estimate-online-covariance

Estimate online covariance with precise

precise provides sklearn-style estimators with a single partial_fit contract. Pure numpy.

pip install precise

The pattern

import numpy as np
from precise import EwaCovariance        # exponentially weighted; recency-biased

est = EwaCovariance(r=0.05)              # r in (0,1]; larger = faster forgetting
for y in stream:                         # y is a 1-D array (one observation)
    est.partial_fit(y)

est.covariance_     # (d, d) ndarray, symmetric and PSD by construction
est.correlation_    # unit-diagonal correlation
est.precision_      # inverse covariance (when well-conditioned)
est.location_       # running mean
est.n_samples_      # observations seen

fit(X) is the batch drop-in (X is 2-D, rows = observations); it resets then replays rows, so it matches sklearn.covariance's call shape.

Choosing the class

all_estimators() lists every estimator; estimator_from_name("LedoitWolfCovariance") looks one up. Sensible defaults by situation:

• general / recency-weighted: EwaCovariance(r=...)
• many variables relative to samples (p/n large) or ill-conditioned: LedoitWolfCovariance, OASCovariance, ShrunkCovariance, FactorCovariance
• heavy tails / outliers: HuberCovariance, TylerCovariance
• regime changes: AdaptiveEwaCovariance, DCCCovariance
• you don't know: use the choose-covariance-estimator skill (suggest(X)).

Notes

• All estimators are truly online — constant work per observation, no growing buffers.
• State is a JSON-able dict: est.get_state() / est.set_state(s) for mid-stream checkpointing.
• covariance_ is always symmetric PSD; don't hand-symmetrize or clip it yourself.
• Named series with a changing universe? Use the keyed-dynamic-universe skill instead.