Refresh generates a completely new model

[CMoranoA]

Project Robyn

Describe issue

I have a model trained with 79 weeks of data. When I refresh the model with 13 more weeks, it generates a completely new model, with new coefficients, hyperparameters, etc., even when setting bounds_freedom = NULL. Particularly, turning off some channels that were supposed to be relevant.

I find this problematic because it’s not easy to communicate contradictory results to stakeholders after just 3 months of data.

I have found similar errors in other threads (#985), but I can’t find a solution in any of them.

Am I missing something?

Environment & Robyn version

packageVersion("Robyn")
[1] ‘3.11.1.9003’

• R version (Please, check and share: sessionInfo() or R.version$version.string)
  

Thanks in advance!

[gufengzhou]

Sorry for the late reply. In general we're aware that the refresh function has flaws. Indeed, refresh is actually remodeling with constraints (on hyperparameters) from the prev. selected model, whereas these constraints will be larger the more periods you're refreshing. The logic behind is this: When adding 13 weeks to refresh, the model is less confident to "extrapolate" for the future, thus wider constraints. In which case you might as well just remodel. If you're only refreshing 1 week, the constraints will be way narrower, thus the results should be way closer. Hope it helps.

I just tried it out with only 1 week refresh with the demo data. They're certainly way closer.

Also closing it because it's an old one. Feel free to reopen.