How do I get an accurate prediction interval after scaling data using Np.log

Thanks for sharing a pretty good model.

I have a batch of values that are relatively large in holidays but relatively small in ordinary times. After scaling y value by Np. log, I predict it, and then restoring data by Np. exp, the confidence interval obtained is very small. Although RMSE value is relatively low, the probability of the real value in the prediction interval is relatively low. Conversely, if I scale y without np.log, the confidence interval is very large, even though the real values are in the predicted interval.

Looking forward to your reply. many thanks.