Comment by noahnoahnoah | Hacker News Reader

Comment by noahnoahnoah | Hacker News Reader

Parent Full thread

Survival analysis is definitely the right way to go for this.

Some good resources in R: http://www.stat.ucdavis.edu/~hiwang/teaching/10fall/R_tutori... http://www.stat.ucdavis.edu/~hiwang/teaching/10fall/R_tutori...

In particular, what it allows you to do is to estimate the future duration -- saying that your average LTV is the 6 months of average revenue from your existing customers as the article suggests is incomplete, because the portion of customers who haven't lapsed will continue to provide revenue. With survival analysis, you can estimate the full durational curve to get this right.

Future value generally doesn't matter a huge amount if you're just looking for gross differences across segments, but anything you're getting particularly precise on (like when you want to figure out allowable cost for advertising based acquisitions), needs to take this into account.

There are tons of other great lessons to apply to "regular" business from actuarial science.

>what it allows you to do is to estimate the future duration -- saying that your average LTV is the 6 months of average revenue from your existing customers as the article suggests is incomplete, because the portion of customers who haven't lapsed will continue to provide revenue.

I'm almost certain we agree, just want to add two clarifying points for those interested:

+You probably want to try experimenting with different combinations of datasets for left-censored, right-censored and full lifetime cases. Or, at least segmenting that way when you perform the survival analysis (Haven't checked out noahnoahnoah's references yet, so they may mention this).

+Also, the other tricky thing is when it's a subscription service (3,6,12 month sort of thing). Main reason this introduces complications is that you need to account for a "unsubscribe signal" that may not sync well with the next prompt to renew. So rather than predict a survival likelihood at the unit of analysis (I assume monthly, maybe weekly) you need to generate a likelihood for the point in time when the user is prompted to renew (and again, you can also vary that to move it up before the real end of the subscription). It's not particularly hard to grasp at a general level, but it takes a lot of careful thinking of edge cases.*

*That description may be a little convoluted, to simplify: What happens when a 12mo subscription shows signs of leaving only 6mo in?

Yep, we agree. Well said.

Reply on news.ycombinator.com