precise 0.1.0

Online covariance and precision estimation

precise

A collection of autonomous online (incremental) covariance matrix estimators.

Install

pip install precise 

Examples

See /examples_basic_usage

Explanation

You use state updaters, then functions of the state.

State updates

All updaters take a prior state and return a posterior state (dict). An empty dict is passed to initialize.

Example:

from precise.synthetic.generate import create_correlated_dataset
from precise.covariance.empirical import emp_pcov

if __name__=='__main__':
    xs = create_correlated_dataset(n=500)
    s = {}
    for x in xs:
        s = emp_pcov(s=s, x=x)
    pprint(s['scov'])

Naming hints:

Shorthand	Intent
scov	Sample covariance
pcov	Population covariance
spre	Sample precision
ppre	Population precision

Method hints:

Shorthand	Meaning
emp	Empirical
ema	Exponential weighted moving average
lz	Le-Zhong variable-by-variable updating
lw	(inspired by) Ledoit-Wolf

Some, such as OAS, don't need their own state tracking. See examples.

State functions & mutations

Three types of utilities exist

1. The covariance/statefunctions are illustrated by the example running_oas_covariance.
2. State covariatnce/statemutations do things like ensuring both covariance and precision matrices exist in the state. Or for instance: s = both_cov(s) ensures both sample and population covariances are present.
3. Miscellaneous /covariance/matrixfunctions functions act directly on matrices.

Updater hyper-parameters

The intent is that methods are parameter free. However some not-quite autonomous methods admit just one additional scalar parameter r and that can make the creation of fully autonomous methods simpler, akin to the tuning of skaters explained here in the timemachines package.

Miscellaneous

• Care to cite, some ideas here
• If you just want univariate, and don't want numpy as a dependency, there is momentum.