1
$\begingroup$

I'm build a model where I need to account for seasonality, but the variable nature of Easter in particular is problematic.

I have about 7 years of sales data (at a weekly level), for 400 stores.

I'm not sure which approach would be best. ARIMA seems like a good option for accounting for Seasonality, but I'm struggling to account for Easter/Christmas and I'm not sure how I would go about levering the 400 parallel timeseries to improve the accuracy of the model.

Are there supervised learning functions that handle seasonality well, potentially using Easter and Christmas as boolean features?

$\endgroup$
1
  • $\begingroup$ post 1 example store and state the beginning date and the country the data is from. If you have any promotion variables provide them as well. $\endgroup$ – Tom Reilly May 2 '17 at 18:00
1
$\begingroup$

I'd suggest you look at causal models, using Boolean dummies for Easter (and possibly for Christmas). You'd need to play around a bit to find the optimum settings for these predictors. Ramp-ups might also be useful.

As for methods, you could use either ARIMAX or regression with ARIMA errors. There is a difference, see Rob Hyndman's ARIMAX model muddle blog post. I personally prefer regression with ARIMA errors - it's easier to explain and understand, and it's implemented by default in the standard R forecasting packages.

This free online forecasting textbook by Hyndman and Athanasopoulos might be useful.

$\endgroup$
1
  • 1
    $\begingroup$ Thanks for your help, very useful. I think that Bayesian Structural Time Series models better allows for regressors. The bsts package in R also appears to have built in support for moving holidays. $\endgroup$ – QuinRiva May 29 '17 at 3:00

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.