skfolio#

skfolio is a Python library for portfolio optimization, factor model construction, and risk management built on top of scikit-learn. It offers a unified interface and tools compatible with scikit-learn to build, fine-tune, cross-validate, and stress-test portfolio models.

It is distributed under the open-source 3-Clause BSD license.

skfolio is backed by Skfolio Labs, which provides enterprise support and SLAs for institutions.

Portfolio optimization examples gallery from skfolio

Installation#

skfolio requires Python 3.10 or later and can be installed with:

pip install -U skfolio

See the installation guide for the full dependency list, for conda-forge and for the mixed-integer solvers.

LLM-friendly documentation#

The documentation follows the llms.txt convention and provides token-efficient Markdown alongside the HTML site:

  • Start with llms.txt to find relevant pages.

  • Read only the Markdown pages needed by appending .md to their HTML URLs, for example factor_models.html.md.

  • Use llms-full.txt only when the complete documentation is required in a single file.

Contribution#

We welcome contributions of all kinds. Whether it’s reporting a bug, suggesting an improvement, or submitting code, your input helps make skfolio better. See the contributing guide to get started.

Key Concepts#

Since the development of modern portfolio theory by Markowitz (1952), mean-variance optimization (MVO) has received considerable attention.

Unfortunately, it faces a number of shortcomings, including high sensitivity to the input parameters (expected returns and covariance), weight concentration, high turnover, and poor out-of-sample performance.

It is well-known that naive allocation (1/N, inverse-vol, etc.) tends to outperform MVO out-of-sample (DeMiguel, 2007).

Numerous approaches have been developed to alleviate these shortcomings (shrinkage, additional constraints, regularization, uncertainty set, higher moments, Bayesian approaches, coherent risk measures, left-tail risk optimization, distributionally robust optimization, factor model, risk-parity, hierarchical clustering, ensemble methods, pre-selection, etc.).

Given the large number of methods, and the fact that they can be combined, there is a need for a unified framework with a machine-learning approach to perform model selection, validation, and parameter tuning while mitigating the risk of data leakage and overfitting.

This framework is built on scikit-learn’s API.

Available models#

  • Portfolio Optimization:
    • Naive:

      • Equal-Weighted

      • Inverse-Volatility

      • Random (Dirichlet)

    • Convex:

      • Mean-Risk

      • Risk Budgeting

      • Maximum Diversification

      • Distributionally Robust CVaR

      • Benchmark Tracker

    • Clustering:

      • Hierarchical Risk Parity

      • Hierarchical Equal Risk Contribution

      • Schur Complementary Allocation

      • Nested Clusters Optimization

    • Ensemble Methods:

      • Stacking Optimization

  • Prior Estimator:

    • Empirical

    • Characteristics-Based Cross-Sectional Factor Model:

      • 46 descriptors across 17 families (e.g. value, size, momentum, profitability)

      • Factor Exposures

      • Cross-Sectional Regression

      • Alpha Estimators

      • Forecast Evaluation

      • Ex-post and Ex-ante Attribution

    • Time-Series Factor Model

    • Black & Litterman

    • Synthetic Data (Stress Test, Factor Stress Test)

    • Entropy Pooling

    • Opinion Pooling

  • Expected Returns Estimator:

    • Empirical

    • Exponentially Weighted

    • Equilibrium

    • Shrinkage

  • Covariance Estimator:

    • Empirical

    • Gerber

    • Denoising

    • Detoning

    • Exponentially Weighted

    • Regime-Adjusted Exponentially Weighted

    • Ledoit-Wolf

    • Oracle Approximating Shrinkage

    • Shrunk Covariance

    • Graphical Lasso CV

    • Implied Covariance

  • Variance Estimator:

    • Empirical

    • Exponentially Weighted

    • Regime-Adjusted Exponentially Weighted

  • Distance Estimator:

    • Pearson Distance

    • Kendall Distance

    • Spearman Distance

    • Covariance Distance (based on any of the above covariance estimators)

    • Distance Correlation

    • Variation of Information

  • Distribution Estimator:
    • Univariate:

      • Gaussian

      • Student’s t

      • Johnson Su

      • Normal Inverse Gaussian

    • Bivariate Copula

      • Gaussian Copula

      • Student’s t Copula

      • Clayton Copula

      • Gumbel Copula

      • Joe Copula

      • Independent Copula

    • Multivariate

      • Vine Copula (Regular, Centered, Clustered, Conditional Sampling)

  • Uncertainty Set Estimator:
    • On Expected Returns:

      • Empirical

      • Circular Bootstrap

    • On Covariance:

      • Empirical

      • Circular Bootstrap

  • Pre-Selection Transformers:

    • Non-Dominated Selection

    • Select K Extremes (Best or Worst)

    • Drop Highly Correlated Assets

    • Select Non-Expiring Assets

    • Select Complete Assets (handle late inception, delisting, etc.)

