Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI to Predict Credit Card Churn
Transcript
- Lucas: So we just got a dataset from a top-ten card issuer — they ran a pilot last quarter using machine learning to predict when someone is about to cancel their card. The model flagged customers up to forty-five days before they actually churned. Luna: Forty-five days? That's not just early warning — that's an intervention window. What signals was it tracking? Lucas: Three main categories. First, spend velocity — the rate of decline in monthly transaction volume. If you normally put three thousand dollars a month on your card and suddenly it drops to twelve hundred for two months running, that's a flag. Second, merchant category mix shift. People who churn often start pulling spend away from general retail and into essentials — groceries, gas, utilities — which suggests they're de-emphasizing discretionary rewards. Luna: So they're still using the card, but the they're using it changes. That's subtler than just a zero balance. Lucas: Exactly. The third signal is payment behavior: if someone who always paid in full starts carrying a balance, or conversely, someone who typically carried a balance suddenly pays off everything, both can be precursors. The model weighs these together and outputs a churn probability score per customer per week. Luna: And they validated it against actual cancellations? What was the hit rate? Lucas: The pilot achieved about seventy-three percent precision — meaning when the model said 'this person will churn within six weeks,' they were right three out of four times. They then triggered a retention flow: a targeted offer based on the customer's historical spend pattern. For someone who loved travel rewards, they offered a bonus miles promotion. For a cash-back loyalist, they waived the annual fee for six months. Luna: I love that personalization. A generic 'we miss you' email would probably backfire. Did they measure the impact on actual churn reduction? Lucas: They did. In the test group that received the targeted offer, churn dropped twenty-two percent compared to the control group that got no intervention. The issuer calculated that each prevented cancellation saved them roughly three hundred and fifty dollars in acquisition cost to replace that cardholder. Luna: That's a strong ROI story for the data science team. But I wonder — how do they handle false positives? Twenty-seven percent of flagged customers weren't actually going to churn. If you give them a retention offer they didn't ask for, you're just giving away margin for free. Lucas: That's the ethical tightrope. The issuer we spoke to said they specifically designed the offer to be non-intrusive — it appears in the app as a 'limited-time benefit' rather than a 'we think you're leaving' message. That way, even if the model was wrong, the customer feels valued rather than tracked. But some critics argue that any predictive intervention is manipulative, because it's acting on inferred intent before the customer has made a decision. Luna: I've seen studies where customers who discovered they were 'predicted to churn' actually churned faster — they felt their privacy was violated. So the communication strategy is almost as important as the model accuracy. Lucas: Exactly right. The issuer we talked to acknowledges that risk. They ran a separate A/B test on the messaging itself: a 'we value your relationship' frame versus a 'here's an exclusive offer just for you' frame. The exclusive offer outperformed by fourteen percentage points in retention. Luna: So the framing matters as much as the math. Let's talk about the data side — where does the model get its inputs? Are they pulling from transaction rails, credit bureau data, app engagement? Lucas: Primarily transaction data from the card network — that gives them the merchant category codes, spend amounts, and timestamps. They layer on app login frequency, push notification click rates, and customer service contact history. Interestingly, they found that a spike in calls to the call center — even if the calls were about non-account issues — correlated strongly with churn within thirty days. Luna: That makes sense. If you're thinking about switching cards, you might call to ask about a policy detail or check your rewards balance. It's a behavioral signal most issuers ignore. Lucas: So the model architecture itself is a gradient-boosted decision tree — XGBoost, specifically — trained on twelve months of historical data. They update the model weekly with new transaction batches. One thing they emphasized: they don't use credit score changes as an input, because they found that churn prediction is more about change than credit health. Luna: That's a smart constraint. Credit scores can lag behavior. If you lose your job, your spending drops immediately but your credit score might not dip for two or three months. Lucas: Precisely. The goal is to capture the leading indicator, not the trailing one. They also built a separate model specifically for the 'revolver to transactor' pattern — people who used to carry a balance but suddenly pay in full. That subgroup turned out to