My OLS model build on full data has heteroscedasticity of residuals. Is it allowed to use a subset of the time series to get around this issue?

What are the possible implications?

  • 1
    $\begingroup$ If you have time series data you should not be using OLS. $\endgroup$ Oct 14, 2019 at 11:16
  • $\begingroup$ How would selecting a subset resolve any heteroscedasticity problems? As indicated by @user2974951, a more important issue concerns the potentially strong serial correlation in the data. $\endgroup$
    – whuber
    Oct 14, 2019 at 12:59
  • 1
    $\begingroup$ @whuber When I run on the 1st 32 data points (there are 40 in total), I am not facing the issue. $\endgroup$
    – Dom Jo
    Oct 14, 2019 at 13:25
  • 1
    $\begingroup$ That sounds a little like cutting off your arm to cure a sore thumb: you already have evidence that the model built from the first 32 points likely isn't appropriate for the last 8 and, by removing the last 8, you make it almost impossible to forecast anything accurately. You might find it more constructive to describe your data, the results you're getting, and your objectives, and then ask what an appropriate form of analysis might be, rather than asking whether subsetting is going to work. $\endgroup$
    – whuber
    Oct 14, 2019 at 13:41


Your Answer

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