1
$\begingroup$

I have daily time series data of almost two years starting for Jan 2018 to Nov 2019 and need to forecast for next two months Dec and Jan. My train data(Jan 2018-Aug 2019)is up to Aug 2019 and its error(MAPE) on test data(Sept to Nov) is 12%. I am using arimax technique.

Should i use these model to forecast for Dec-Jan which seems to too far ahead considering daily data or shall i build model on whole data again and then forecast for Dec -Jan?

My concern is if i am using first model to forecast then it is going too much into future and hence less confidence about the prediction. On the other hand if i use 2nd model then it's capturing recent changes and hopefully will give better forecast. My 2nd model is having all those regressors of 1st model whose p valus is less than 0.05.

Please suggest

Improve this question
$\endgroup$
2
  • $\begingroup$ What's your reason for not using the more recent data? Are you worried that including the more recent data between Sept and Nov 2019 will decrease the performance? If your time series is stationary, in the sense that its statistical properties do not change over time, then you should use the most recent data to make forecasts. If you change your model (i.e. only selecting only some of the regressors), I would re-evaluate my model to estimate its performance. But if you're happy with the performance estimate of the first model, why not just use that? $\endgroup$
    – mloning
    Commented Dec 3, 2019 at 12:01
  • $\begingroup$ @mloning With 1st model i know model has not over fitted. Train error is 10 and test error is 12 but as i use it to forecast for Dec 1st two days error has gone up to 25% and 28%. whereas with 2nd model i am not sure whether it will over fit or not in long run but has given error of 9 and 7% respectively. whats the best practise? $\endgroup$
    – joy_1379
    Commented Dec 3, 2019 at 12:59

1 Answer 1

Reset to default
0
$\begingroup$

with 23 values , I would simply form a useful model using all 23 ... then I would use that model from period 18 ..estimating parameters based upon the 18 historical values and predict the 19th and the 20th .. similarly i would use the useful model and use 19 values to estimate parameters and predict the 20th and the 21st ...

everything being reasonable I would then estimate parameters for the useful model using all 21 historical values and predict the 24th and 25th periods.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.