Login
Login

Question about ARIMA forecast

Home 21090308 Forums Numerical Method Statistics Question about ARIMA forecast

  • This topic is empty.
Viewing 4 posts - 1 through 4 (of 4 total)
  • Author
    Posts
  • #1949
    murari
    Member

    Hi,
    I am a newer of SuanShuers and I want to use ARIMA model in SuanShu to forecast next value of time series just like last topic “ARIMA forecast”. After reading the reply of last topic, I have a try using my own data and the codes are listed below. However, some new questions appear.

    My Data:

    Source Code:

    and the output is:

    My questions are:
    1. the forecast accuracy is too low, prediction value is 4.246 but real value is 2.78
    2. how to determine p,d,q automatically?
    3. why error != |real-prediction|?

    Thanks for your kindly attention and Waitng for your help.

    #2258
    valondon
    Member

    Dear wwssttt,

    1 The forecast error is related to sample size (in your case 18, which is very small), and the true underlying model (is the model you used to generate the data an ARIMA model?) It is helpful that you can provide the true underlying model and its parameters, and also the estimated parameters. If the parameter are not estimated accurately, the forecast will naturally be far way from real.

    2 This is an open question, there is no standard answer. One way to determine the order is to check the PACF and ACF of the series. Here is a good explanation: http://people.duke.edu/~rnau/411arim2.htm.

    3 Please refer to the javadoc of the meaning of forecast.var. The var term here is not |real – forecast|. You should know this from your program. You did not pass the real value (2.78) to the model “arima” or the forecaster “forecast”, therefore |real – forecasst| cannot be computed.

    Best wishes

    #2259
    murari
    Member

    Thanks for your detailed reply

    #2260
    murari
    Member

    BTW, my data is from applications in real life not generated by some models

Viewing 4 posts - 1 through 4 (of 4 total)
  • You must be logged in to reply to this topic.