0
$\begingroup$

Background: I'm working with log returns for about 400 tech stocks. I want to use PCA to reduce these into principal components (Internet companies, software developers, circuit board manufacturers, etc), which then are to be used as sector-related indices. Parallel analysis is my method of determining how many factors to extract in the first place. Rotation is done via Oblimin with Kaiser Normalization in SPSS.

Questions: How do I determine which variables need to be trimmed? Those with low communality? Or should I look at the pattern matrix and start cutting variables which happen to load on multiple factors? After that, can I use the figures in the pattern matrix to weight each stock's contribution to each component's variance? For example, if the rotation yields .71 for Stock A and .10 for Stock B, I could multiply the return for each stock in each period, T, by its respective weight.

Improve this question
$\endgroup$
3
  • $\begingroup$ Why communality bothers you? Is that because your aim is to do factor analysis and not PCA? $\endgroup$
    – ttnphns
    Commented Mar 17, 2015 at 20:00
  • $\begingroup$ I think that factor analysis, in particular CFA, is a better approach in this case than PCA. It seems to me that the OP wants to reduce dimensionality by considering latent variables. $\endgroup$
    – Aleksandr Blekh
    Commented Mar 17, 2015 at 20:09
  • $\begingroup$ Why do you want to "trim" any variables at all? $\endgroup$
    – amoeba
    Commented Mar 17, 2015 at 22:56

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.