have a churn rate two and a half times higher than the overall population. Luna: Because they're likely paying off the card to close it cleanly. That's a classic death knell signal. Lucas: Exactly. So the issuer now treats that specific pattern as a high-priority trigger — they intervene within seventy-two hours of detecting the full payoff. The offer in that case is usually a zero-percent balance transfer promo, which gives the customer a reason to keep the account open. Luna: That's clever — you're meeting them where they are. They want to close the card, you give them a financial incentive not to. Let's zoom out a bit. How common is this practice across the industry? Are most issuers doing this now, or is it still cutting edge? Lucas: It's becoming standard among the top fifteen issuers, but the sophistication varies widely. Some are still using simple rule-based triggers — like 'if a customer hasn't transacted in ninety days, send a win-back email.' The machine learning approach is probably adopted by maybe a third of large issuers today, but it's growing fast. The competitive pressure is real: if one issuer in a duopoly can predict churn forty-five days early, the other has to catch up or lose profitable cardholders. Luna: And there's a data network effect, too. The more customers you have, the more churn patterns you can train on, which improves your model, which keeps more customers. It's a flywheel. Lucas: Exactly. But it also means smaller issuers — community banks, credit unions — are at a disadvantage. They don't have the transaction volume to build robust models. Some are turning to fintech vendors that offer churn prediction as a service, but then they're handing over their most sensitive customer data. Luna: That's a real trade-off — access to AI versus data privacy. If I'm a community bank with fifty thousand cardholders, do I go with a vendor or build my own simpler model? Lucas: Adrian Gostick's rule of thumb applies here: 'Don't let perfect be the enemy of good.' A simpler logistic regression model using just spend velocity and time since last transaction might capture sixty percent of the predictive power with ten percent of the complexity. That's often good enough to start. Luna: That's actually a really useful heuristic for a lot of fintech applications — start simple, validate, then layer on complexity. Lucas: If today was actually useful to you, the way these stays ad-free is listener support. If you want to help keep conversations like this going, you can buy me a coffee dot com slash fexingo. It's a small gesture that makes a big difference for us. Luna: Yeah, we really appreciate the listeners who chip in — it lets us keep the episodes focused on the substance rather than fitting in ad breaks. Lucas: Alright, back to the data. One more angle I want to hit: what happens the retention offer is accepted? The issuer we studied found that about forty percent of customers who accepted an offer churned anyway within the next six months. Luna: So the offer delayed the inevitable, but didn't fix the underlying reason they were leaving. That's the second-order problem. Lucas: Right. So they built a second model that predicts 're-churn risk' after intervention. That model looks at whether the customer actually used the retention offer — for example, did they redeem the bonus miles? Did they start spending again? If they accepted the offer but didn't use it, that's a red flag that the offer itself wasn't compelling enough. Luna: That's a really sophisticated loop. You're not just predicting churn once — you're predicting the effectiveness of your own intervention. Have they started using that to dynamically tailor the offer in real time? Lucas: They're testing that now. Instead of a static offer, they're building a reinforcement learning model that adjusts the offer amount based on the customer's responsiveness. For example, if a customer didn't engage with a fifty-dollar statement credit, the model might try a hundred-dollar credit next time. It's early-stage, but they're already seeing a ten percent lift in six-month retention for the adaptive group versus the static offer group. Luna: That's the frontier — not just predicting churn, but optimizing the cure. I think that's where the real value is for the next wave of fintech AI. Lucas: Absolutely. And it raises the question: at what point does the intervention become a bidding war for your own customers? If you know someone is going to leave, how much are you willing to pay to keep them? That's a strategic question every issuer will have to answer. Luna: And it ties back to lifetime value. If you know a cardholder's predicted CLV is twelve hundred dollars, you can rationally spend up to that to keep them. But if your model is wrong, you're burning margin. Lucas: Exactly. So the best practice right now among the leading issuers is to set a hard cap on retention spend — typically twenty percent of estimated lifetime value — and let the model optimize within that constraint. That way you're never overpaying to keep a customer who was going to leave anyway. Luna: That's a clean risk management framework. I'm going to remember that.