I am working on a paper part of which is the application of validation rules to find how many principal components give us the least generalization error.

The concept goes more or less like this: "Given that the dimension of the model is reduced, we reset our window size to 60 days to avoid overfitting problem. After running multivariate linear regression within the first 20 components using 60 days training set, we find that the first 12 components give the smallest generalization error on the 30 testing days. Finally, we compute the in-sample and out-of-sample residuals."

Note that, the dimension of the full principal components matrix is 483 (days) X 482 (values).

The implementation is trivial indeed, yet I am strangling on how to "choose" the correct sub-matrices for any of these cases.

All suggestions (including matlab code) are welcome.

EDIT: Using the hint of cross validation as Arnold Neumaier mentioned below, and provided that I have already implemented the function that splits the initial dataset into parts, does the following solution solve partially the problem? What should I do next?

for i=1:1:10
  training_set = ex1_data_txt(find(split_assignments(:,i)==0),:);
  test_set = ex1_data_txt(find(split_assignments(:,i)==1),:);

  % determine weights from the training set
  phi_train=[training_set(:,1).^(0) training_set(:,1).^(1)];
  phi_test=[test_set(:,1).^(0) test_set(:,1).^(1)];
  % apply learned weights to the test set and compute MSE

1 Answer 1


By cross validation. Divide your data set into 5 random batches of about equal size. Apply the estimation procedure 5 times with always one batch left out, and work out for each choice of effective dimension the generalization error by testing it on the remainder. Average these generalization errors .(If you want to be cautios, consider tin place of the averages the averages plus $k$ times the standard deviations, for some $k$ that you feel addresses your level of risk aversion.)

This gives you a curve from which you can read off the answer.

If the curve has a pronounced minimum, it gives you the effective dimension.

If, instead, the curve has an essentially flat bottom region, it gives you an indifference interval of effective dimensions, and by Ockahm's razor, you'd choose the smallest value in the indifference interval as the effective dimension.

Afterwards, you can do with this effective dimension the PCA on the full data set, and can expect a similar generalization errro on unseen data. But of course only when the original data were representative of the unseed data.

Window sizes, etc. can be chosen along similar lines.

  • $\begingroup$ Just updated the question above. Could you please give me some further and practical hints? $\endgroup$
    – eualin
    Jul 27, 2012 at 17:18
  • $\begingroup$ My description in words describes well enough what needs to be done, packaged in simple steps. But I don't give programming help. $\endgroup$ Jul 27, 2012 at 17:32
  • $\begingroup$ hmmm, then let me ask you in a theoretical context: Applying a 5-cross-validation results 5 MSEs, right? How is that connected with the optimum selection of principal components? It stills not quite clear to me. $\endgroup$
    – eualin
    Jul 27, 2012 at 17:39
  • $\begingroup$ It gives a sample of 5 MSE's which can be used to compute an estimate for the mean MSE for generalization and its standard deviation. This is the basis of the plot against the effective dimension, which allows you to find that effective dimension for which this estimate (or a robust variant) is smallest. $\endgroup$ Jul 27, 2012 at 18:15

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.