How to further improve ARIMA model



Hi All,
I am new to time series forecasting and ARIMA.
I am fitting ARIMA model and used auto.arima() in R.
I got following residual plot and I can see lot of significant autocorrelation and partial-autocorrelation.

I need help in interpretation of residual ACF and PACF graphs and how can I further improve the mode? Following is the residual display.


Check this. Start from the top before trying Auto Arima


What model are you using?
This is a handy summary of rules for identifying ARIMA models. If the season is 24 and the plots are negative check Rule 13


Hi Aless,
I am using ARIMA(5,1,0)(2,0,0)[24] model. Time series is hourly data.
I have chosen frequency of time series as 24.

Also could you please provide summary of the rules for identifying ARIMA which you mentioned.



Ops, sorry. I was referring to this link: guidelines arima.

In any case i suggest going back to your initial data and apply transformations (in your case seasonal differentiation, maybe 1st differentiation if you have a trend, log if variance changes, etc). Read a couple of tutorials. I am working myself at a kernel in kaggle called sarimax-on-mean-visits, which could be a good primer on ARIMA models in python. However, this is better :slight_smile: You can certainly find other tutorials for R


Looking at the plot, here are my suggestion.

Check if your time series is stationary.
Decompose your time series and have a look at its components.
next check if you apply some transformation such log transformation helps in reducing the variance.

Post that apply the auto arima function.