search

kamilest / conformal-rnn

Implementation for Stankevičiūtė et al. "Conformal time-series forecasting", NeurIPS 2021.
65 stars 13 forks source link

Question on unintuitive results for Bonferroni correction when alpha = 0.5 #5

Open bethrice44 opened 1 year ago

bethrice44 commented 1 year ago

I have been involved in implementing a conformal interval wrapper based on this method in sktime.

I am confused about how the Bonferroni correction works and wondered if you could help. There are more details in this issue and this PR but essentially I think the implementation is as in the paper but the results when alpha=0.5 are unintuitive (the values get much larger and it throws the forecasts off).

We don't use the factor of (n_calibration + 1.0)/n_calibration as you do in this code but that is close to 1 so I don't think it would make the difference. Do you have any pointers to what could be going wrong here?

Thanks