I've seen a couple of questions asked here about seasonality "left over" (so to speak) after differencing like this and this, but unfortunately those answers don't help my situation.

I have ~ 17000 hourly observations which seem to exhibit daily (frequency=24) and weekly (frequency=168) seasonalities, as well as general nonstationary unit root (d=1). But even after removing these three components I'm still seeing troubling autocorrelation at lag 24 and 168. I'm not sure if even may have overdifferenced. Below are the ACF plots:

Original Data ACF

enter image description here

First-Order Differencing

enter image description here

After seasonal difference of 168

enter image description here

After seasonal difference of 24

enter image description here

The data is taken from Kaggle in the hour.csv file, using the first 8645 observations to model the cnt variable.

I'm at a loss at how to resolve this. I'm trying to make 8 hour ahead forecasts. Can anyone please help?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.