    • Drop Zero Variance

  • Cross-Sectional Transformers:

    • Standard Scaler (z-score)

    • Percentile Rank Scaler

    • Gaussian Rank Scaler (rank gaussianization)

    • Winsorizer (percentile clipping)

    • Tanh Shrinker (smooth outlier shrinkage)

  • Cross-Validation and Model Selection:

    • Compatible with all sklearn methods (KFold, etc.)

    • Walk Forward

    • Combinatorial Purged Cross-Validation

    • Multiple Randomized Cross-Validation

    • Covariance Forecast Evaluation

    • Online Predict and Online Score

  • Hyper-Parameter Tuning:

    • Compatible with all sklearn methods (GridSearchCV, RandomizedSearchCV)

    • Online Grid Search and Online Randomized Search

  • Risk Measures:

    • Variance

    • Semi-Variance

    • Mean Absolute Deviation

    • First Lower Partial Moment

    • CVaR (Conditional Value at Risk)

    • EVaR (Entropic Value at Risk)

    • Worst Realization

    • CDaR (Conditional Drawdown at Risk)

    • Maximum Drawdown

    • Average Drawdown

    • EDaR (Entropic Drawdown at Risk)

    • Ulcer Index

    • Gini Mean Difference

    • Value at Risk

    • Drawdown at Risk

    • Entropic Risk Measure

    • Fourth Central Moment

    • Fourth Lower Partial Moment

    • Skew

    • Kurtosis

  • Optimization Features:

    • Minimize Risk

    • Maximize Returns

    • Maximize Utility

    • Maximize Ratio

    • Transaction Costs

    • Management Fees

    • L1 and L2 Regularization

    • Weight Constraints

    • Group Constraints

    • Budget Constraints

    • Tracking Error Constraints

    • Turnover Constraints

    • Cardinality and Group Cardinality Constraints

    • Threshold (Long and Short) Constraints

Quickstart#

The code snippets below are designed to introduce the functionality of skfolio so you can start using it quickly. It follows the same API as scikit-learn.

Imports#

from sklearn import set_config
from sklearn.model_selection import (
    GridSearchCV,
    KFold,
    RandomizedSearchCV,
    train_test_split,
)
from sklearn.pipeline import Pipeline
from scipy.stats import loguniform

from skfolio import RatioMeasure, RiskMeasure
from skfolio.datasets import (
    load_factors_dataset,
    load_sp500_dataset,
    make_synthetic_characteristics,
)
from skfolio.descriptor import (
    BookToPrice,
    CashFlowToPrice,
    EWMarketBeta,
    EWMomentum,
    EWResidualVolatility,
    EWVolatility,
    LogMarketCap,
    SalesToPrice,
)
from skfolio.distribution import VineCopula
from skfolio.factor_exposure import (
    DerivedFactor,
    FixedWeightedFactor,
    GlobalFactor,
    OneHotCategoricalFactors,
)
from skfolio.model_selection import (
    CombinatorialPurgedCV,
    WalkForward,
    cross_val_predict,
)
from skfolio.moments import (
    DenoiseCovariance,
    DetoneCovariance,
    EWMu,
    GerberCovariance,
    ShrunkMu,
)
from skfolio.optimization import (
    MeanRisk,
    HierarchicalRiskParity,
    NestedClustersOptimization,
    ObjectiveFunction,
    RiskBudgeting,
)
from skfolio.pre_selection import SelectKExtremes
from skfolio.preprocessing import prices_to_returns
from skfolio.prior import (
    BlackLitterman,
    CharacteristicsFactorModel,
    EmpiricalPrior,
    EntropyPooling,
    TimeSeriesFactorModel,
    OpinionPooling,
    SyntheticData,
)
from skfolio.uncertainty_set import BootstrapMuUncertaintySet

Load Dataset#

prices = load_sp500_dataset()

Train/Test split#

X = prices_to_returns(prices)
X_train, X_test = train_test_split(X, test_size=0.33, shuffle=False)

Minimum Variance#

model = MeanRisk()

Fit on Training Set#

model.fit(X_train)

print(model.weights_)

Predict on Test Set#

portfolio = model.predict(X_test)

print(portfolio.annualized_sharpe_ratio)
print(portfolio.summary())

