I am conducting some time series forecasts using quite limited data, 13 years annually. Basically, I am trying to forecast companies emission totals using historical values. The historical data however, seems to be nothing but white noise, e.g no significant lags on ACF, however Box Ljung <0.05.

My question is if it would be helpful to include correlated variables in a ARIMA model, with xreg..? For example including GDP, World Emissions, and Energy Prices. Would this improve the forecasts at all, or would any improvement on the test accuracy only be coincidences (limited data).

  • 1
    $\begingroup$ 1. if it is white noise, adding predictors won't help. 2. if it only seems like once noise, you will have to clarify what you mean by this. This question will be hard to answer without looking at data, so precision in your description will be helpful. $\endgroup$ – Taylor Jun 26 '19 at 23:04

13 data points is extremely little. There is a very rough rule of thumb that you need 20 data points per parameter to be estimated, so you just have enough for an intercept term. Also, you really don't have enough data to do holdout testing - if you hold out, say, the last three data points, then you only have ten points to fit your model to, and how much would you want to rely on a holdout set of just three points?

By all means, try including regressors you believe might drive your time series. (Remember that you need to forecast future values of your regressors, too - this adds additional uncertainty!) Per above, I would not expect this to be very helpful.

Best method for short time-series may be useful.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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