combination() delegates to math.comb: Replaces the hand-rolled nCr loop. Signature and results unchanged; the unused multichoose kwarg alias was dropped (use repetition).
• npNaN alias removed internally: All modules now use np.nan directly. No public API change.
• All npX numpy import aliases removed (26 aliases, e.g. npSqrt, npArange, npNdArray): modules now use import numpy as np + np.x directly. Internal only; no public API change.
• Explicit package namespace: pandas_ta_classic/__init__.py, core.py, and utils/__init__.py no longer use star imports — every re-export is explicit with __all__ defined. All documented names (utility functions, metrics, Strategy, lazy-loaded indicators) are unchanged. Incidental namespace leaks (e.g. ta.np, ta.pd, ta.DataFrame, ta.logger) are gone; import these from their real homes.
• Lint-clean repo: ruff check . passes with zero errors and zero per-file ignores. Ambiguous variables renamed (O → O_ in candle loops, l → low in tests); sys.path bootstrap hacks removed from tests/fixtures/, tools/, and docs/conf.py (the editable install makes them redundant); notebook cells reordered to import-first.
• linear_regression() always uses the numpy implementation: the "r" key is now consistently the Pearson correlation coefficient. Previously, environments with scikit-learn installed silently used LinearRegression.score() (R²) instead, so results depended on the environment. Affects pure_profit_score.
• optimal_leverage() returns a float: the result is no longer truncated with int(), the capital argument is documented, the unused **kwargs parameter is removed, and the "Incomplete. Do NOT use" docstring warning is removed. Zero-variance input now raises ValueError (previously crashed with OverflowError).
• Optional-dependency extras restructured: the integration extra is split into data (yfinance, alpha-vantage — the only integrations the library imports, lazily via utils/data) and backtest (backtesting, vectorbt, backtrader — docs/examples only). integration is retained as an alias for data,backtest. The [all] extra no longer pulls data/backtest, so pip install "pandas-ta-classic[all]" stays installable on newer Python/OS combinations that lack wheels for these unmaintained, platform-fragile packages; install them explicitly when needed.
• tulipy oracle frozen to a golden file: tulipy (unmaintained, no wheels for CPython ≥3.12) is removed from the oracle extra and no longer imported at test time. Its deterministic outputs are snapshotted in tests/fixtures/tulipy_oracle.json (regenerate with tests/fixtures/generate_tulipy_oracle.py on CPython <3.12), and test_oracle_tulipy.py now compares native output against the frozen arrays on every Python version. This fixes the testing-oracle CI job on Python 3.12–3.14, where the tulipy install previously failed. ta-lib remains the live primary oracle. (The MSW oracle tolerance was loosened to 1e-3 to stay robust to ~1e-4 numpy-version float drift in that recursive trig indicator.)
• tal_ma() no longer needs TA-Lib and fails fast: the MA-name→MA_Type mapping is now a plain dict of the (frozen) TA-Lib enum values, so tal_ma("ema") == 1 regardless of whether TA-Lib is installed (previously it silently returned 0/SMA when TA-Lib was absent). Unknown names now raise ValueError and non-string input raises TypeError instead of silently defaulting to SMA. Callers (bbands, ppo, apo) already normalise mamode to a valid string, so their behaviour is unchanged.

Deprecated

• Built-in data fetching (df.ta.ticker(), ta.yf(), ta.av()): emit a FutureWarning and will be removed in a future release. Data fetching is out of scope for a technical-analysis library. Fetch OHLCV with yfinance / alpha-vantage directly and pass the DataFrame to pandas-ta-classic; see the new examples/fetch_market_data.py for the replacement patterns. The data optional extra remains available during the deprecation window.
• ichimoku() tuple return: returning a (visible, span) tuple is deprecated and now emits a DeprecationWarning. Pass as_dataframe=True to opt in to the single-DataFrame return (add append_span=True for the projected span rows), or as_dataframe=False to keep the tuple without warning. The accessor df.ta.ichimoku() already returns a DataFrame and is unaffected.
• CDL_PATTERN_NAMES: Use ALL_PATTERNS instead. Accessing the old name emits a DeprecationWarning.
• tsignals drift parameter: Passing drift now emits a DeprecationWarning and the value is ignored — trade signals are always computed with a 1-period difference. The parameter will be removed in a future release.
• Inactive drift parameters on cfo, inertia, kst, rsx, chop, accbands, kvo: These indicators accepted a drift parameter that never entered their calculation (dead since the upstream template). Passing drift now emits a DeprecationWarning; the value has never affected the result and the parameter will be removed in a future release. Misleading docstrings were corrected (e.g. cfo's "short period", chop's ATR(drift) pseudocode).

