Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI for Real-Time Purchase APR Optimization
Transcript
- Lucas: Luna, I want to talk about something that sounds a bit niche but actually affects almost anyone with a credit card or a buy now pay later account: the interest rate you're offered at the moment of purchase. Luna: You mean the APR, right? But that's set when you open the account, typically. How could it change per purchase? Lucas: That's exactly the old model. A handful of fintechs are now using AI to offer what they call 'real-time purchase APR' — a rate that's dynamically calculated for each transaction based on risk signals that include not just your credit score but also the merchant, the time of day, the device you're using, even the category of purchase. Luna: So, a lower rate if I'm buying groceries at 10 a.m. from a known store, but a higher rate if I'm buying a gaming console at 2 a.m. from a new merchant? That feels a little intrusive, but also kind of smart. Lucas: Exactly that. And the data that fuels these models goes far beyond traditional credit bureau data. Companies like Upstart and Affirm have been moving in this direction, though they frame it differently. Upstart calls it 'risk-based pricing in real time' and Affirm emphasizes transparency — you see the exact APR before you swipe. Luna: Right, Affirm shows you a 'you'll pay this much total' figure. But is the APR actually changing minute by minute for the same user? Lucas: It can. The model ingests hundreds of variables. For example, if you're buying from a merchant with a high historical return rate or high dispute rate, the risk score goes up. If it's a repeat purchase at a store you've used before without issues, the score goes down. The APR might move by a few hundred basis points between two transactions five minutes apart. Luna: That's a huge shift from the 'one APR for life' model. How do consumers react when they see a lower rate for one purchase and a higher one for another? Lucas: Early data suggests they actually engage more. A fintech I spoke with — let's call them Lender A — saw a 12 percent increase in approval rates and an 18 percent reduction in charge-offs within six months of rolling this out. That's because the pricing better aligns with risk. Low-risk users get rewarded with lower rates, so they use the product more. High-risk users either see a rate that reflects their risk and take it, or they decline — which actually saves the lender from a bad loan. Luna: So it's a win-win if the model is accurate. But what about fairness? If a user's data profile penalizes them for shopping at certain merchants or at odd hours, isn't that potentially discriminatory? Lucas: That's the big regulatory question. The Consumer Financial Protection Bureau has been looking at this. The key is that the variables must be 'creditworthy' and not proxies for race, gender, or age. Using merchant category codes — like 'grocery' vs 'electronics' — is generally considered okay if it's statistically linked to repayment behavior. But using location or time of day could raise red flags if it correlates with protected classes. Luna: I've read that some fintechs are using device fingerprinting — the make and model of your phone — as a signal. That seems like it could be problematic. Lucas: It is. Device type can correlate with income and, indirectly, with race. So regulators are watching closely. But the fintechs argue that these models are more inclusive, not less, because they consider hundreds of signals beyond a thin credit file. Someone with a low credit score but a stable purchase pattern at consistent merchants might get a better rate than a traditional FICO model would give them. Luna: That's a strong argument. So, concretely, how does the AI work under the hood? Are we talking about gradient boosting models, neural networks? Lucas: Mostly gradient boosting — XGBoost, LightGBM — because they're interpretable enough for regulatory review and they handle mixed data types well. The model is retrained frequently, sometimes daily, to capture shifting patterns. The feature engineering is the real secret sauce: things like 'time since last purchase,' 'average transaction amount over last 30 days,' 'number of merchants visited,' and even 'hour of day' as a cyclical feature. Luna: And the inference has to happen in milliseconds — at the point of sale — so it's a real-time system. Lucas: Exactly. The latency requirement is brutal. You have maybe 200 milliseconds to compute the risk score and return an APR. That means the feature store has to be precomputed and cached. The model itself is usually a lightweight ensemble that runs on a GPU or even a CPU if optimized well. Luna: What about the macroeconomic environment? Do these models incorporate things like the current Fed rate or unemployment data? Lucas: They do. A good real-time pricing model includes macro features that are updated daily — like the ten-year Treasury yield, the unemployment rate, or consumer confidence index. That way, if the economy is tightening, the model automatically shifts the base rate upward for all users, even without changing the underwriting criteria. It's a much more responsive approach than the old quarterly portfolio review. Luna: It sounds like these fintechs are effectively running a mini central bank for their own portfolio. Adjusting rates in real time based on supply and demand of credit. Lucas: That's a great way to put it. And the benefits show up in the numbers. One study from a large fintech lender showed that dynamic APR pricing increased net interest margin by about 1.5 percent over a static APR model, while keeping the charge-off rate flat. That's a huge lift in profitability. Luna: So, is this the future of all consumer lending? Will static APRs disappear? Lucas: I think we'll see a hybrid model for a while. Traditional credit cards still operate on a fixed APR range, but some issuers are experimenting with offering a 'smart APR' option where you opt in to dynamic pricing. The regulatory uncertainty is the biggest hurdle. But the technology is proven, and the consumer benefit — lower rates for responsible borrowers — is compelling. Luna: It also requires a level of financial literacy. If consumers don't understand why their rate changed, they might feel cheated. Lucas: Absolutely. Transparency is key. The firms that do this well, like Affirm, show you the total cost upfront and explain that the rate is based on the transaction and your history. They don't just say 'your APR is X.' They say 'you'll pay Y dollars in interest if you take 3 months.' That's much more digestible. Luna: It really is. And speaking of things that make a difference, we should mention that episodes like this, and the entire Fexingo network, are supported by listeners. A couple of dollars a month is genuinely what keeps these going — buy me a coffee dot com slash fexingo, if you've gotten something out of them. Lucas: Yeah, it's a small thing that makes a big difference for us. So if you're finding these conversations useful for what you're building or running, that's the place. Luna: Back to APRs — one more angle: how do these models handle fraud? If a fraudster gets your card and tries to make a purchase, the dynamic APR might actually be very low because the model sees a 'normal' pattern initially, right? Lucas: That's a great point. The real-time APR model is usually paired with a real-time fraud model. If the fraud score is high, the APR might be set very high — effectively denying the transaction — or the system might decline the purchase outright. The two models work in parallel. The APR model assumes the user is legitimate and prices risk; the fraud model decides if the user is who they say they are. Luna: So they complement each other. The fraud model acts as a gate, and the APR model prices the risk for legitimate users. Lucas: Exactly. And when both models are trained on the same data pipeline, you get a very robust system. I know of one fintech that reduced fraud losses by 22 percent just by adding the real-time APR model as an additional risk signal — because a sudden spike in APR request can indicate abnormal behavior. Luna: That's fascinating. So the APR itself becomes a data point for fraud detection. Lucas: Precisely. It's a virtuous cycle. The more data you have, the better both models get. And as these systems become more common, we'll see the line between pricing and fraud detection blur even further. Luna: Alright, Lucas, any predictions? Will the major card networks like Visa and Mastercard get into this space? Lucas: They already are, indirectly. Visa has been piloting real-time risk scoring for transactions, and Mastercard has their Decision Intelligence product. But they're not yet offering dynamic APR at the network level — that would require cooperation from issuers. I think we'll see the first major bank offering dynamic purchase APRs within two years. It's too powerful a tool to ignore. Luna: It really does feel like the next frontier in consumer lending. Thanks for breaking it down. Lucas: Always a pleasure. Until next time.