←back to thread

693 points hienyimba | 2 comments | | HN request time: 0.409s | source
Show context
pc ◴[] No.28523805[source]
(Stripe cofounder.)

Ugh, apologies. Something very clearly went wrong here and we’re already investigating.

Zooming out, a few broader comments:

* Unlike most services, Stripe can easily lose very large amounts of money on individual accounts, and thousands of people try to do so every day. We are de facto running a big bug bounty/incentive program for evading our fraudulent user detection systems.

* Errors like these happen, which we hate, and we take every single false rejection that we discover seriously, knowing that there’s another founder at the other end of the line. We try to make it easy to get in touch with the humans at Stripe, me included, to maximize the number that we discover and the speed with which we get to remedy them.

* When these mistaken rejections happen, it’s usually because the business (inadvertently) clusters strongly with behavior that fraudulent users tend to engage in. Seeking to cloak spending and using virtual cards to mask activity is a common fraudulent pattern. Of course, there are very legitimate reasons to want to do this too (as this case demonstrates).

* We actually have an ongoing project to reduce the occurrence of these mistaken rejections by 90% by the end of this year. I think we’ll succeed at it. (They’re already down 50% since earlier this year.)

replies(25): >>28524033 #>>28524044 #>>28524048 #>>28524050 #>>28524154 #>>28524171 #>>28524182 #>>28524398 #>>28524413 #>>28524431 #>>28524441 #>>28524749 #>>28525580 #>>28525617 #>>28525758 #>>28526933 #>>28527035 #>>28527043 #>>28527233 #>>28527269 #>>28527682 #>>28528656 #>>28529788 #>>28530370 #>>28537774 #
yjftsjthsd-h ◴[] No.28524182[source]
> We actually have an ongoing project to reduce the occurrence of these mistaken rejections by 90% by the end of this year. I think we’ll succeed at it. (They’re already down 50% since earlier this year.)

How can you tell? It seems, naively as an outsider, like the problem is precisely that you can't tell if they should have been rejected, in which case you can't tell how often it happens?

replies(1): >>28524264 #
1. pc ◴[] No.28524264[source]
Yeah, good question. First, we aren’t trying to calculate the absolute rate, just relative changes. (The absolute rate would be nice to know but it’s not needed to know whether we’re getting better or worse.) Methodologically, we sample/scrutinize rejections manually and also look at the occurrence of discovered false rejections. But you’re right that there could be some dark matter that we never become aware of.
replies(1): >>28528778 #
2. mdoms ◴[] No.28528778[source]
Well it looks like the Brains Trust inside Stripe has found a way to duke your OKRs, because this guy's appeal was denied and he was cut off anyway. No wonder your "incorrectly identified as fraud" metric is trending down if your staff are simply doubling down on incorrect accusations instead of copping them. Sounds like Goodhart's Law in action - do you happen to tie bonuses to OKRs?