Removed

• Dead code (audit, zero callers): BasePandasObject and _check_na_columns() in core.py; CPR helpers round_to_strike, calculate_option_strikes, detect_cpr_breakout, detect_cpr_rejection; math utils geometric_mean, log_geometric_mean, and the hand-rolled erf; time utils df_dates, df_month_to_date/mtd, df_quarter_to_date/qtd; category_files(); six unused CandleArrays helper methods; the pkg_resources Python < 3.8 fallback in _meta.py; tests/context.py and other unused test scaffolding.
• crossany(): zero callers and zero tests; use cross(a, b, above=True) | cross(a, b, above=False) if needed. Docs entry removed.
• high_low_range() / real_body() wrappers (utils/_candles.py): call sites now use non_zero_range() directly (real_body(open_, close) ≡ non_zero_range(close, open_)).
• ytd alias: use df_year_to_date(). Internal _camelCase2Title inlined at its single call site.
• Unused dev/optional dependencies: stochastic (never imported), cython (no .pyx sources), pytest-cov and pytest-benchmark (no coverage/benchmark runs), isort (black owns formatting). Corresponding Imports probes and the Makefile isort step removed.
• Unused hard dependencies: scipy, scikit-learn, and statsmodels dropped from [project.dependencies] — nothing in the package imports them anymore. _linear_regression_sklearn() removed (see Changed). Stale Imports probes pruned (backtrader, matplotlib, mplfinance, numba, scipy, sklearn, statsmodels, vectorbt, yaml); numba acceleration is unaffected (utils/_njit.py self-detects numba).

Fixed

• tsignals drift corrupted trade signals: trends.diff(drift) computed entries/exits against the state drift bars ago instead of the previous bar, so drift != 1 reported spurious entries mid-trend and missed real crossovers (contradicting the documented diff() behavior). Now always uses a 1-period difference. Output at the default (drift=1) is unchanged.
• Cross-package indicator imports: A submodule import could overwrite a re-exported function of the same name on the parent package (e.g. from pandas_ta_classic.volatility import atr occasionally returned the atr module instead of the function). Resolved.
• Test-fixture regen no longer drops cmo_14 entries when running tests without tulipy installed. Generators now merge with existing JSON instead of overwriting.

[0.6.52] - 2026-06-25

Added

• SMC Liquidity Sweep (smc_sweep) (PR #123 by @Adhyansinghgupta): Bullish and bearish smart-money liquidity sweep indicator for momentum analysis. Available as df.ta.smc_sweep().
• ichimoku append_span (PR #120): New append_span parameter on the ichimoku accessor for span-appending control.

Changed

• CPR numeric encoding (PR #117): calculate_price_position() and calculate_cpr_width() now return np.int8 instead of string labels. CPR_POSITION: 1 (above TC), 0 (inside CPR), -1 (below BC). CPR_WIDTH_CLASS: 1 (wide), 0 (medium), -1 (narrow). All indicator columns are now numeric.
• AnalysisIndicators.__call__ fail-fast: Removed except BaseException: pass swallowing all indicator errors. Indicator exceptions now propagate to callers instead of silently returning None. Code that relied on df.ta.rsi() never raising must add its own error handling.

Fixed

• ichimoku accessor returns DataFrame (PR #116): df.ta.ichimoku() now returns a DataFrame instead of a tuple.
• macdext silent double-fallback (PR #121): Eliminated silent double-fallback for KAMA/MAMA matypes in macdext.

Documentation

• long_run / short_run / xsignals (PR #119): Clarified fast/slow as pre-computed Series; added xa/xb range guidance.
• beta / correl benchmark (PR #118): Documented that benchmark is required for non-None output.

Dependencies

• actions/checkout v7 (PR #122): Bumped from v6 to v7.

