I have two years of data - I want to forecast how the 2nd year of traffic will behave based off the previous year of data.

The x axis is time, where left -> right is moving forward in calendar year

Dataset 1:

enter image description here

Dataset 2:

enter image description here

What is the best model for forecasting here? Would it be a simple polynomial regression fitted with 2017 data, and then applied to 2018? Or is a time-series approach better here?

I want to choose an approach that best preserves the weekly cyclic nature of traffic, while capturing the overall shape of the curve.


You don’t need any forecasting model. 2018 data closely mirrors 2017 data , so the best forecasting model for 2018 is 2017. 2 years worth of history is a very short series for any time series models. Try searching Naive method for forecasting they are extremely useful for forecasting short series.

  • $\begingroup$ Do you know of any naive bayes tutorials for using 2+ years of data? The only one I can find uses a single year to predict. $\endgroup$
    – NBC
    Jul 22 '18 at 23:51
  • $\begingroup$ Hi @NBC, you don’t need naive bayes, since it’s time series you could use naive time series forecasting. You could also use simple linear regreasion as well, the only problem you may face is that you have 52 seasonal dummies, which is quite large for 2 years or 104 weeks. $\endgroup$
    – forecaster
    Jul 23 '18 at 0:04
  • $\begingroup$ @forecaster do you mean naive or seasonal naive? seasonal naive seems like a better idea here. $\endgroup$
    – Skander H.
    Jul 23 '18 at 17:23
  • $\begingroup$ @Alex seasonal naive in this case, since the data is clearly seasonal. $\endgroup$
    – forecaster
    Jul 23 '18 at 17:44

Polynomial regression assumes the samples are i.i.d, which is not the case for time series data. You can still use it, but it's something to be aware of.

If you are only concerned with making good predictions I would look toward methods in machine learning like Random Forests. If you're comfortable with mathematics there are plenty of nice papers that deal with time series data in the field.


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.