Covariance Estimator

A covariance estimator estimates the covariance matrix of the assets.

It follows the same API as scikit-learn’s estimator: the fit method takes X as the assets returns and stores the covariance in its covariance_ attribute.

X can be any array-like structure (numpy array, pandas DataFrame, etc.)

covariance_ is expressed in the periodicity of X (daily returns give a daily covariance) and is consumed as is by the optimizers, without annualization (see Periodicity Convention).

Available estimators are:

• EmpiricalCovariance

• EWCovariance

• RegimeAdjustedEWCovariance

• GerberCovariance

• DenoiseCovariance

• DetoneCovariance

• LedoitWolf

• OAS

• ShrunkCovariance

• GraphicalLassoCV

• ImpliedCovariance

For online learning and streaming workflows, EWCovariance and RegimeAdjustedEWCovariance support incremental updates with partial_fit. They also support NaN-aware updates with active_mask, which helps distinguish assets that belong to the universe but have missing returns, (e.g. holidays) or assets outside the universe (e.g. during pre-listing or post-delisting periods). See Online Learning for the full online workflow, including covariance forecast evaluation and online hyper-parameter tuning. See Missing Data and Changing Universes for the full convention on NaNs, universe membership, estimator warmup and investability.

Example:

from skfolio.datasets import load_sp500_dataset
from skfolio.moments import EmpiricalCovariance
from skfolio.preprocessing import prices_to_returns

prices = load_sp500_dataset()
X = prices_to_returns(prices)

model = EmpiricalCovariance()
model.fit(X)
print(model.covariance_)