---

[0.6.20] - 2026-05-21

Added

• apply_offset / apply_fill helpers (PR #105): Extracted into utils/ and all indicators migrated to use them. Removes ~200 lines of duplicated offset/fill logic.
• Comprehensive candle pattern tests (PR #106): 59 TA-Lib candle patterns covered by CI tests.
• Fluent API chaining (PR #113): DataFrame accessor supports method chaining.
• Property-based testing (PR #114): Hypothesis-based tests added to test suite.
• Full Indicator Name Comments: All indicator files include full indicator names as comments on line 1 (format: # Full Indicator Name (ABBREVIATION)).
• UV Package Manager Support: All documentation includes uv install instructions alongside pip.
• Native Candlestick Patterns: Added native cdl_doji and cdl_inside implementations (no TA-Lib required). Accessible via df.ta.cdl_doji(), df.ta.cdl_inside(), or df.ta.cdl_pattern().

Changed

• Wilder smoothing shared utility (PR #112 remediation): Extracted wilder_smooth() into utils/_wilder.py for TA-Lib-exact cumulative smoothing. Used by dm.py for PLUS_DM/MINUS_DM parity.
• Chained EMA helper (PR #112 remediation): Added _ema_chain() to overlap/ema.py for consistent NaN-stripping in DEMA, TEMA, T3. Reduces repetitive boilerplate by ~60 lines.
• Fixture auto-regeneration: tests/__init__.py now regenerates expected_values.json and regression_snapshots.json on import when TA-Lib is available. make fixtures and make test-all targets added to Makefile.
• tal → talib rename: All test files now import talib directly instead of aliasing as tal, matching source code convention.
• Dead code removal: Removed ~370 instances of unused imports (F401), 8 unused local variables (F841), 1 duplicate method (test_custom_a), 65 useless f-string prefixes (F541), and 279 unnecessary UTF-8 encoding declarations (UP009).
• Code modernization: Applied pyupgrade (UP) and flake8-return (RET) fixes — Optional[X]→X|None, List→list, removed redundant else after return.
• Linreg TA-Lib dispatch: Replaced 6-branch if/elif chain with _TALIB_DISPATCH lookup dictionary.
• PSAR cleanup: Removed import numpy as _np from function body; uses module-level np instead.
• test_strategy: Fixed duplicate test_custom_a method (was shadowed, never ran). Fixed stale column count assertion.

Fixed

• TA-Lib-exact native implementations (PR #112): Corrected native paths for many indicators to match TA-Lib output exactly.
• stdev variance calculation (PR #109): ddof default updated to False for population variance.
• Modularity refactor (PR #108): Issue #46 modularity improvements.
• Accessor usability (PR #107): Issue #48 usability and docs fixes.
• Alpha Vantage integration (PRs #110, #111): Support for both alpha_vantage and alphaVantage libraries; improved import checks and error handling.
• Feature fix PR #112 (PR #115): Follow-up corrections to the TA-Lib-exact implementation from PR #112.

---

[0.5.44] - 2026-04-30

Added

• TA-Lib / tulipy parity indicator set (PR #104): Added wrappers and native implementations for msw, fosc, macdext, macdfix, rocp, rocr, rocr100, stochf, avgprice, medprice, typprice, linregangle, linregintercept, linregslope, mavp, md, stderr, dx, edecay, plus_dm, minus_dm, sarext, avolume, cvi, hvol, emv, marketfi, vosc, and wad.
• Math operator namespace: Added pandas_ta_classic/math/ exposing arithmetic operators (add, sub, mult, div), rolling operators (rolling_max, rolling_min, rolling_sum), and math transforms for TA-Lib/tulipy compatibility.
• Oracle parity suites: Added tests/test_oracle_talib.py and tests/test_oracle_tulipy.py for cross-library validation on shared SPY fixtures.
• 60 native CDL pattern files (PR #87): Added candles/cdl_*.py implementations for 60 patterns. Combined with cdl_doji and cdl_inside, total accessible via cdl_pattern() is 62. TA-Lib is never used for CDL patterns — native implementations take priority regardless of TA-Lib installation. Added shared _cdl_math.py helper.
• 5 Hilbert Transform cycle indicators (PR #83): ht_dcperiod, ht_dcphase, ht_phasor, ht_sine, ht_trendmode. Shared _hilbert.py helper. Cycles category grows from 2 to 7.
• MAMA / FAMA (PR #84): MESA Adaptive Moving Average with FAMA output. Uses Ehlers' adaptive phase computation.
• HT_TRENDLINE (PR #84): Hilbert Transform Instantaneous Trendline. Added to overlap category.
• TSF (PR #85): Time Series Forecast — linear regression projected one period ahead. Matches TA-Lib TSF.
• Beta (PR #86): Asset volatility relative to a benchmark series. Matches TA-Lib BETA.
• CORREL (PR #86): Pearson Correlation Coefficient between two series. Matches TA-Lib CORREL.
• ADXR (PR #89): Average Directional Movement Index Rating — smoothed average of ADX. Matches TA-Lib ADXR.
• CPR (PR #77): Central Pivot Range with 4 calculation methods (classic, camarilla, fibonacci, woodie).
• Chandelier Exit (ce): Trailing stop. CE_L = rolling_max(high, length) - multiplier × ATR, CE_S = rolling_min(low, length) + multiplier × ATR. Default length=22, multiplier=3.0. Returns DataFrame with CE_L_{length}_{multiplier} and CE_S_{length}_{multiplier}.
• 9 New Technical Indicators (Issue #29): LRSI (Laguerre RSI), PMAX (Price Max), VFI (Volume Flow Indicator), MMAR (Madrid Moving Average Ribbon), Rainbow (Rainbow Charts), PO (Projection Oscillator), DSP (Detrended Synthetic Price), TRIXH (TRIX Histogram), VWMACD (Volume Weighted MACD).

