Another way to do this is based on a paper by Fader & Hardie [0] that uses a probabilistic model, instead of a regression-based one. It's highly readable, and very intuitive. The technique workss particularly well when there's only limited cohort data (e.g. early days of a startup) and you can't just blend years of prior cohort data. A former colleague and I made a web-based implementation as a little weekend project if anyone wants to try it out: http://retentionizer.herokuapp.com
[0] https://marketing.wharton.upenn.edu/files/?whdmsaction=publi...