Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use Real-Time Transaction Data for Fraud Prevention
Transcript
- Lucas: If these conversations are useful for what you're building or running, you're in the right place. I'm Lucas. Luna: And I'm Luna. July 30, 2026 — and Lucas, I saw that number from the Fed this morning. Ten point seven billion dollars in card fraud losses last year. That's staggering. Lucas: Yeah, and it's up almost a third from pre-pandemic levels. The bad guys have gotten better — they're faster, they're using synthetic identities — so the counterpunch has to happen in real time. And that's exactly what a bunch of fintechs are doing: using live transaction data to decide whether to approve or block a purchase, often in less than a hundred milliseconds. Luna: So how do they actually do that without flagging half my coffee purchases as suspicious? Lucas: Great question. The classic model was rules-based — if you spend over $500 in a foreign country, block it. But that's too blunt. Now they feed millions of data points per second into machine learning models. Things like: where you're shopping, what device you're using, how fast you're entering your card number, even the angle at which you tap your phone at a terminal. Luna: Wait — the angle of the tap? That's a real signal? Lucas: Some fintechs are experimenting with it. The idea is that your physical behavior — the way you hold your phone, the pressure you apply — has a unique pattern. Combined with location and typical purchase velocity, it becomes a biometric signature. Not every company uses it, but the more signals, the better the model. Luna: But more signals also means more complexity. How do they keep the false positive rate low? Lucas: That's the balancing act. Take Stripe Radar for example. Stripe claims that their machine learning models reduce false positives by about 30 percent compared to traditional rules. They do it by training on millions of transactions and constantly updating — so if the model sees a pattern that looks sketchy but later turns out to be legitimate, it adjusts. It's a feedback loop that runs in real time. Luna: So the system gets smarter after every decision. That feels like the only way to keep up with fraudsters who also learn. Lucas: Exactly. And it's not just about blocking — sometimes the system will let a borderline transaction through but add extra authentication, like a push notification asking 'did you just buy this?' That's less friction than a hard decline, and the data from that interaction goes back into the model. Luna: Is there a trade-off? I mean if you're too aggressive, you lose customers. Too lenient, you lose money on chargebacks. Lucas: Right. Some fintechs say they aim for a sweet spot where the authorization rate stays above 96 percent while fraud losses stay below 0.1 percent of volume. A digital bank I spoke with said they stopped about one hundred fifty million dollars in fraudulent transactions last quarter — but they also approved a bunch of transactions that would have been declined under their old system. That improved customer satisfaction. Luna: One hundred fifty million — that's real money. But it's also a cat and mouse game. What's the next frontier? Lucas: I think we'll see real-time registries of compromised cards. Right now, when your card is stolen, it can take days to propagate that info. Some fintechs are sharing fraud signals in real time via consortiums. Also, quantum-safe encryption is on the horizon — because once quantum computers arrive, current encryption is toast. Luna: Before we go deeper into that — and I know this is a little meta — but the same kind of small, real-time contributions that power these fraud models also keep this podcast going. A couple of dollars a month from listeners genuinely makes a difference. It's like those micro-signals: individually small, collectively huge. Lucas: Yeah, we keep the show ad-free because of people who throw a few bucks into the hat at buy me a coffee dot com slash fexingo. If you've gotten something out of these episodes, that's the place. Luna: It adds up — just like all those transaction data points. So back to the quantum piece: are fintechs actually preparing for that now, or is it still a distant threat? Lucas: A few are already testing post-quantum cryptographic standards from NIST. They worry that fraudsters could harvest encrypted transaction data today and decrypt it later — so-called 'harvest now, decrypt later.' The timeline is maybe five to ten years, but the migration takes years, so starting now makes sense. Luna: Interesting. So the war on fraud has a long horizon. Any other data streams you're watching? Lucas: Open banking data is a big one. If a fintech can see your current account balance and recent transactions in real time — with your permission — they can make much smarter decisions about whether a transaction is normal for you. One European fintech I looked at reduced false declines by 40 percent using open banking signals alongside traditional card data. Luna: Forty percent — that's huge for merchants who lose sales on false declines. And it's all happening in milliseconds. Lucas: Right. And the models keep learning. That's the beauty — as long as the data keeps flowing, the models get harder to fool. It's a constant arms race, but the fintechs have the advantage of speed and scale. Luna: Alright, next episode we'll look at another real-time data use case. Thanks Lucas. Lucas: Thanks Luna. See you next time.