Yeah I am still seeing a big difference with the 30% default. While the total for ci_lo is around the historical KPI range, the mean is really high although I am using the default 0.9 confidence_interval

the response curves do not look reasonable

The response curve shows each channel’s incremental outcome at increased levels of spend. Thus it is important for the budget allocator. The first step here is to make sure you have reasonable response curves. Some considerations,

• It can be difficult to correctly infer the response curve for ranges of media execution beyond what’s observed in the data. If a channel has low variation in the media argument of InputData, it may be the case that the budget allocator is going beyond what’s observed in the data. Prevent this by decreasing the budget constraints for any such channels to a value that ensures budget allocation doesn't go beyond what's observed in the data.
• Consider a custom ec_m prior that encourages more saturated response curves. A prior with more mass at smaller values will encourage this.
• Consider setting the paid_media_prior_type argument to 'mroi'. A more informative prior on mROI will shrink posterior optimal budget allocation towards historical.

Thank you for the considerations @alexwakim. Tested the approaches you suggested and it doesn't seem to make a big difference in truncating the response curves.

• To your first point, the InputData I am using goes back to around two and a half years with sufficient variance across media channels.
• The custom ec_m prior did help truncate the scale of the response curves by a little bit, but the volume is still higher than historical.
• I have been modeling towards kpi so setting paid_media_prior_type argument to 'mroi'. might not make sense.