PyPortfolioOpt

What it is

PyPortfolioOpt is the library most people meet first when they go looking for Markowitz in Python. The shape is deliberately scikit-learn-ish: expected_returns and risk_models turn a price DataFrame into a mean vector and a covariance matrix, EfficientFrontier takes the pair and solves for weights under max_sharpe(), min_volatility(), efficient_risk() or efficient_return(), and DiscreteAllocation turns the resulting fractions into how many shares to buy with the cash you actually have.

Around that core: shrinkage covariance (Ledoit-Wolf, oracle approximating, exponentially weighted), L2 regularisation and sector constraints through objective_functions, BlackLittermanModel for folding absolute or relative views into the prior, HRPOpt for hierarchical risk parity, CLA for Markowitz's critical line algorithm, and efficient semivariance, CVaR and CDaR frontiers in their own modules.

The project changed hands in 2026, and that is the fact worth knowing before you depend on it. The repository now lives under the PyPortfolio organisation — the old robertmartin8 URL redirects — and carries a GC.OS sponsorship badge, the same stewardship arrangement as sktime. Version 1.6.0 landed on 26 February 2026 after 1.5.6 in December 2024, and its notes read like a rescue rather than a feature release: Python 3.14 and pandas 3 support, Python 3.8 and 3.9 retired, soft dependencies moved out of the core install, CI rebuilt. The published package metadata still names Robert Andrew Martin as maintainer; the release was cut by Franz Király.

Pricing

Free and MIT, with nothing for sale — no pro build, no hosted tier, no data bundle. What is not free of charge in the licence sense is the all-extras depset, which adds matplotlib, plotly, ecos and cvxopt. The last two are GPLv3. That is irrelevant if you are running research on your own machine and material if you redistribute a binary.

Data & coverage

None, and that is by design. Grep the installed package for a network call and you find nothing: no bundled dataset, no fetcher, no vendor client. Asset classes and markets are whatever your price series covers, so the library is as global as your data source is.

Integrations

cvxpy does the solving, so anything cvxpy supports is available by passing solver=. Hard dependencies are numpy, pandas, scipy, scikit-learn and — new in 1.6.0 — scikit-base, the base-class package from the sktime ecosystem, which is the clearest signal of where the project now sits. Weights come back as a dict or an OrderedDict; save_weights_to_file writes csv, json or txt, and everything else is a pandas object you export yourself.

Limitations

• No data and no backtest. If you arrived from a backtesting framework expecting to test a rebalancing rule, this is the wrong shelf — it computes one weight vector per call, and looping it over time is your code.
• Mean-variance is only as good as the expected returns you feed it, and the README says so. Historical means are a poor forecast, which is the reason Black-Litterman and shrinkage are in the package at all.
• DiscreteAllocation.lp_portfolio() needs a mixed-integer solver. It currently defaults to ECOS_BB when ecos is installed and warns that in 1.7.0 the default becomes cvxpy's own choice — code written against today's behaviour will need an explicit solver= argument.
• PyPI still classifies it Development Status 4 - Beta, eight years in.
• 82 open issues and a release cadence measured in years, not weeks. The new maintainers are real but the backlog is real too.

Alternatives

Riskfolio-Lib covers far more risk measures and is the one to reach for if CVaR, drawdown-at-risk or Kelly objectives are the point. skfolio is the scikit-learn-native take, with cross-validation and hyper-parameter search around the optimiser rather than bolted on. Portfolio Optimizer does the same mathematics over HTTP for people who are not writing Python at all, and Portfolio Visualizer is the no-code version of the question.