Description
While using the budget allocator, I was seeing unusually high adstock for certain channels - levels that didn't match the other outputs, and worse still, noticing that historical adstock was being added to channels that only became active during the budget allocation window, e.g. a channel with spend from March-December showing adstock when the window starts in January.
When investigating the adstock_weibell function, I noticed that the scale parameter is increased based on the windlen variable - calculated based on the length of the input data. I'm not sure exactly what this is supposed to be, but it definitely looks incorrect, since the object whose length it checks is not the filtered timeframe used for modelling, but the length of the input file itself, which could be significantly longer.
Parallel runs of Robyn with only one change - a zero-padded dt_input fed to robyn_inputs - while having a window_start and window_end that should exclude the zero-padding, will produce different results for budget allocation.