I am conducting common factor analysis for several time series. The range of observation is 1995M07 to 2010M06. I would like to do a rolling window analysis with n = 120 observations:
1995M07 2005M06
1995M08 2005M07
...
The technique used for factor extraction is:
- method: principal factors
number of factors: parallel analysis (mean)
random generator: Knuth
inital communalities: squared multiple correlation
method: VARIMAX
starting values: unrotated
The resulting factor loadings are within the degree of communlaties that I expected. However, I was quite surprised to find significant differences between rotated factor loadings from different rolling windows (e.g. between 1995M07 2005M06 and 1995M08 2005M07), despite the fact that only one in 120 obervations was different. Most time series would load on the same factor. For a few time series, however, the rotated factor loadings could differ significantly. In rare instances the number of intial extracted factors differed as well (although never by more than one).
It appears as if those time series with large differences in common factor loadings between different rolling windows would display small or insignificant loadings alltogether (even after rotation). As a result, they appear somewhat 'indifferent' as to their 'preferred loading' when maximizing on a particular factor in Varimax. As a result, even small changes in time series data may result in large changes in factor loadings.
Here comes my question: Is there a setting (e.g. as part of the initial assumptions) or estimation method that would mitigate the impact from time series with insignificant factor loadings. Put differently, I would like to use the results from the initial factor estimate (e.g. for 1995M07 2005M06) as a basis for the next observation period, so that the results change only gradually over time. I appreciate any input on this. Thanks in advance