Changed

• QQE output columns (PR #97): qqe() now returns 6 columns instead of 3. New columns: QQEb_l (long band), QQEb_s (short band), QQEd (±1 trend direction). Breaking change: code relying on a fixed column count or positional indexing must be updated.
• Updated indicator counts: 164 indicators in Category (was 151); total 224 with native CDL patterns (was 213).
• linreg breaking default change: degrees kwarg now defaults to True (was False) to match TA-Lib. Callers using linreg(close, angle=True) without degrees=False will now receive degrees instead of radians.
• stdev/variance breaking default change: ddof now defaults to 0 (population, was 1 sample) to match TA-Lib. Callers relying on sample variance must pass ddof=1 explicitly.
• natr breaking default change: mamode now defaults to 'rma' (Wilder smoothing, was 'ema') to match TA-Lib. Callers relying on EMA-based NATR must pass mamode='ema' explicitly.
• Native preferred by default for all indicators with talib parameter: Native implementation used by default across all 59 indicators. Callers wanting TA-Lib output must pass talib=True explicitly.
• Enhanced RVGI: Relative Vigor Index now includes histogram column (RVGI − Signal).
• Dynamic Category Discovery: Category dict in _meta.py now built dynamically from filesystem structure. Auto-discovers previously undocumented indicators (cdl_doji, cdl_inside, hwma, ma, drawdown, dm, vp).
• Python Version Support: Rolling 5-version support policy (LATEST-4 through LATEST). Managed dynamically via LATEST_PYTHON_VERSION in .github/workflows/ci.yml.
• Development Status: Changed from Beta to Production/Stable in pyproject.toml.

Performance

• numpy vectorization (PR #88): 15 indicators (QQE, PSAR, HWC, HT_TRENDLINE, SSF, squeeze, squeeze_pro, RVGI, TD_SEQ, TOS_STDEVALL, ALMA, SINWMA, SWMA, TRIMA, VIDYA) now use sliding_window_view instead of pandas .iloc loops. Adds shared _sliding_weighted_ma() utility.
• Numba JIT acceleration (PR #99): 10 indicators (SSF, MCGD, HWMA, RSX, PSAR, Supertrend, QQE, and 3 more) gain optional @njit(cache=True) via numba. Graceful no-op fallback in utils/_njit.py. Enable with pip install pandas-ta-classic[performance]. Speedups: RSX 230×, HWMA 70×, MCGD 43×, SSF 42×, Supertrend 13×, QQE 10×, PSAR 6×.

Fixed

• Oracle test policy: Replaced skip-based oracle tests with explicit assertions for value equivalence or documented divergence. Zero skipped tests.
• TA-Lib compatibility paths: Added/updated talib=True behavior for macdfix, psar, stochrsi, plus_dm, minus_dm.
• Indicator formula parity: Corrected edecay (multiplicative exponential decay), emv scaling (divisor=10000).

