search

mateuszbaran / CovarianceEstimation.jl

Lightweight robust covariance estimation in Julia
MIT License
42 stars 7 forks source link

Supporting passing in weights #10

Open rofinn opened 5 years ago

rofinn commented 5 years ago

It might be nice to support a lower level fit and transform API on top of just cov. This would allow folks to pass in an existing covariance matrix that they'd like to apply an estimation method to (e.g., a weighted covariance).

mateuszbaran commented 5 years ago

I don't quite understand how would that work. Implemented formulas for covariance shrinkage assume normal covariance and a specific shrinkage target. Wouldn't using weighted covariance lead to wrong shrinkage coefficients?

rofinn commented 5 years ago

Hmmm, good point. I guess the R corpcor just lets you pass in the weights with a data matrix rather than precomputed values.

mateuszbaran commented 5 years ago

StatsBase.jl has its own approach to weights and it looks nicer. I think it would be good to have a similar design here.

nickrobinson251 commented 5 years ago

Supporting weights, following the StatsBase API of cov(ce::CovarianceEstimator, X::AbstractMatrix, w::AbstractWeights), would be really useful :)

mateuszbaran commented 5 years ago

Yes, it would be very useful, but I still don't know if it's even possible to have generic implementation for all types of weights. Frequency weights should be relatively easy to support though. What kind of weights do you need?

nickrobinson251 commented 5 years ago

I believe AnalyticWeights. But also I am not sure weighted Lediot-Wolf is possible (I was hoping it was)

mateuszbaran commented 5 years ago

Unfortunately, I don't know how to modify Ledoit-Wolf to make it work with AnalyticWeights. If you find a paper about it, I will definitely like to see it.

nickrobinson251 commented 5 years ago

My colleague tells me it is not possible (or at least the proof in the LW paper breaks down when weights are added) ...but i've not done the working myself nor seen a paper on it.

rofinn commented 5 years ago

NOTE: a potential work around could be to collect a weighted sampling first and then estimate the covariance from that.