Can I use bsts with external predictors such as promotions, holidays etc? How do I do it?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
7
You can if you correctly pre-specify the lead and lag structures around each holiday. Also care needs to be taken to properly specify and level shifts or trnd changes that may have occurred. Also you need to make sure that there are no anomalies ( 1 time pulses) in your data. Also you need to specify if all of the data should be used or just a recent portion reflecting transience in the model parameters. You also need to be concerned with any non-constant error variance is lurking and needs to be remedied. Additionally there may be some ARIMA structure that may be necessary to render a final set of model residuals to be free of structure .. always a good idea. Finally if there are particular days or weeks in the month or months in the year that are consistently different then appropriate dummy variables need to be pre-specified.
Finally there may be pre-promotion response that might need to be identified and incorporated to reflect assignable cause.
Finally the way you identify pre and post effects around each holiday is to either spend a lot of time looking at graphs or use a sophisticated search engine to effectively suggest this via a form of AI.
-
$\begingroup$ So in your opinion, does bsts provide a better time series forecast (both in terms of accuracy and robustness) versus exponential smoothing? $\endgroup$ Commented Jul 25, 2022 at 4:08
-
1$\begingroup$ I would think so as it is more general . $\endgroup$ Commented Jul 25, 2022 at 11:26
-
$\begingroup$ That comes with the caveat that we would need to fill in all the parameters you mentioned above, and many more, correctly? Which is a tough task on its own. $\endgroup$ Commented Jul 25, 2022 at 13:25
-
$\begingroup$ or preferably do a stepdown and only use the statistically significant parameters $\endgroup$ Commented Jul 25, 2022 at 20:50
-
1$\begingroup$ Stepdown is when you delete insigificant parameters from a model and just retain the statistically significant ones. $\endgroup$ Commented Jul 26, 2022 at 16:42