ruin

A Polars-first risk metrics library for quant hedge funds.

Installation

pip install pyruin
# or with optional SciPy extras (reserved for future use):
pip install "pyruin[stats]"

Distributed on PyPI as pyruin (the name ruin was taken). The import name is unchanged: import ruin.

Quickstart

import polars as pl
import ruin

returns = pl.Series([0.01, -0.02, 0.03, -0.01, 0.02, ...])  # daily returns

sr = ruin.sharpe_ratio(returns, periods_per_year=252)
mdd = ruin.max_drawdown(returns)
df = ruin.summary(returns, periods_per_year=252)

Module Overview

Composition Principle

This library provides building blocks. If you want a bundled metric, call summary() or compose the primitives yourself. Every function returns exactly one thing. We will not add bundled convenience functions.

Composition in Polars is cheap. A custom bundle is two lines:

import polars as pl
import ruin

# Rolling Sharpe with bootstrap confidence intervals
returns = pl.Series(your_returns)

# Rolling Sharpe (returns a Series)
rs = ruin.rolling_sharpe(returns, window=60, periods_per_year=252)

# Bootstrap CI on the scalar Sharpe
point, lo, hi = ruin.bootstrap_metric(
    lambda r: ruin.sharpe_ratio(r, periods_per_year=252),
    returns,
    n_samples=1000,
    confidence=0.95,
)
print(f"Sharpe: {point:.3f} ({lo:.3f}, {hi:.3f})")

Call multiple functions and compose:

metrics = {
    "sharpe": ruin.sharpe_ratio(returns, periods_per_year=252),
    "sortino": ruin.sortino_ratio(returns, periods_per_year=252),
    "max_dd": ruin.max_drawdown(returns),
    "cvar": ruin.conditional_value_at_risk(returns),
}

What This Library Is Not

The following are explicitly out of scope. Do not open issues requesting them.

• Money-weighted returns / IRR / fee crystallization — needs cash flows; belongs in a sibling library.
• Portfolio construction — optimization, risk parity, mean-variance, Black-Litterman.
• Factor models — Fama-French, Barra-style attribution, Brinson attribution.
• Backtesting engine — trade simulation, transaction costs, execution modeling.
• Data fetching — no yfinance, no Bloomberg, no CSV readers.
• Time series hygiene — resampling, calendar alignment, gap filling, corporate actions.
• Stress testing / Monte Carlo / scenario analysis.
• Forecasting — no GARCH, EWMA, regime models, volatility prediction.
• Plotting — use Polars + Altair/Plotly/Matplotlib directly.
• Pandas support — use Polars. pd.Series.to_numpy() converts in one call if needed.
• Cornish-Fisher VaR — reserved for a future ruin[stats] extra.

Design Principles

1. Returns in, numbers out. No prices, no positions, no trades, no I/O.
2. Polars-first, minimal deps. Runtime: polars + numpy only.
3. Pure functions. No side effects, no hidden state.
4. One function, one metric, one return type. Scalar → float, rolling → pl.Series.
5. Explicit over implicit. periods_per_year is always explicit. risk_free is per-period.
6. NaN handling. Drop by default; strict=True in summary() raises instead.

Sign Conventions

• Drawdowns are non-positive: -0.23 means 23% drawdown.
• VaR and CVaR are positive loss magnitudes (desk convention): 0.02 means "lose at most 2%."

See docs/conventions.md for library-wide behaviour (sign conventions, NaN policy, input types, alignment). See docs/metrics.md for per-metric definitions, formulas, assumptions, and citations.