Open Source
Portfolio Optimization & Risk Management

Load raw inputs such as prices, volumes, alternative datasets, or externally discovered alphas and signals. Handle heterogeneous data sources and formats with a data-agnostic design, ensuring compatibility across research and production workflows.

Encode prior beliefs and distributional assumptions before optimization using empirical estimates, Black–Litterman formulations, factor models or pooling methods such as entropy and opinion pooling. Priors may also be built from synthetic or stressed distributions, allowing scenario views to feed directly into the optimization process.

Apply a broad range of methods, from mean–risk models such as mean–variance, CVaR, and CDaR, to risk budgeting and maximum diversification, as well as hierarchical approaches like HRP and HERC. Utilise distributionally robust techniques built on ambiguity sets or combine estimators through ensemble methods. Methods support a diverse set of objective functions, constraints and regularizations.

Calibrate and fine-tune estimators and optimization hyper-parameters with methods such as grid search or randomized search, combined with time-aware cross-validation to prevent look-ahead bias. Parallelise computations to accelerate evaluation, ensuring leakage-safe selection, fair model comparison and robust out-of-sample performance.

Assess model quality with time-aware validation methods such as walk-forward, combinatorial purged cross-validation and resampling-based backtests, designed to prevent look-ahead bias. Extend evaluation beyond standard performance and risk metrics to include stress tests, scenario generation with vine copulas and factor exposure for a comprehensive view of out-of-sample robustness.

Estimate return distributions with advanced methods such as vine copulas to capture tail dependencies and improve extrapolation of joint extreme events. Enable conditional sampling for scenario-specific simulations, supporting stress tests that are both extreme and plausible.

Estimate expected returns using approaches ranging from historical and exponentially weighted averages to equilibrium models and shrinkage techniques, or inject custom alpha signals directly. These estimators provide stable inputs for portfolio construction, reducing noise and improving the reliability of optimization outcomes.

Estimate covariance matrices with a variety of methods including Ledoit–Wolf, Oracle Approximating Shrinkage, Graphical Lasso, Gerber covariance, Denoising and Detoning techniques or directly ingest external covariance matrices. These estimators reduce error, filter out noise and multicollinearity, and provide more reliable risk inputs for portfolio optimization.

Integrate any estimator that follows the scikit-learn API, from built-in models to fully customized implementations. Users can plug in external algorithms, transformations, or data-driven models as native components without breaking compatibility.

Compute dependence structures with distance estimators such as variation of information or Kendall distance. These methods capture both linear and nonlinear relationships between assets, supporting clustering and advanced portfolio construction techniques.

Quantify uncertainty in expected returns and covariances with empirical and bootstrap techniques to capture sampling variability. Define ambiguity sets to model parameter uncertainty explicitly, providing a framework for distributionally robust optimization and more stable allocations under estimation risk.

Extend the scikit-learn framework with time-aware methods tailored for financial data. Go beyond standard k-fold with walk-forward validation, combinatorial purged cross-validation and multiple randomized backtests, designed for leakage-safe training and fair model selection under temporal dependence. Apply calendar-based rebalancing rules to align with real-world considerations.

The explicit output of the pipeline, preserving a complete audit trail of every step for inspection, modification and reproducibility. A portfolio object consolidates weights, analytics, and performance, supports visualisation and reporting and is fully picklable for efficient storage and seamless reuse in future workflows.