1

Anyone have any links or resources on pros/cons of building a timeseries model with overlapping data points? Generally, a lot of text build models on daily returns, but let's say the daily variable is just too noisy and I'd prefer to smooth it out a bit by doing a rolling 7 day or 30 day value. After all, I'd also prefer to predict the next 30 day value as opposed to the next day.

What are the pitfalls of using a daily 30day rolling value as opposed to 30day rolling values spaced out by 30 days? Another consideration is I really don't have that many data points, maybe 1 year worth of good data and 1 year worth of questionable data (so 2 max).

I know the former will have much smaller standard deviation because you are using overlapping data points, but at the same time, if I'm trying to predict what the 30day value of something will be, I feel like that's more realistic - or am I wrong?

My goal is to get a mean value and I like the AR model in that the next 30 days probably is best predicted by the current 30 days.

None of the links here work: Time series regression with overlapping data

Thanks!

Richard Hardy
  • 67,272
confused
  • 3,253
  • As I noted in a recent comment to your previous questions, I have taken care to fix the links. They do work. – Richard Hardy Apr 09 '22 at 07:29
  • 1
    None of these models attempt to represent, or even care about, "overlapping data." You are free to construct a new time series from the original data in any way you like, including a rolling sum or mean (for instance). Just be aware that this tends to create strong and long-range serial correlations in the new time series. And don't expect any magic to occur: no amount of such processing will create any genuinely new information, so if you begin with small amounts of data, you will end up with correspondingly large uncertainty in your inferences and forecasts. – whuber Apr 09 '22 at 15:13

0 Answers0