Closed dom-devel closed 6 years ago
Hello! It was sort of intentional, the reasoning was as follows.
If you are not using exogenous variables as your control then you should be providing a custom model rather than the default one. Which means you can do it - see section 8 in https://github.com/jamalsenouci/causalimpact/blob/master/GettingStarted.ipynb (if you build a custom model without an exog parameter it should do what you want)
It's slightly prescriptive but I was hoping it was a bit more explicit about the model that is being used.
Hello! Been really enjoying the library as it's saved me having to either re-learn R or get R and pandas talking.
One notable question:
In analysis.py you currently block any attempt to use the model without a predictor time series. In R it's still possible to use it with a single Series and it remains useful (although obviously without a strong control actually measuring uplift for example is very hard).
Was just wondering about the decision behind this, is it a functional thing or just a spare time thing? (with this presumably just being a fun side project).
I can bypass it by looking for the first nonzero value in the first column rather than the second, or just forcing to take the max from the pre-period, but this just gives me a flat line prediction unlike in R and at this point I'm reaching the end of my skills.