pomata

A Polars-native quant toolkit — technical indicators, PnL accounting, and performance & risk metrics. Each is a composable pl.Expr, so an entire study is one lazy Polars pipeline, from price to performance.

And it doesn't ask you to trust its numbers — it proves them, each family against the yardstick that catches its bugs: indicators to the float64 floor against an independent reference, PnL and metrics at the edges where every degenerate input has a defined, tested behavior — all under 100% branch coverage.

Install

pip install pomata
# or
uv add pomata

From source:

git clone https://github.com/ilpomo/pomata
cd pomata && uv sync

Dependencies

• Runtime — polars only (>= 1.43). Nothing else is pulled into your environment.
• Python — 3.12, 3.13, 3.14.
• Optional — the differential group (TA-Lib) powers the cross-reference parity tier; the other groups are the contributor gate. See CONTRIBUTING.md.

The data

Every snippet below runs on the same sample: a quarter of real daily bars for AAPL, GOOG, and NVDA, checked into the repo at docs/_static/ohlcv_sample.parquet. Clone the repo to run the snippets as-is, or point read_parquet at your own OHLCV frame:

import polars as pl

ohlcv = pl.read_parquet("docs/_static/ohlcv_sample.parquet")
ohlcv.head(9)

shape: (9, 7)
┌────────────────────────────────┬────────┬────────┬────────┬────────┬────────┬───────────┐
│ datetime                       ┆ ticker ┆ open   ┆ high   ┆ low    ┆ close  ┆ volume    │
│ ---                            ┆ ---    ┆ ---    ┆ ---    ┆ ---    ┆ ---    ┆ ---       │
│ datetime[μs, America/New_York] ┆ str    ┆ f64    ┆ f64    ┆ f64    ┆ f64    ┆ i64       │
╞════════════════════════════════╪════════╪════════╪════════╪════════╪════════╪═══════════╡
│ 2024-01-02 17:00:00 EST        ┆ AAPL   ┆ 185.06 ┆ 186.33 ┆ 181.83 ┆ 183.56 ┆ 82488700  │
│ 2024-01-02 17:00:00 EST        ┆ GOOG   ┆ 138.38 ┆ 139.39 ┆ 136.54 ┆ 138.34 ┆ 20071900  │
│ 2024-01-02 17:00:00 EST        ┆ NVDA   ┆ 49.16  ┆ 49.21  ┆ 47.51  ┆ 48.08  ┆ 411254000 │
│ 2024-01-03 17:00:00 EST        ┆ AAPL   ┆ 182.16 ┆ 183.8  ┆ 181.38 ┆ 182.19 ┆ 58414500  │
│ 2024-01-03 17:00:00 EST        ┆ GOOG   ┆ 137.39 ┆ 139.86 ┆ 137.22 ┆ 139.13 ┆ 18974300  │
│ 2024-01-03 17:00:00 EST        ┆ NVDA   ┆ 47.4   ┆ 48.1   ┆ 47.24  ┆ 47.48  ┆ 320896000 │
│ 2024-01-04 17:00:00 EST        ┆ AAPL   ┆ 180.11 ┆ 181.04 ┆ 178.86 ┆ 179.87 ┆ 71983600  │
│ 2024-01-04 17:00:00 EST        ┆ GOOG   ┆ 138.63 ┆ 139.41 ┆ 136.8  ┆ 136.83 ┆ 18253300  │
│ 2024-01-04 17:00:00 EST        ┆ NVDA   ┆ 47.68  ┆ 48.41  ┆ 47.42  ┆ 47.91  ┆ 306535000 │
└────────────────────────────────┴────────┴────────┴────────┴────────┴────────┴───────────┘

Technical Indicators

75 indicators, each a pl.Expr you compute straight on your price frame, each verified against an independent oracle — and 58 of the 75 cross-checked against TA-Lib too, at a relative 1e-10 (the other 17 have no TA-Lib twin or a documented divergence). On a multi-ticker panel, wrap the call in .over("ticker") so each symbol warms up on its own history (null until the window fills, never a fabricated value):

from pomata.indicators import rsi

ohlcv.with_columns(
    rsi=rsi(pl.col("close"), 14).over("ticker").round(2),
).select("datetime", "ticker", "close", "rsi").tail(9)

