I am trying to forecast an index option's implied volatility using a sliding window regression and I'm a little confused on how I can go about cross validating with respect to the training and test set size. What I have so far is I cross validate the training window size by fixing the test set to be 1. I then do my sliding regression across time as such: enter image description here

Now this generates my predictions over all the points after my 1st training window and I calculate the $R^{2}$ across this set of predictions and get an optimal training window size from this. I then fix my training window to be this size and do the same cross validation procedure but this time varying my test set size, from 1 to 100s.

Intuitively, it makes sense that my $R^{2}$ is going to generally be monotonically decreasing as I expand the test set size since I'm not incorporating new, relevant info quick enough. But I am wondering what are the cons of choosing a test set size of just 1?

I'm doing OLS so computational intensity of retraining the model at every step is negligible and I don't see how this would be overfitting the model.


1 Answer 1


Maybe I'm not understanding correctly but your test set size is all your data points (minus the initial training set) right? The size 1 is the forecasting horizon, how many data points forward you forecast for a given training dataset. There's no issue setting that to 1 if it makes sense for your application.

  • $\begingroup$ Yes, this is correct. I was just wondering if there's any direct issue with my forecasting horizon being 1 apart from cost of retraining at every step, but I think it makes sense for my purposes $\endgroup$ Oct 13, 2023 at 16:48

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.