I'm seeing negative values where we shouldn't have any. Is there a calculation error?
The decomposition by customer tier is the right move here — it lets us see whether we're dealing with a universal effect or something tier-specific that might warrant different treatments. See /u/plg_lucas_sanche/p/plot-0016.
a rebasing in the back half of the series? (edited to fix a unit)
Can we see this by cohort? (edited to fix a unit)
the y axis on this chart
The zero line should stand out more.
Same conclusion here.
What's the retention rate? See /u/plg_lucas_sanche/p/plot-0016.
The 2004 break is an outlier — it shows up in every series from that source.