Hirenkumar,

your question assumes that the problem is the amount of training data. There are many factors that could be influencing the performance of your model e.g. after taking the 1st difference the data may not be sufficiently stationary, you might have inadequate model orders, you might not be "integrating" the data after model prediction. If you believe that the problem is the amount of training data, then perhaps you could test the residuals for correlation. If they are uncorrelated (white) then you should have sufficient data. If the residuals are not white, then there are many potential pitfalls.

Luis added some really important points. Consider them and then maybe think about some of the following.

Understanding the data is critical to ARIMA models.  Changing the autoregressive AR or moving average (MA) elements can make a significance.  

It would help us if you go into greater detail.  For example, it is confusing when you say you are "...recording application workload every 5 minutes after an hour or two..."  It sounds lie sometimes you are skipping an observation (I.E., after an hour or two..."). An ARIMA model is sensitive to the sequence of the data.  Skipping an observation can cause serious problems.  Think about the implications for the MA and AR elements. 

In addition, why is the 5 minutes after the hour important?  Does the data fluctuate every minute? Also, how far into the future are you trying to forecast?  Are you forecasting for 5 minutes after each hour?  Sometimes ARIMA models are able to capture turning points when other models are not capable of doing so.  

I hope this helps.

Rob

hi,

thanks,

here my interval is 5 min so its 12 observations per 1 hour ok. so 120 for 10 hours so at now my criteria is as below.

1 - 120 use as training data

121 - 240  here i start predicting at every 5 min 

241- 360 same as 121 -240 predicting every 5 min 

so here one question is for 241 - 360 should i consider history from 1 -240 or 121 - 240 ?