shape: (9, 4)
┌────────────────────────────────┬────────┬────────┬───────┐
│ datetime                       ┆ ticker ┆ close  ┆ rsi   │
│ ---                            ┆ ---    ┆ ---    ┆ ---   │
│ datetime[μs, America/New_York] ┆ str    ┆ f64    ┆ f64   │
╞════════════════════════════════╪════════╪════════╪═══════╡
│ 2024-03-26 17:00:00 EDT        ┆ AAPL   ┆ 168.02 ┆ 37.64 │
│ 2024-03-26 17:00:00 EDT        ┆ GOOG   ┆ 150.37 ┆ 64.6  │
│ 2024-03-26 17:00:00 EDT        ┆ NVDA   ┆ 92.4   ┆ 65.66 │
│ 2024-03-27 17:00:00 EDT        ┆ AAPL   ┆ 171.59 ┆ 45.62 │
│ 2024-03-27 17:00:00 EDT        ┆ GOOG   ┆ 150.61 ┆ 64.95 │
│ 2024-03-27 17:00:00 EDT        ┆ NVDA   ┆ 90.09  ┆ 60.06 │
│ 2024-03-28 17:00:00 EDT        ┆ AAPL   ┆ 169.78 ┆ 42.64 │
│ 2024-03-28 17:00:00 EDT        ┆ GOOG   ┆ 150.93 ┆ 65.43 │
│ 2024-03-28 17:00:00 EDT        ┆ NVDA   ┆ 90.2   ┆ 60.24 │
└────────────────────────────────┴────────┴────────┴───────┘

Multi-output indicators (bollinger_bands, macd, stochastic_slow, …) return a single pl.Struct — pick a line with .struct.field(...) or expand every line with .struct.unnest().

PnL Accounting

18 functions that turn a signal into money. An indicator becomes a signed weight; returns_gross / cost_proportional / returns_net turn that into a costed return, and the .shift(1) on the signal is the whole no-look-ahead story — a decision at the close acts on the next bar. Every degenerate input (null / NaN / 0 / ±inf / warm-up) has a defined, documented behavior:

from pomata.pnl import returns_simple, returns_gross, returns_net, cost_proportional

pnl = (
    ohlcv
    .with_columns(
        weight=(rsi(pl.col("close"), 14) > 50).cast(pl.Float64).shift(1).over("ticker"),
        asset_returns=returns_simple(pl.col("close")).over("ticker"),
    )
    .with_columns(
        net=returns_net(
            returns_gross(pl.col("weight"), pl.col("asset_returns")),
            cost_proportional(pl.col("weight"), rate=0.001).over("ticker"),
        ),
    )
)

pnl.select("datetime", "ticker", "weight", pl.col("net").round(4)).tail(9)

shape: (9, 4)
┌────────────────────────────────┬────────┬────────┬─────────┐
│ datetime                       ┆ ticker ┆ weight ┆ net     │
│ ---                            ┆ ---    ┆ ---    ┆ ---     │
│ datetime[μs, America/New_York] ┆ str    ┆ f64    ┆ f64     │
╞════════════════════════════════╪════════╪════════╪═════════╡
│ 2024-03-26 17:00:00 EDT        ┆ AAPL   ┆ 0.0    ┆ -0.0    │
│ 2024-03-26 17:00:00 EDT        ┆ GOOG   ┆ 1.0    ┆ 0.0036  │
│ 2024-03-26 17:00:00 EDT        ┆ NVDA   ┆ 1.0    ┆ -0.0257 │
│ 2024-03-27 17:00:00 EDT        ┆ AAPL   ┆ 0.0    ┆ 0.0     │
│ 2024-03-27 17:00:00 EDT        ┆ GOOG   ┆ 1.0    ┆ 0.0016  │
│ 2024-03-27 17:00:00 EDT        ┆ NVDA   ┆ 1.0    ┆ -0.025  │
│ 2024-03-28 17:00:00 EDT        ┆ AAPL   ┆ 0.0    ┆ -0.0    │
│ 2024-03-28 17:00:00 EDT        ┆ GOOG   ┆ 1.0    ┆ 0.0021  │
│ 2024-03-28 17:00:00 EDT        ┆ NVDA   ┆ 1.0    ┆ 0.0012  │
└────────────────────────────────┴────────┴────────┴─────────┘