Maximum Sortino Ratio#

model = MeanRisk(
    objective_function=ObjectiveFunction.MAXIMIZE_RATIO,
    risk_measure=RiskMeasure.SEMI_VARIANCE,
)

Denoised Covariance & Shrunk Expected Returns#

model = MeanRisk(
    objective_function=ObjectiveFunction.MAXIMIZE_RATIO,
    prior_estimator=EmpiricalPrior(
        mu_estimator=ShrunkMu(), covariance_estimator=DenoiseCovariance()
    ),
)

Uncertainty Set on Expected Returns#

model = MeanRisk(
    objective_function=ObjectiveFunction.MAXIMIZE_RATIO,
    mu_uncertainty_set_estimator=BootstrapMuUncertaintySet(),
)

Weight Constraints & Transaction Costs#

model = MeanRisk(
    min_weights={"AAPL": 0.10, "JPM": 0.05},
    max_weights=0.8,
    transaction_costs={"AAPL": 0.0001, "RRC": 0.0002},
    groups=[
        ["Equity"] * 3 + ["Fund"] * 5 + ["Bond"] * 12,
        ["US"] * 2 + ["Europe"] * 8 + ["Japan"] * 10,
    ],
    linear_constraints=[
        "Equity <= 0.5 * Bond",
        "US >= 0.1",
        "Europe >= 0.5 * Fund",
        "Japan <= 1",
    ],
)
model.fit(X_train)

Risk Parity on CVaR#

model = RiskBudgeting(risk_measure=RiskMeasure.CVAR)

Risk Parity & Gerber Covariance#

model = RiskBudgeting(
    prior_estimator=EmpiricalPrior(covariance_estimator=GerberCovariance())
)

Nested Cluster Optimization with Cross-Validation and Parallelization#

model = NestedClustersOptimization(
    inner_estimator=MeanRisk(risk_measure=RiskMeasure.CVAR),
    outer_estimator=RiskBudgeting(risk_measure=RiskMeasure.VARIANCE),
    cv=KFold(),
    n_jobs=-1,
)

Randomized Search of the L2 Norm#

randomized_search = RandomizedSearchCV(
    estimator=MeanRisk(),
    cv=WalkForward(train_size=252, test_size=60),
    param_distributions={
        "l2_coef": loguniform(1e-3, 1e-1),
    },
)
randomized_search.fit(X_train)

best_model = randomized_search.best_estimator_

print(best_model.weights_)

Grid Search on Embedded Parameters#

model = MeanRisk(
    objective_function=ObjectiveFunction.MAXIMIZE_RATIO,
    risk_measure=RiskMeasure.VARIANCE,
    prior_estimator=EmpiricalPrior(mu_estimator=EWMu(half_life=40)),
)

print(model.get_params(deep=True))

gs = GridSearchCV(
    estimator=model,
    cv=KFold(n_splits=5, shuffle=False),
    n_jobs=-1,
    param_grid={
        "risk_measure": [
            RiskMeasure.VARIANCE,
            RiskMeasure.CVAR,
            RiskMeasure.CDAR,
        ],
        "prior_estimator__mu_estimator__half_life": [10, 20, 30, 40],
    },
)
gs.fit(X)

best_model = gs.best_estimator_

print(best_model.weights_)

Black & Litterman Model#

views = ["AAPL - BBY == 0.03 ", "CVX - KO == 0.04", "MSFT == 0.06 "]
model = MeanRisk(
    objective_function=ObjectiveFunction.MAXIMIZE_RATIO,
    prior_estimator=BlackLitterman(views=views),
)

Characteristics-Based Factor Model#

month = 21
quarter = 3 * month
half_year = 6 * month
year = 12 * month

characteristics = make_synthetic_characteristics(
    n_assets=500,
    n_observations=2000,
    random_state=0,
)

# Global factor
market_factor = GlobalFactor(family="market")

# Industry factors
industry_factors = OneHotCategoricalFactors(
    category="industry",
    family="industry",
)

# Style factors
beta_factor = FixedWeightedFactor(
    descriptors=[
        ("market_beta", EWMarketBeta(half_life=year)),
    ],
    transform_by_group="industry",
)

momentum_factor = FixedWeightedFactor(
    descriptors=[
        ("momentum", EWMomentum(half_life=half_year, skip=month)),
    ],
    transform_by_group="industry",
)

size_factor = FixedWeightedFactor(
    descriptors=[("log_market_cap", LogMarketCap())],
    transform_by_group="industry",
)

