varlab

Python package for Value at Risk and Expected Shortfall.

📦 Installation

Install from PyPI:

pip install varlab

🧰 What varlab Provides

varlab focuses on practical market risk workflows:

• Value at Risk (var)
• Expected Shortfall (es)
• Rolling and expanding estimators (rolling_var, rolling_es, expanding_var, expanding_es)
• VaR backtesting diagnostics with a concise ASCII report and structured outputs
• PIT-based distribution diagnostics (including randomized PIT for empirical/discrete cases)

Both VaR and ES support:

• method="empirical" (historical)
• method="parametric" with distribution="normal" or distribution="t"
• Optional exponential weighting in empirical mode via lamb
• 1-day and multi-day horizons (n_days)
• Optional sample mean component (mean="sample")

⚡ Quick Start

from varlab import rolling_var

var_estimates = rolling_var(
    returns,
    window=60,
    method="empirical",
    confidence=0.99,
    n_days=1,
    distribution="normal",
    mean="sample",
).shift(1).fillna(0)

Parametric Example (Student-t)

If df is not provided, degrees of freedom are estimated by MLE.

from varlab import var, es

var_t = var(
    returns,
    method="parametric",
    distribution="t",
    confidence=0.99,
    mean="sample",
)

es_t = es(
    returns,
    method="parametric",
    distribution="t",
    confidence=0.99,
    mean="sample",
)

Exponentially Weighted Historical Risk

For empirical VaR/ES you can apply exponential weighting by setting lamb:

from varlab import var, es

var_ew = var(returns, method="empirical", confidence=0.99, lamb=0.97)
es_ew = es(returns, method="empirical", confidence=0.99, lamb=0.97)

Backtesting Suite

Run a full backtesting diagnostic:

from varlab import backtesting

back_res = backtesting.diagnostic.run(
    returns=returns,
    exceedances=exceedances,
    confidence=0.99,
    window_type="rolling",
    pit_case="discrete",
    alpha=0.05,
    window=60,
)

Example report:

==================================================================
                  VALUE AT RISK BACKTEST REPORT
==================================================================
Confidence level: 99.00% | Sample size: 4445 | Date: 2026-03-08
Overall result: FAIL

COVERAGE (4/4)
------------------------------------------------------------------
Exact Binomial                                          PASS
Kupiec Pof                                              PASS
Christoffersen Conditional                              PASS
Basel Traffic Light                                     PASS

DISTRIBUTION (0/3)
------------------------------------------------------------------
Uniformity                                              FAIL
Independence                                            FAIL
Berkowitz                                               FAIL

INDEPENDENCE (2/2)
------------------------------------------------------------------
Christoffersen                                          PASS
Loss Quantile                                           PASS

The diagnostic object is typed and machine-friendly, so you can inspect details programmatically:

back_res.results['independence']["christoffersen"]

Test Families Included

• Coverage: exact binomial, Kupiec POF, Christoffersen conditional coverage, Basel traffic light (when confidence is 99%)
• Distribution: PIT-based diagnostics (uniformity, independence, Berkowitz)
• Independence: Christoffersen exceedance independence, loss-quantile independence

Probability Integral Transform (PIT)

Distribution diagnostics rely on PIT values:

• pit_case="continuous" for parametric settings
• pit_case="discrete" for empirical settings (randomized PIT is used)

Window logic can be:

• window_type="rolling"
• window_type="expanding"

Public API

from varlab import (
    var,
    es,
    rolling_var,
    rolling_es,
    expanding_var,
    expanding_es,
    backtesting,
)

Notes

• Inputs can be plain arrays/lists or pandas objects.
• Risk values are returned as positive loss magnitudes.
• For multi-asset parametric workflows, pass a return matrix and portfolio weights.
• Shift rolling/expanding estimates by one step when using them for out-of-sample backtests.

License

MIT © 2025 — Developed with ❤️ by Lorenzo Santarsieri