---

[0.4.47] - 2026-03-17

Fixed

• Dependency cleanup, pandas 3.0 compat, Windows pool fix (PR #79).
• Initialization and edge-case bugs across 11 indicators (PR #80).
• Numerical bugs in linreg, tsi, brar, bbands, cti (PR #94).
• TA-Lib reference alignment for 8 indicators (PR #81).
• Rolling stats cross-version determinism: Replaced pandas rolling stats with numpy (PR #82).
• None-guards to prevent crashes on short/invalid input across 26 files (PR #95).
• print() → logging across library code (PR #93). Breaking: code catching print output must switch to log capture.
• Strategy dataclass bugs (PR #96). Breaking: some Strategy API surface changed.
• Code of Conduct Contact: Updated enforcement contact from original maintainer email to GitHub Issues.
• PyPI Release Version: Set SETUPTOOLS_SCM_PRETEND_VERSION from the release tag to prevent .dev0 versions being published.
• Version Scheme: Changed from post-release to default scheme. Tagged releases get clean versions (e.g., 0.4.47); commits after a tag get .devN suffix.
• PyPI Image Display: Updated README.md to use absolute GitHub URLs so images render correctly on the PyPI package page.
• CI/CD Shallow Clone: Added fetch-depth: 0 to all checkout steps to ensure full git history for setuptools-scm.
• Version Fallback: Changed fallback from 0.0.0.dev0 to 0.0.0 (PEP 440 compliant).

Added

• Automatic Version Management: Version now determined from git tags via setuptools-scm. Development builds get .dev suffix; tagged releases use the tag exactly. See CONTRIBUTING.md for documentation.

---

[0.3.78] - 2026-02-27

Changed

• Type hints (PR #67 by @rmarcink): Type annotations added across all public function signatures.

---

Pre-0.3.78 — Historical

Items below describe the state of the library as inherited from the original pandas-ta project and pre-2026 development. Preserved for historical reference.

General

• A Strategy Class to help name and group your favorite indicators.
• If a TA Lib is already installed, Pandas TA will run TA Lib's version. (BETA)
• Some indicators have had their mamode kwarg updated with more moving average choices with the Moving Average Utility function ta.ma(). For simplicity, all choices are single source moving averages. This is primarily an internal utility used by indicators that have a mamode kwarg. This includes indicators: accbands, amat, aobv, atr, bbands, bias, efi, hilo, kc, natr, qqe, rvi, and thermo; the default mamode parameters have not changed. However, ta.ma() can be used by the user as well if needed. For more information: help(ta.ma)
  • Moving Average Choices: dema, ema, fwma, hma, linreg, midpoint, pwma, rma, sinwma, sma, swma, t3, tema, trima, vidya, wma, zlma.
• An experimental and independent Watchlist Class located in the Examples Directory that can be used in conjunction with the new Strategy Class.
• Linear Regression (linear_regression) is a new utility method for Simple Linear Regression using Numpy or Scikit Learn's implementation.
• Added utility/convience function, to_utc, to convert the DataFrame index to UTC. See: help(ta.to_utc) Now as a Pandas TA DataFrame Property to easily convert the DataFrame index to UTC.

Breaking / Depreciated Indicators

• Trend Return (trend_return) has been removed and replaced with tsignals. When given a trend Series like close > sma(close, 50) it returns the Trend, Trade Entries and Trade Exits of that trend to make it compatible with vectorbt by setting asbool=True to get boolean Trade Entries and Exits. See help(ta.tsignals)

