Thanks Ajith kumara. Need to find a way to do this enhancsement using eviews.

Hi,

It seems that your model is not yet the best model. I guess that you've taken monthly data for your analysis. Sometimes there may be seasonality in the data although the series is stationary. Under seasonality ARIMA model alone does not perform well. In a simple way, you can include monthly dummy variables to see if the seasonality is there. You need to include these dummy variables in the automatic ARIMA selection in Eviews. If the seasonality is deterministic this method work well, hopefully you will have white noise. If it doesn't work try SARIMA method (include seasonal AR and seasonal MA terms in your model. If you are using this method you need to include seasonal AR and seasonal MA terms in automatic ARIMA selection. If seasonality is an issue in your data, one of these methods will help you to get a better model with white-noise error term.

Kind regards

Thushara

Thanks a lot Thushara. Hope this will help to sort out my problem. Since there is no unit root and no seasonality according to the graphs I didn't consider seasonal factors to forcast the model.i will recapture model using SARMA.

I considered seasonal factors also then my ARIMA model changed as ARMA(4,4)(2,1). Still q statistic give p value less than 0.05 which means there is correlation with residuals. I used eviews automatic ARIMA fearture and according to the system it allow to select max SAR mam max SMA only up to 2. Please advise to clear this matter.

Hi Thakshila,

Did you use a stationary variable in your model? if the series is not stationary it's hard to remove AC or PAC.

Kind reagrds

Thushara

yes Thushara. I have done unit root testing and there was no unit root for 1st order level. Therefore I considered 1st order difference based values for ARMA model. Really unable to find the reason why there is auto correlation for residuals. Therefore can I check that only for short term predictions ? I am having limited time to complete my research work.

Hi: it's kind of odd that you would need obtain an arma(4,4). generally speaking, if you use an arma model you shouldn't need orders larger than 2 ( maybe 3 but it's rare ) for both the AR piece and MA piece. also, "automatic" selection procedures can be problematic because one doesn't know the details of the criteria being used for model selection. I would use a brute force approach by running all 9 models manually using orders 1 to 3 for both the AR and MA part. Then study the results along with residual autocorrelation. Finally, as other people said, check for seasonality. That may be the reason why you are getting an arma(4,4). Oh: one other thing: using root testing can be problematic. it's best to plot the response and just look to see if it seems stationary. it may need differencing even thought the unit root test says it doesn't. the unit root testing world is not a perfect one.

Ajit: The context at the link is very different from the context that the OP is dealing with. The OP is already imposing "structure" by using an ARIMA model. The link is refereeing to the case where one is fitting a linear model which is extremely different.

Thank you Ajith Kumara. I can present the model based on the mentioned AIC, AICC and BIC value regardless of the residual correlation for this phase. But for the learning points have to consider that factor too for have a better model. Thank you for the information.

The model you have selected is ARMA(4,4). I suspect that you've restricted maximum AR and MA to 4 in your automatic ARIMA selection in Eviews. If it is so, try to increase the maximum number of AR and MA terms, that may improve model. If it selects higher number of AR and MA terms, make sure that the model is invertible.

Kind regards

Thushara

Thank you Thushara. Will try on that.

Thank you Mark Leeds for your descriptive answer with your years of experience. I am not having that much time to reconsider about manual ARIMA process only considering AR, MA only max 2 values because I have applied automatic ARIMA model for more than 6 factors. I will consider about your factors as future enhancement.

Hi: Youre welcome and I hope that something I said helps. I don't know eviews but it's pretty easy to estimate ARIMA models in R without using automatic procedures. R has a steep learning curve but, once you get used to it, it's very powerful. All the best and good luck with project.