Polars Extension for Ta-Lib

Version 0.2.1 bundles TA-Lib 0.8.1: all 201 batch indicators, including SuperTrend, VWAP, HMA, KDJ and Heikin-Ashi. Python 3.10+; Polars 1.20+ and 2.x. Binary wheels include the C library.

• Upgrade notes, platform matrix and source builds
• Install the skill for Codex, Claude Code or Cursor
• Reproducible performance results

Native AI plugin installation

The repository supports the same native marketplace/install flow as Shioaji:

# Claude Code
claude plugin marketplace add Yvictor/polars_ta_extension
claude plugin install polars-talib@polars-ta-extension

# Codex
codex plugin marketplace add Yvictor/polars_ta_extension
codex plugin add polars-talib@polars-ta-extension

See skill installation for invocation, updates, local checkouts, standalone installation, and Cursor support. The plugin supplies coding guidance; install the Python library separately.

Getting Started

pip install 'polars-talib==0.2.1'
# or, in a uv project
uv add 'polars-talib==0.2.1'

and

import polars
import polars_talib as plta

Usage

single symbol usage

df.with_columns(
    pl.col("close").ta.ema(5).alias("ema5"),
    pl.col("close").ta.macd(12, 26, 9).struct.field("macd"),
    pl.col("close").ta.macd(12, 26, 9).struct.field("macdsignal"),
    pl.col("open").ta.cdl2crows(pl.col("high"), pl.col("low"), pl.col("close")).alias("cdl2crows"),
    pl.col("close").ta.wclprice("high", "low").alias("wclprice"),
)

typed namespace access

plta.col(...) returns the same .ta namespace with a static return type, so editors and type checkers resolve every indicator method:

df.with_columns(
    plta.col("close").ema(5).alias("ema5"),
    plta.col(pl.col("close") * 2).rsi(14).alias("rsi_doubled"),
)

multiple symbol usage using over syntax

Sort by symbol and timestamp before computing history-dependent indicators.

df.with_columns(
    pl.col("close").ta.ema(5).over("symbol").alias("ema5"),
    pl.col("close").ta.macd(12, 26, 9).over("symbol").struct.field("macd"),
    pl.col("close").ta.macd(12, 26, 9).over("symbol").struct.field("macdsignal"),
    pl.col("open").ta.cdl2crows(
        pl.col("high"), pl.col("low"), pl.col("close")
    ).over("symbol").alias("cdl2crows"),
    pl.col("close").ta.wclprice("high", "low").over("symbol").alias("wclprice"),
)

missing values

A Polars null marks a missing observation. Leading nulls are skipped as warm-up. After that, every output row that can depend on a null input is null: to the end of the partition for recursive indicators such as RSI, EMA, MACD or OBV, and for the indicator's lookback window for window-bounded ones such as MAX, WILLR or the math operators. A null never turns into a plausible number, and under .over("symbol") it only affects its own symbol. NaN and infinity are ordinary float values and are passed to TA-Lib unchanged, so its results for them are undefined; use pl.col("close").fill_nan(None) when NaN means missing. To restart an indicator after each gap instead, or to skip gaps:

# restart after every gap, per symbol
gap = pl.col("close").is_null().cum_sum()
df = df.with_columns(pl.col("close").ta.rsi(14).over("symbol", gap).alias("rsi_restart"))

# compute over observed rows only (e.g. market-closed days), then join back
observed = df.filter(pl.col("close").is_not_null()).with_columns(
    pl.col("close").ta.rsi(14).over("symbol").alias("rsi_observed")
)
df = df.join(observed.select("symbol", "date", "rsi_observed"), on=["symbol", "date"], how="left")

usage just like talib.abstract with more flexible

df.with_columns(
    plta.ht_dcperiod(),
    plta.ht_dcperiod(pl.col("close")),
    plta.aroon(),
    plta.aroon(pl.col("high"), pl.col("low"), timeperiod=10),
    plta.wclprice(),
    plta.wclprice(
        pl.col("high"), pl.col("low"), pl.col("close"), 
        timeperiod=10
    ),
)

Performance

See the 0.2.0 release benchmark for a controlled comparison against 0.1.6, raw samples and a reproducible script. The notebook comparison below is historical and measures a different workload.

Polars with polars_talib

%%timeit
df = p.with_columns(
    plta.sma(timeperiod=5).over("Symbol").alias("sma5"),
    plta.macd(fastperiod=10, slowperiod=20, signalperiod=5).over("Symbol").alias("macd"),
    plta.stoch(pl.col("high"), pl.col("low"), pl.col("close"), fastk_period=14, slowk_period=7, slowd_period=7).over("Symbol").alias("stoch"),
    plta.wclprice().over("Symbol").alias("wclprice"),
).with_columns(
    pl.col("macd").struct.field("macd"),
    pl.col("macd").struct.field("macdsignal"),
    pl.col("macd").struct.field("macdhist"),
    pl.col("stoch").struct.field("slowk"),
    pl.col("stoch").struct.field("slowd"),
).select(
    pl.exclude("stoch")
).filter(
    pl.col("Symbol") == "AAPL"
).collect()

135 ms ± 5.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Pandas with talib

%%timeit
df["sma5"] = df.groupby("Ticker")["close"].transform(lambda x: ta.SMA(x, timeperiod=5))
df["macd"] = df.groupby("Ticker")["close"].transform(lambda x: ta.MACD(x, fastperiod=10, slowperiod=20, signalperiod=5)[0])
df["macdsignal"] = df.groupby("Ticker")["close"].transform(lambda x: ta.MACD(x, fastperiod=10, slowperiod=20, signalperiod=5)[1])
df["macdhist"] = df.groupby("Ticker")["close"].transform(lambda x: ta.MACD(x, fastperiod=10, slowperiod=20, signalperiod=5)[2])
df["slowk"] = df.groupby("Ticker").apply(lambda x: ta.STOCH(x, fastk_period=14, slowk_period=7, slowd_period=7)).droplevel(0)["slowk"] 
df["slowd"] = df.groupby("Ticker").apply(lambda x: ta.STOCH(x, fastk_period=14, slowk_period=7, slowd_period=7)).droplevel(0)["slowd"]
df["wclprice"] = df.groupby("Ticker").apply(lambda x: ta.WCLPRICE(x)).droplevel(0)
df.loc["AAPL"]

19.2 s ± 367 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

That historical notebook measured about 150x for this workload; see basic.ipynb

Supported Indicators and Functions

import polars_talib as plta

# list of functions
plta.get_functions()

# dict of functions by group
plta.get_function_groups()

The complete 201-indicator catalog includes every function in the pinned upstream release. New examples:

df.with_columns(
    plta.supertrend().alias("supertrend"),
    plta.vwap().alias("vwap"),
    pl.col("close").ta.hma(timeperiod=20).alias("hma"),
    plta.kdj().alias("kdj"),
)

Multi-output indicators return Struct columns. For example, plta.supertrend().struct.field("trend") returns the Int32 trend flag.