Cross Validation/Bootstrap on Time Series Segement

#1
I have a small time series segment (~100 to ~500 points) and would like to fit a function to it. I'm not interested in the out of sample performance, but I want to make sure that I avoid overfitting.
I read online that cross validation and bootstrap don't work on time series since the data are not independent. Is this correct? How can I calculate a prediction error?
 

hlsmith

Not a robit
#2
I don't have a direct answer for this question, but while waiting for a reply you should review the prior threads. I believe within the past month or so, there was a similar post.
 

Link

Ninja say what!?!
#4
It's true that in time series data, there is normally correlation. However, this doesn't rule out boostrapping or CV. Some thought should be given into perhaps modelling the change over time. If your model accounts for this, it may still be fine undergoing bootstrapping or CV.

Regarding your question "How can I calculate a prediction error?", you SHOULD be interested in out of sample performance. By the very definition of prediction error, you're looking at how the model you've fit performs predicting on new data. CV usually gives you a good estimate of this.

Hope that helps.