portlab

A portfolio risk and optimization library for Python, built around one principle: a risk number you cannot backtest is an opinion. Every estimator here is cross-checked against an independent implementation and ships with the tests that pin its conventions down ��� see VALIDATION.md for the verification record, including the defects that process found and fixed.

Install

pip install portlab

Requires Python 3.10+. Core dependencies: numpy, pandas, scipy, scikit-learn, statsmodels, plotly. The CVaR/CDaR LP solvers use scipy; optional solver backends (cvxpy, OSQP, Gurobi) are imported lazily when requested. Market-data loaders are also lazy: install portlab[data] for the yfinance / pandas-datareader / pyarrow stack used by portlab.data.

From source instead:

pip install git+https://github.com/Realosunboy6/portlab.git

Quick start

import pandas as pd
from portlab.risk.var import historical_var
from portlab.risk.backtest import exceptions, kupiec_pof, christoffersen_ind

rets = pd.Series(...)                       # daily simple returns
var = historical_var(rets, alpha=0.05, window=250)
exc = exceptions(rets, var)                 # nullable boolean Series
print(kupiec_pof(exc, 0.05))                # coverage test
print(christoffersen_ind(exc))              # gap-aware independence test

Reproduce the published validation anchor

import pandas as pd
from portlab.risk.var import historical_var
from portlab.risk.backtest import exceptions, kupiec_pof, christoffersen_ind

px = pd.read_csv("https://stooq.com/q/d/l/?s=spy.us&i=d",
                 parse_dates=["Date"], index_col="Date")["Close"].sort_index()
rets = px.loc["2019-01-02":].pct_change().dropna()

exc = exceptions(rets, historical_var(rets, alpha=0.05, window=250))
print(kupiec_pof(exc, 0.05))
print(christoffersen_ind(exc))

Recorded on the validation dataset (through 2026-10-09): n=1,703, 86 exceptions, Kupiec LR = 0.0089, Christoffersen LR_ind = 10.589773. Live data adds recent days, so counts grow slightly ��� the LR values should still land on the anchors to ~3 decimals.

SPY's VaR passes the coverage test but fails independence ��� the exceptions bunch in Feb���Apr 2020. That distinction is the point of the library: cite conditional coverage, not Kupiec alone.

What's inside

Conventions (read this before trusting a number)

• Inputs are simple returns (positive = gain). VaR/ES are reported as positive numbers = losses. An exception is returns < -VaR.
• alpha means two things, by module family ��� in portlab.risk.* it is the tail probability (0.05 = 95% VaR); in portlab.metrics.* and the CVaR optimizer it is the confidence level (0.95). Every docstring states which.
• Rolling estimators use forecast alignment: the value stamped at day t is built only from data strictly before t. Backtests score exactly the stamped days ��� no look-ahead, no warmup padding.
• exceptions() returns a nullable-boolean Series. Days that cannot be scored (VaR warmup, missing returns, gaps) are <NA>; Kupiec counts only scored days and Christoffersen counts only transitions between genuinely adjacent scored days.

Scope

As of 0.2.0 this is the complete toolkit ��� the earlier public core plus the modules that were initially held back (monitoring, tax, Monte Carlo, factors, tactical, ALM, derivatives, data, reporting), each verified before release; see VALIDATION.md for the 0.2.0 verification record. What is not in this repo, by design: broker credentials, live account data, and the Sentinel web application that consumes this library. The algorithms are standard textbook methods (Kupiec 1995; Christoffersen 1998; Rockafellar & Uryasev 2000; Black & Litterman 1992) ��� the contribution is the implementation discipline and the validation suite.

Citation

If you use portlab in research, please cite it via CITATION.cff (GitHub's "Cite this repository" button reads the same file).

License

MIT ��� see LICENSE.

Author

Ibrahim Oyeyinka ��� PhD candidate, Industrial & Systems Engineering, Northern Illinois University.