x******a
发帖数: 6336 | 1 A question regarding using PCA to capture the pairwise covariance matrix.
Suppose I have 10 times series and 250 data point for each time series in
the format of a matrix. Let's call it A of shape 10*250,Let us call the
covariance matrix COV and it is a 10*10 nonnegative defined matrix.
I would like to capture this matrix COV with a one-factor model,
0.Is PCA the right direction on this?
1.Assuming 0. is right. We find the largest eigenvalue lamdbda_M and the
correponding vectors v_M of COV. Then is it true that one fact is modeled as
Y= sqrt{lambda_M}* + epsilon? I tried this in two dimension case,
the result is far from correct.
2. Suppose we have some missing values for the data A, when I calculated the
pairwise covariance, I used the available data. what is a good algorithm to
capture the covariance in this case(it could be non-positive defined in
this case)? I saw the imputation algorithm, EM algorithm, which one is
better?
Thanks a lot! | c********h
发帖数: 330 | 2 你的Anormalize过吗，印象中要想用你1里的公式，要先怎么弄下，至少是subtract
mean吧 | k***n
发帖数: 997 | 3 do stock returns often have 0 averages so no need to center them before
applying pca?
what's Y in the model? | x******a
发帖数: 6336 | 4 mean was substracted. thanks for pointing out. |
|
