I am trying to understand how the ICA by A. Hyvärinen, J. Karhunen, E. Oja (2001) works in practice. In particular, I have problems trying to understand its application to factor-analysis-kind of models. In their book (2001) and in another paper, the authors refer to the fact that they apply ICA to the cash flows of 40 stores/shops of which they have 140 time-observations each ("Kiviluoto, Oja, 1998, Independent component analysis for parallel financial time series"). They say they reduce first the dimension of teh data (the 40 time-series) to only 4 using PCA through which they also whiten the data. Then, they apply their FastICA algorithm to extract the sources from those new 4 time-series. The idea is to extract the underlying sources common to all the 40 shops and investigate them.

I am trying something similar using the FastICA provided in scikit-learn in Python. Using the fit_transform method, I should get back the "fitted" values for the original time series, but actually the ouput is 4 time series only. Now my question is: can I actually apply the ICA method by choosing a number of sources (much) smaller than the signals we observe in the data (in this case, the cash flows of shops) and then, once I have my mixing matrix, multiply it with the matrix of sources to get back fitted values (or predictions or transformed values) for my initial input matrix?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.