non_linear_size_factor = DerivedFactor(
    source="size",
    func=lambda x: x**3,
    transform_by_group="industry",
)

value_factor = FixedWeightedFactor(
    descriptors=[
        ("book_to_price", BookToPrice()),
        ("sales_to_price", SalesToPrice()),
        ("cash_flow_to_price", CashFlowToPrice()),
    ],
    weights=[0.8, 0.1, 0.1],
    transform_by_group="industry",
)

volatility_factor = FixedWeightedFactor(
    descriptors=[
        ("vol", EWVolatility(half_life=quarter)),
        (
            "residual_vol",
            EWResidualVolatility(
                half_life=quarter,
                beta_half_life=quarter,
            ),
        ),
    ],
    transform_by_group="industry",
)

# Characteristics factor model
model = CharacteristicsFactorModel(
    factors=[
        ("market", market_factor),
        ("industry", industry_factors),
        ("beta", beta_factor),
        ("momentum", momentum_factor),
        ("size", size_factor),
        ("non_linear_size", non_linear_size_factor),
        ("value", value_factor),
        ("volatility", volatility_factor),
    ],
    neutralize_against={
        "volatility": ["beta"],
        "non_linear_size": ["size"],
    },
    constrained_families=[("industry", None)],
    exposure_lag=1,
    inv_idio_variance_weight_shrinkage=0.5,
    n_jobs=-1,
)

model.fit(characteristics=characteristics)

factor_model = model.factor_model_
print(factor_model.summary())

For complete workflows, see the Factor Models user guide and the Factor Models tutorials.

Time-Series Factor Model#

factor_prices = load_factors_dataset()

X, factors = prices_to_returns(prices, factor_prices)
X_train, X_test, factors_train, factors_test = train_test_split(
    X, factors, test_size=0.33, shuffle=False
)

model = MeanRisk(prior_estimator=TimeSeriesFactorModel())
model.fit(X_train, factors=factors_train)

print(model.weights_)

portfolio = model.predict(X_test)

print(portfolio.calmar_ratio)
print(portfolio.summary())

Time-Series Factor Model & Covariance Detoning#

model = MeanRisk(
    prior_estimator=TimeSeriesFactorModel(
        factor_prior_estimator=EmpiricalPrior(covariance_estimator=DetoneCovariance())
    )
)

Black & Litterman Time-Series Factor Model#

factor_views = ["MTUM - QUAL == 0.03 ", "VLUE == 0.06"]
model = MeanRisk(
    objective_function=ObjectiveFunction.MAXIMIZE_RATIO,
    prior_estimator=TimeSeriesFactorModel(
        factor_prior_estimator=BlackLitterman(views=factor_views),
    ),
)

Pre-Selection Pipeline#

set_config(transform_output="pandas")
model = Pipeline(
    [
        ("pre_selection", SelectKExtremes(k=10, highest=True)),
        ("optimization", MeanRisk()),
    ]
)
model.fit(X_train)

portfolio = model.predict(X_test)

K-fold Cross-Validation#

model = MeanRisk()
mpp = cross_val_predict(model, X_test, cv=KFold(n_splits=5))
# mpp is the predicted MultiPeriodPortfolio object composed of 5 Portfolios (1 per testing fold)

mpp.plot_cumulative_returns()
print(mpp.summary())

Combinatorial Purged Cross-Validation#

model = MeanRisk()

cv = CombinatorialPurgedCV(n_folds=10, n_test_folds=2)

print(cv.summary(X_train))

population = cross_val_predict(model, X_train, cv=cv)

population.plot_distribution(
    measure_list=[RatioMeasure.SHARPE_RATIO, RatioMeasure.SORTINO_RATIO]
)
population.plot_cumulative_returns()
print(population.summary())

Minimum CVaR Optimization on Synthetic Returns#

vine = VineCopula(log_transform=True, n_jobs=-1)
prior = SyntheticData(distribution_estimator=vine, n_samples=2000)
model = MeanRisk(risk_measure=RiskMeasure.CVAR, prior_estimator=prior)
model.fit(X)
print(model.weights_)

Stress Test#

vine = VineCopula(log_transform=True, central_assets=["BAC"], n_jobs=-1)
vine.fit(X)
X_stressed = vine.sample(n_samples=10_000, conditioning = {"BAC": -0.2})
ptf_stressed = model.predict(X_stressed)

