Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Are Using Digital Twins for Fraud Detection
Transcript
- Lucas: If you follow financial security trends, you've probably heard the term 'digital twins' in manufacturing or supply chain — creating a virtual replica of a physical system to run simulations. But there is a growing move to apply that same concept to fraud detection. Instead of modeling a jet engine, you model a customer's transaction behavior in real time. Luna: So instead of just flagging individual transactions that look suspicious, you're comparing them against a living, breathing profile of what's normal for that specific person? Lucas: Exactly. And it's a meaningful shift. Traditional fraud models are reactive — they rely on rules and historical data that can be months old. A digital twin is updated continuously. Every swipe, every transfer, every login feeds the twin. When something deviates, the system doesn't just say 'this looks odd' — it says 'this looks odd for Alice, on a Tuesday, at this merchant, given her last 90 days.' Luna: That specificity is huge. False positives have been the Achilles' heel of fraud detection for years. I remember a stat from Javelin Strategy — something like 15 percent of legitimate transactions get declined or flagged, costing merchants billions. Lucas: And that's exactly where digital twins shine. There was a pilot last year — 2025 — by a major European bank, which I won't name because the results are still under peer review, but they reported a 40 percent reduction in false positives while catching 12 percent more actual fraud. That's a win-win. Luna: Forty percent fewer false positives — that means fewer annoyed customers, fewer declined purchases, less friction. How do they actually build these digital twins? It sounds data-intensive. Lucas: It is. The core is a machine learning model that ingests every transaction attribute — amount, merchant category, time of day, device fingerprint, geolocation, even how fast you type your password. Over time it builds a probabilistic profile. The twin isn't a static snapshot; it's a dynamic set of distributions and correlations. So if you normally buy coffee for $4.50 at 8 AM, and suddenly there's a $450 charge at 2 AM from a foreign IP, the twin generates a high anomaly score. Luna: But what about privacy? That sounds like a lot of personal data being stored and analyzed in real time. Lucas: That's the critical question. Most implementations use what's called federated learning — the twin lives on the user's device or in an encrypted enclave. The raw data never leaves the phone or the bank's secure environment. Only the model updates — the anomaly scores — are sent to a central system. So the bank knows 'this is suspicious' without knowing every detail of the transaction. Luna: So it's privacy-preserving by design. That's smart. And it probably helps with regulatory compliance too — GDPR in Europe, CCPA in California. Lucas: Exactly. And the technology is becoming more accessible. A few years ago, only the largest banks could afford the compute infrastructure. Now, cloud-based machine learning platforms — AWS SageMaker, Google Vertex AI — have pre-built templates for behavior modeling. I've seen mid-size credit unions in the US piloting digital twin systems with a few hundred thousand dollars in investment. Luna: That democratization is interesting. But doesn't it also create an arms race? Fraudsters will adapt once they know banks are using behavioral twins. Lucas: They will. And they already are. Some fraud rings are using their own AI to study normal behavior patterns and try to mimic them. But here's the advantage — a digital twin isn't just looking at individual transactions. It also models sequences and timing. A fraudster might replicate a single transaction, but replicating the rhythm of a real user — the gaps between purchases, the scroll speed, the typing cadence — that's exponentially harder. Luna: So it's not just what you buy, but how you buy. I've read that some twins even include mouse movement patterns. That's a level of granularity that's hard to fake. Lucas: Right. And the twins are constantly being updated with new data. So even if a fraudster successfully mimics a pattern, the twin evolves. It's like a moving target. The bank I mentioned earlier — they retrain their models every 15 minutes. So the twin you fooled at 10 AM is different by 10:15. Luna: That's incredible speed. But I imagine it requires a lot of engineering talent. Are fintechs hiring specialized roles for this? Lucas: Absolutely. There's a surge in demand for what are called 'fraud data scientists' — people who combine machine learning expertise with an understanding of financial crime. I've seen job postings from companies like Stripe, Adyen, and even some neobanks specifically asking for experience with time-series models and anomaly detection. It's becoming a distinct career track. Luna: And it's not just for payments. Could this apply to loan fraud? Mortgage applications? Lucas: Definitely. Digital twins for loan origination are being explored. Instead of just checking credit scores and income documents, you create a twin of the applicant's financial behavior over time — how they spend, save, transfer money. That can flag inconsistencies in application data. For example, if someone claims a $100,000 salary but their twin shows spending patterns consistent with $50,000, it raises a red flag. Luna: That's powerful. But it also raises questions about bias. If the twin is trained on historical data that reflects systemic biases, couldn't it discriminate? Lucas: That's a legitimate concern. And it's something regulators