Performance & Risk Metrics

60 metric pl.Expr — 43 reduce the whole history into the figure you report (Sharpe, Sortino, Calmar, VaR/CVaR, capture, benchmark-relative, and more), and 17 are their series-valued rolling and running twins, the row-wise drawdown among them. A null is skipped by the reducing metrics; a non-null NaN poisons their result loudly, rather than passing a plausible lie downstream (the rolling twins follow the windowed policy on each page):

from pomata.pnl import equity_curve
from pomata.metrics import sharpe_ratio, total_return, max_drawdown

report = (
    pnl
    .group_by("ticker", maintain_order=True)
    .agg(
        sharpe=sharpe_ratio(pl.col("net"), periods_per_year=252).round(2),
        total_return=total_return(equity_curve(pl.col("net"))).round(4),
        max_drawdown=max_drawdown(equity_curve(pl.col("net"))).round(4),
    )
)

report

shape: (3, 4)
┌────────┬────────┬──────────────┬──────────────┐
│ ticker ┆ sharpe ┆ total_return ┆ max_drawdown │
│ ---    ┆ ---    ┆ ---          ┆ ---          │
│ str    ┆ f64    ┆ f64          ┆ f64          │
╞════════╪════════╪══════════════╪══════════════╡
│ AAPL   ┆ -4.14  ┆ -0.083       ┆ -0.0788      │
│ GOOG   ┆ -0.45  ┆ -0.0286      ┆ -0.1359      │
│ NVDA   ┆ 4.34   ┆ 0.5076       ┆ -0.087       │
└────────┴────────┴──────────────┴──────────────┘

The whole study, one lazy query

Every step above is a pl.Expr, so the four of them fuse into a single lazy pipeline — no intermediate frames, no glue, no second dependency between the steps. Run it on .lazy() and .collect() the same three-row verdict:

report = (
    ohlcv.lazy()
    .with_columns(
        weight=(rsi(pl.col("close"), 14) > 50).cast(pl.Float64).shift(1).over("ticker"),
        asset_returns=returns_simple(pl.col("close")).over("ticker"),
    )
    .with_columns(
        net=returns_net(
            returns_gross(pl.col("weight"), pl.col("asset_returns")),
            cost_proportional(pl.col("weight"), rate=0.001).over("ticker"),
        ),
    )
    .group_by("ticker", maintain_order=True)
    .agg(
        sharpe=sharpe_ratio(pl.col("net"), periods_per_year=252).round(2),
        total_return=total_return(equity_curve(pl.col("net"))).round(4),
        max_drawdown=max_drawdown(equity_curve(pl.col("net"))).round(4),
    )
    .collect()
)

shape: (3, 4)
┌────────┬────────┬──────────────┬──────────────┐
│ ticker ┆ sharpe ┆ total_return ┆ max_drawdown │
│ ---    ┆ ---    ┆ ---          ┆ ---          │
│ str    ┆ f64    ┆ f64          ┆ f64          │
╞════════╪════════╪══════════════╪══════════════╡
│ AAPL   ┆ -4.14  ┆ -0.083       ┆ -0.0788      │
│ GOOG   ┆ -0.45  ┆ -0.0286      ┆ -0.1359      │
│ NVDA   ┆ 4.34   ┆ 0.5076       ┆ -0.087       │
└────────┴────────┴──────────────┴──────────────┘

Same three numbers, arrived at as one query: the indicator fed the signal, the signal fed the PnL, the PnL fed the metrics, and the optimizer fused the lot. Momentum paid in a quarter NVDA ran and cost a little where AAPL slid — the .over("ticker") keeps a three-name panel, or a five-hundred-name one, equally separate.

Correctness

Verified, not asserted. Every function is written twice: the shipped pl.Expr, and a second, independent oracle that shares no code with it. The two must agree — on fixed series, frozen golden masters, and thousands of fuzzed inputs, under 100% branch coverage — or the build is red.

Each family is then held to the yardstick that catches its bugs: indicators to the digit against an independent oracle — and, for the 58 of 75 with a twin, against the public TA-Lib reference too; PnL and metrics at the edges, where every degenerate input has a defined behavior.

The full account — the precision guarantee, the receipts, and exactly what is and is not claimed — is on the correctness page.

Where pomata fits

pomata is for the quant already working in Polars. Each function is a free-standing pl.Expr with polars as the only runtime dependency, composable across eager, lazy, single-series, and grouped (.over) contexts — so the everyday primitives live in one coherent toolkit instead of a wired-together stack.

It is vectorized analytics and accounting: indicators, total mark-to-market PnL, and metrics. It is not an execution engine — no order fills, no event loop, no lot accounting.

Project

• Contributing — see CONTRIBUTING.md; the full gate (ruff and codespell, mypy / pyright / pyrefly gating with ty advisory, doctests, 100% branch coverage) runs on every commit.
• License — MIT, see LICENSE.
• Citation — CITATION.cff.