We are working with 3 years of 15minute time-interval media data (1million+ entries) and have 14 external regressors (daypart, weekdays, holiday, genres). Objective is to forecast for next 15-minute intervals for given values of external regressors. What we did: - Used auto.arima (with xreg) from forecast package on this data. Result attached enter image description here

  • Created a sub-set of 5% of xreg data to check forecasting accuracy. When we compare actual media GRP values with forecasted values, there is high deviation of ~100% for certain data points.

    1. Is auto.arima the right approach for such large data?
    2. How can we improve accuracy of this model?

1 Answer 1


It might be better to model your using a or a model, rather than with dummies. Then again, this might run into performance issues, given your large dataset.

You may be interested in How to know that your machine learning problem is hopeless?

Other than that, I think we simply don't have enough information to be overly helpful.

  • $\begingroup$ Thanks Stephan for replying. However, we cannot remove these dummies, since the objective is to predict for given values of daypart, day, genre combinations. Eg: objective is to be able to predict TRP for a Comedy program (genre) aired on Tuesday (weekday) at Night (daypart).. Need your help how can I improve forecast here? also should I use Neural Networks? $\endgroup$ Commented Mar 21, 2018 at 14:16
  • $\begingroup$ A BATS or TBATS model will allow you to forecast for a given future time point, depending on the day of week and time of day that is. I suggest you take a look at that. Of course you can also use NNs. $\endgroup$ Commented Mar 21, 2018 at 14:22

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.