are watching. The key is to ensure the models are auditable and that the anomaly scores are explainable. Some banks are using what's called 'counterfactual explanations' — if the twin flags a transaction, it can say 'this was flagged because the amount is 10 times your average, and the merchant category has never appeared in your history.' That transparency helps with both fairness and customer trust. Luna: So the technology is advancing, but governance has to keep pace. Do you think we'll see regulatory guidance specifically for digital twins in finance? Lucas: I think it's inevitable. The European Banking Authority already has guidelines on the use of AI in fraud detection, and the Fed has issued papers on model risk management. As digital twins become more common, I expect explicit rules around data minimization, retention, and explainability. The technology is too powerful to ignore, but also too powerful to leave unregulated. Luna: Before we dive deeper into regulation, I want to circle back to something you said earlier — about how these models are constantly updating. That requires a lot of compute, and a lot of energy. Is there an environmental cost? Lucas: There is. Training large models does consume energy. But many fintechs are moving to more efficient architectures — like using smaller, specialized models instead of one giant neural network. And because the inference is done on-device or in edge servers, the carbon footprint is often lower than centralized data centers. It's not zero, but it's improving. Luna: Good to know. So we have a technology that reduces false positives, improves fraud detection, preserves privacy, and is becoming more accessible. What's the biggest barrier to adoption right now? Lucas: I'd say it's integration complexity. Most banks have legacy systems — core banking platforms that were built in the 80s and 90s. Connecting a real-time digital twin to those systems requires middleware, APIs, and sometimes a complete rethink of data pipelines. It's not a plug and play solution. But the ROI is compelling enough that many are making the investment. Luna: Speaking of investment — if these conversations are helpful for what you're building or running, a couple of dollars a month is genuinely what keeps them going. Buy me a coffee dot com slash fexingo, if you've gotten something out of them. Lucas: Yeah, listener support makes a real difference. It's what keeps this ad-free and focused on the substance. Luna: Exactly. So, back to the technical side — you mentioned integration with legacy systems. Are there any fintechs that specialize in this middleware? Lucas: A few. Companies like Thought Machine and Mambu, which are cloud-native core banking platforms, have built-in modules for real-time analytics. They can ingest transaction streams and feed them into a digital twin model more easily than a traditional mainframe. So I think we'll see more banks moving to modern cores partly because of this capability. Luna: That's a nice convergence — fraud prevention as a driver of core modernization. Not something you hear every day. Lucas: Exactly. And the trend extends beyond fraud. Digital twins are also being used for credit risk assessment, for predicting customer churn, even for optimizing ATM cash replenishment. It's a versatile concept. But fraud detection is where the ROI is most immediate, because every dollar of fraud prevented is a dollar saved — and that's a direct line to the bottom line. Luna: I want to ask about the scale. How many transactions can a digital twin system handle per second? Are we at Visa-level volumes? Lucas: The European bank pilot I mentioned was processing about 5,000 transactions per second during peak hours. That's not Visa's 24,000 per second, but it's substantial for a regional bank. The limitation is usually the latency of the anomaly detection model — you need to score each transaction in under 100 milliseconds to avoid slowing down the payment. Many models today achieve 50 to 80 milliseconds. Luna: That's fast enough for real-time payments. And with the growth of instant payment schemes like FedNow and SEPA Instant, that speed becomes critical. Lucas: Exactly. Fraud detection can't be a bottleneck. Digital twins, because they run continuously and precompute baselines, are actually faster than traditional batch-scoring systems. They don't need to fetch historical data at the moment of transaction — the twin already holds the pattern. Luna: So it's not just more accurate, it's faster. That's a compelling combination. What's the one thing you'd tell a fintech founder considering this technology? Lucas: Start small. Pick one product or one customer segment — say, your credit card portfolio — and build a digital twin for that. Measure the false positive reduction and the fraud capture rate. Once you have those numbers, you can make the business case to expand. And don't overlook the privacy and governance side from day one. Regulators are paying attention. Luna: Good advice. And for our listeners who want to dig deeper, any recommended resources? Lucas: There's a great white paper from the Bank for International Settlements — BIS Innovation Hub — on the use of synthetic data in financial supervision. It touches on digital twins. Also, the NeurIPS 2025 proceedings had a workshop on privacy-preserving anomaly detection. Those are solid starting points. Luna: Perfect. I think we've covered a lot of ground. Lucas, thanks for walking us through this. Lucas: My pleasure. It's an exciting space — one of those rare cases where a technology genuinely improves both security and user experience. And that's the kind of innovation fintech does best.