Minimum CVaR Optimization on Synthetic Factors#

vine = VineCopula(central_assets=["QUAL"], log_transform=True, n_jobs=-1)
factor_prior = SyntheticData(
    distribution_estimator=vine,
    n_samples=10_000,
    sample_args=dict(conditioning={"QUAL": -0.2}),
)
factor_model = TimeSeriesFactorModel(factor_prior_estimator=factor_prior)
model = MeanRisk(risk_measure=RiskMeasure.CVAR, prior_estimator=factor_model)
model.fit(X, factors=factors)
print(model.weights_)

Factor Stress Test#

factor_model.set_params(factor_prior_estimator__sample_args=dict(
    conditioning={"QUAL": -0.5}
))
factor_model.fit(X, factors=factors)
stressed_dist = factor_model.return_distribution_
stressed_ptf = model.predict(stressed_dist)

Entropy Pooling#

entropy_pooling = EntropyPooling(
    mean_views=[
        "JPM == -0.002",
        "PG >= LLY",
        "BAC >= prior(BAC) * 1.2",
    ],
    cvar_views=[
        "GE == 0.08",
    ],
)
entropy_pooling.fit(X)
print(entropy_pooling.relative_entropy_)
print(entropy_pooling.effective_number_of_scenarios_)
print(entropy_pooling.return_distribution_.sample_weight)

CVaR Hierarchical Risk Parity optimization on Entropy Pooling#

entropy_pooling = EntropyPooling(cvar_views=["GE == 0.08"])
model = HierarchicalRiskParity(
    risk_measure=RiskMeasure.CVAR,
    prior_estimator=entropy_pooling
)
model.fit(X)
print(model.weights_)

Stress Test with Entropy Pooling on Factor Synthetic Data#

# Regular Vine Copula and sampling of 100,000 synthetic factor returns
factor_synth = SyntheticData(
    n_samples=100_000,
    distribution_estimator=VineCopula(log_transform=True, n_jobs=-1, random_state=0)
)

# Entropy Pooling by imposing a CVaR-95% of 10% on the Quality factor
factor_entropy_pooling = EntropyPooling(
    prior_estimator=factor_synth,
    cvar_views=["QUAL == 0.10"],
)

factor_model = TimeSeriesFactorModel(factor_prior_estimator=factor_entropy_pooling)
factor_model.fit(X, factors=factors)

# We retrieve the stressed distribution:
stressed_dist = factor_model.return_distribution_

# We stress-test our portfolio:
stressed_ptf = model.predict(stressed_dist)

Opinion Pooling#

# We consider two expert opinions, each generated via Entropy Pooling with
# user-defined views.
# We assign probabilities of 40% to Expert 1, 50% to Expert 2, and by default
# the remaining 10% is allocated to the prior distribution:
opinion_1 = EntropyPooling(cvar_views=["AMD == 0.10"])
opinion_2 = EntropyPooling(
    mean_views=["AMD >= BAC", "JPM <= prior(JPM) * 0.8"],
    cvar_views=["GE == 0.12"],
)

opinion_pooling = OpinionPooling(
    estimators=[("opinion_1", opinion_1), ("opinion_2", opinion_2)],
    opinion_probabilities=[0.4, 0.5],
)

opinion_pooling.fit(X)

Docker#

You can also spin up a reproducible JupyterLab environment using Docker:

Build the image:

docker build -t skfolio-jupyterlab .

Run the container:

docker run -p 8888:8888 -v <path-to-your-folder-containing-data>:/app/data -it skfolio-jupyterlab

Browse:

Open localhost:8888/lab and start using skfolio

Recognition#

We would like to thank all contributors to our direct dependencies, such as scikit-learn and cvxpy, as well as the contributors of the following resources:

Citation#

If you use skfolio in a scientific publication, we would appreciate citations:

The library:

@software{skfolio,
  title     = {skfolio},
  author    = {Delatte, Hugo and Nicolini, Carlo and Manzi, Matteo},
  version   = {1.0.0},
  year      = {2026},
  doi       = {10.5281/zenodo.16148630},
  url       = {https://doi.org/10.5281/zenodo.16148630}
}

The above uses the concept DOI, which always resolves to the latest release. If you need precise reproducibility, especially for journals or conferences that require it, you can cite the version-specific DOI for the exact release you used. To find it, go to our Zenodo project page, locate the release you wish to reference (e.g. “v1.0.0”), and copy the DOI listed next to that version.

The paper:

@article{nicolini2025skfolio,
  title         = {skfolio: Portfolio Optimization in Python},
  author        = {Nicolini, Carlo and Manzi, Matteo and Delatte, Hugo},
  journal       = {arXiv preprint arXiv:2507.04176},
  year          = {2025},
  eprint        = {2507.04176},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2507.04176}
}