Hacker News Reader

Comment by sbov | Hacker News Reader

I've never used stripe specifically, but in my experience its always good to have your own layer of fraud prevention, because if you have too much fraud reach whoever is providing payment processing, they will terminate your account.

Beyond that, a few simple checks will make most fraudsters move onto other websites. It's not about being bulletproof, it's about being harder to abuse than the others.

View on HN

I think the point is that it's very hard to do as good a job as providers like Stripe and SiftScience will do (assuming machine learning against fraud is not your core competency) - they take into account a lot of the signals you might take into account anyway if you feed them clean data. Better to focus on sending them the clean data.

totally false. Machine learning is not my companies core competency, but we do have a team of data scientists (myself included). Most of the data we get from our data providers is garbage by itself, but using our own data along with features from providers, we are able to pick out a huge amount of fraud.

These types of models are actually quite simple to build and implement if you actually have data.

Exactly the opposite. Generic fraud checks will perform worse

They're not generic, though. Those providers build custom ML models specifically from your data (and they pull in a ton of signals for this), and then have the data of many other merchants to work with.

At a previous employer we found both of those products to be effective.

These systems are the most frustrating to me, because it is the unusual transactions that are more important than the usual ones.

If I have eggs delivered every week, and one of those transaction fails, who cares?

If I buy a plane ticket for a flight that leaves in 30 minutes because of an emergency, I will be fucking enraged if that transaction is blocked.

Similarly if I can't pay the hotel at check-out and because of this delay I miss my flight.

Or when I arrive in a new country and can't pay for the taxi ride, and the ATMs don't work either.

All these events have happened to me.

Effective at preventing fraud - sure But what about false positives?

If the question is whether you'll have any - of course you will! You will no matter what method you go with. At scale it's a pretty easy curve to optimize revenue on top of though, and you have to be willing to accept that bit. You know your customer LTV, you know your fraud rate/cost.

Sift gives you direct control of expected false positive rate, Radar hides that optimization under the hood I think?

Reply on news.ycombinator.com