New Indicators

• Arnaud Legoux Moving Average (alma) uses the curve of the Normal (Gauss) distribution to allow regulating the smoothness and high sensitivity of the indicator. See: help(ta.alma)
• Draw Down (drawdown) calculates the percentage decline from the peak equity of a trading account, or fund. See help(ta.drawdown)
• Candle Patterns (cdl_pattern) If TA Lib is installed, then all those Candle Patterns are available. See the list and examples above on how to call the patterns. See help(ta.cdl_pattern)
• Candle Z Score (cdl_z) normalizes OHLC Candles with a rolling Z Score. See help(ta.cdl_z)
• Correlation Trend Indicator (cti) is an oscillator created by John Ehler in 2020. See help(ta.cti)
• Cross Signals (xsignals) was created by Kevin Johnson. It is a wrapper of Trade Signals that returns Trends, Trades, Entries and Exits. Cross Signals are commonly used for bbands, rsi, zscore crossing some value either above or below two values at different times. See help(ta.xsignals)
• Directional Movement (dm) developed by J. Welles Wilder in 1978 attempts to determine which direction the price of an asset is moving. See help(ta.dm)
• Even Better Sinewave (ebsw) measures market cycles and uses a low pass filter to remove noise. See: help(ta.ebsw)
• Jurik Moving Average (jma) attempts to eliminate noise to see the "true" underlying activity.. See: help(ta.jma)
• Klinger Volume Oscillator (kvo) was developed by Stephen J. Klinger. It is designed to predict price reversals in a market by comparing volume to price.. See help(ta.kvo)
• Schaff Trend Cycle (stc) is an evolution of the popular MACD incorportating two cascaded stochastic calculations with additional smoothing. See help(ta.stc)
• Squeeze Pro (squeeze_pro) is an extended version of "TTM Squeeze" from John Carter. See help(ta.squeeze_pro)
• Tom DeMark's Sequential (td_seq) attempts to identify a price point where an uptrend or a downtrend exhausts itself and reverses. Currently exlcuded from df.ta.strategy() for performance reasons. See help(ta.td_seq)
• Think or Swim Standard Deviation All (tos_stdevall) indicator which returns the standard deviation of data for the entire plot or for the interval of the last bars defined by the length parameter. See help(ta.tos_stdevall)
• Vertical Horizontal Filter (vhf) was created by Adam White to identify trending and ranging markets.. See help(ta.vhf)

Updated Indicators

• Acceleration Bands (accbands) Argument mamode renamed to mode. See help(ta.accbands).
• ADX (adx): Added mamode with default "RMA" and with the same mamode options as TradingView. New argument lensig so it behaves like TradingView's builtin ADX indicator. See help(ta.adx).
• Archer Moving Averages Trends (amat): Added drift argument and more descriptive column names.
• Average True Range (atr): The default mamode is now "RMA" and with the same mamode options as TradingView. See help(ta.atr).
• Bollinger Bands (bbands): New argument ddoff to control the Degrees of Freedom. Also included BB Percent (BBP) as the final column. Default is 0. See help(ta.bbands).
• Choppiness Index (chop): New argument ln to use Natural Logarithm (True) instead of the Standard Logarithm (False). Default is False. See help(ta.chop).
• Chande Kroll Stop (cksp): Added tvmode with default True. When tvmode=False, cksp implements "The New Technical Trader" with default values. See help(ta.cksp).
• Chande Momentum Oscillator (cmo): New argument talib will use TA Lib's version and if TA Lib is installed. Default is True. See help(ta.cmo).
• Decreasing (decreasing): New argument strict checks if the series is continuously decreasing over period length with a faster calculation. Default: False. The percent argument has also been added with default None. See help(ta.decreasing).
• Increasing (increasing): New argument strict checks if the series is continuously increasing over period length with a faster calculation. Default: False. The percent argument has also been added with default None. See help(ta.increasing).
• Klinger Volume Oscillator (kvo): Implements TradingView's Klinger Volume Oscillator version. See help(ta.kvo).
• Linear Regression (linreg): Checks numpy's version to determine whether to utilize the as_strided method or the newer sliding_window_view method. This should resolve Issues with Google Colab and it's delayed dependency updates as well as TensorFlow's dependencies as discussed in Issues #285 and #329.
• Moving Average Convergence Divergence (macd): New argument asmode enables AS version of MACD. Default is False. See help(ta.macd).
• Parabolic Stop and Reverse (psar): Bug fix and adjustment to match TradingView's sar. New argument af0 to initialize the Acceleration Factor. See help(ta.psar).
• Percentage Price Oscillator (ppo): Included new argument mamode as an option. Default is sma to match TA Lib. See help(ta.ppo).
• True Strength Index (tsi): Added signal with default 13 and Signal MA Mode mamode with default ema as arguments. See help(ta.tsi).
• Volume Profile (vp): Calculation improvements. See Pull Request #320 See help(ta.vp).
• Volume Weighted Moving Average (vwma): Fixed bug in DataFrame Extension call. See help(ta.vwma).
• Volume Weighted Average Price (vwap): Added a new parameter called anchor. Default: "D" for "Daily". See Timeseries Offset Aliases for additional options. Requires the DataFrame index to be a DatetimeIndex. See help(ta.vwap).
• Volume Weighted Moving Average (vwma): Fixed bug in DataFrame Extension call. See help(ta.vwma).
• Z Score (zscore): Changed return column name from Z_length to ZS_length. See help(ta.zscore).