Payment Data Analytics

TL;DR: We built a transformer-based payments foundation model. It works. For years, Stripe has been using machine learning models trained on discrete features (BIN, zip, payment method, etc.) to improve our products for users. And these feature-by-feature efforts have worked well: +15% conversion, -30% fraud. But these models have limitations. We have to select (and therefore constrain) the features considered by the model. And each model requires task-specific training: for authorization, for fraud, for disputes, and so on. Given the learning power of generalized transformer architectures, we wondered whether an LLM-style approach could work here. It wasn’t obvious that it would—payments is like language in some ways (structural patterns similar to syntax and semantics, temporally sequential) and extremely unlike language in others (fewer distinct ‘tokens’, contextual sparsity, fewer organizing principles akin to grammatical rules). So we built a payments foundation model—a self-supervised network that learns dense, general-purpose vectors for every transaction, much like a language model embeds words. Trained on tens of billions of transactions, it distills each charge’s key signals into a single, versatile embedding. You can think of the result as a vast distribution of payments in a high-dimensional vector space. The location of each embedding captures rich data, including how different elements relate to each other. Payments that share similarities naturally cluster together: transactions from the same card issuer are positioned closer together, those from the same bank even closer, and those sharing the same email address are nearly identical. These rich embeddings make it significantly easier to spot nuanced, adversarial patterns of transactions; and to build more accurate classifiers based on both the features of an individual payment and its relationship to other payments in the sequence. Take card-testing. Over the past couple of years traditional ML approaches (engineering new features, labeling emerging attack patterns, rapidly retraining our models) have reduced card testing for users on Stripe by 80%. But the most sophisticated card testers hide novel attack patterns in the volumes of the largest companies, so they’re hard to spot with these methods. We built a classifier that ingests sequences of embeddings from the foundation model, and predicts if the traffic slice is under an attack. And it does this all in real time so we can block attacks before they hit businesses. This approach improved our detection rate for card-testing attacks on large users from 59% to 97% overnight. Perhaps even more fundamentally, it suggests that payments have semantic meaning. Just like words in a sentence, transactions possess complex sequential dependencies and latent feature interactions that simply can’t be captured by manual feature engineering. Turns out attention was all payments needed!

Stripe 𝗯𝘂𝗶𝗹𝘁 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝘁𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿 𝗺𝗼𝗱𝗲𝗹 𝘁𝗿𝗮𝗶𝗻𝗲𝗱 𝗼𝗻 𝗽𝗮𝘆𝗺𝗲𝗻𝘁 𝗱𝗮𝘁𝗮! Not for text and NOT for code, BUT for billions of payments. Think GPT, but instead of learning language, it learned the structure, behavior, and patterns behind every transaction: ⬇️ 𝗛𝗲𝗿𝗲 𝗶𝘀 𝘄𝗵𝗮𝘁 Stripe 𝗷𝘂𝘀𝘁 𝗱𝗶𝗱? For years, Stripe used traditional ML — separate models for fraud, disputes, and authorizations. Each one relied on handpicked features like BIN codes, ZIP codes, email addresses, and payment methods. That worked — but it was narrow, manually intensive, couldn’t scale and most importantly, it missed the bigger picture. So Stripe trained a transformer, just like GPT — but instead of learning language, it learned from billions of transactions. Each payment — from a coffee in Paris to a subscription in Tokyo — was turned into a dense vector: a numerical fingerprint capturing its behavior and context. 𝗧𝗵𝗲 𝗼𝘂𝘁𝗰𝗼𝗺𝗲? ➜ Transactions with similar behavior cluster naturally — by issuer, merchant, location, or risk ➜ Suspicious patterns emerge organically — without handcrafted rules ➜ Fraud becomes easier to detect — not because it was labeled, but because it’s "understood" This foundation model captures now the structure and relationships between transactions — in real time — the way GPT models understand the flow of words in a sentence. Stripe no longer needs a different model for every use case. They’ve built one that generalizes across many — and keeps learning. 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁? They tested it on one of the hardest problems in the space: Card testing attacks that hide in legitimate traffic. ➜ Traditional ML: 59% detection ➜ Transformer-based model: 97% — overnight Visionary work by Stripe! BUT this approach has implications far beyond payments. Great example to see that foundation models aren’t limited to text. The next phase of AI will probably focus more on transformer architectures trained on high-value, underexplored data domains: transactions, supply chains, behavioral signals, scientific processes — even spreadsheets. 𝗜 𝗮𝗺 𝗽𝗿𝗲𝘁𝘁𝘆 𝘀𝘂𝗿𝗲 𝘄𝗲 𝘄𝗶𝗹𝗹 𝘀𝗲𝗲 𝗺𝘂𝗰𝗵 𝗺𝗼𝗿𝗲 𝗱𝗼𝗺𝗮𝗶𝗻-𝘀𝗽𝗲𝗰𝗶𝗳𝗶𝗰 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹𝘀 — 𝗽𝘂𝗿𝗽𝗼𝘀𝗲-𝗯𝘂𝗶𝗹𝘁 𝘁𝗼 𝗼𝗽𝗲𝗿𝗮𝘁𝗲 𝗶𝗻𝘀𝗶𝗱𝗲 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗹𝗶𝗸𝗲 𝗳𝗶𝗻𝗮𝗻𝗰𝗲, 𝗵𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲, 𝗮𝗻𝗱 𝗲𝗻𝗲𝗿𝗴𝘆. 𝗙𝗼𝗿 𝘆𝗲𝗮𝗿𝘀, 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗵𝗮𝘀 𝗯𝗲𝗲𝗻 𝗳𝗼𝗰𝘂𝘀𝗲𝗱 𝗼𝗻 𝗹𝗮𝗯𝗲𝗹𝗶𝗻𝗴 𝗽𝗿𝗼𝗯𝗹𝗲𝗺𝘀. 𝗡𝗼𝘄, 𝘄𝗲'𝗿𝗲 𝗲𝗻𝘁𝗲𝗿𝗶𝗻𝗴 𝗮 𝗽𝗵𝗮𝘀𝗲 𝘄𝗵𝗲𝗿𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 𝗯𝗲𝗴𝗶𝗻 𝘁𝗼 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱 𝘁𝗵𝗲𝗺 — 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗮𝗹𝗹𝘆, 𝗰𝗼𝗻𝘁𝗲𝘅𝘁𝘂𝗮𝗹𝗹𝘆, 𝗮𝗻𝗱 𝗮𝘁 𝘀𝗰𝗮𝗹𝗲. Full story in the comments. 𝗣.𝗦. 𝗜 𝗷𝘂𝘀𝘁 𝗹𝗮𝘂𝗻𝗰𝗵𝗲𝗱 𝗮 𝗳𝗿𝗲𝗲 𝗻𝗲𝘄𝘀𝗹𝗲𝘁𝘁𝗲𝗿 𝗼𝗻 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗮𝗻𝗱 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 — 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗿𝗲𝗮𝗱 𝗯𝘆 𝟮𝟬,𝟬𝟬𝟬+. 𝗝𝗼𝗶𝗻 𝗵𝗲𝗿𝗲: https://lnkd.in/dbf74Y9E

What Is Your Payments Data Worth in an Agentic Commerce World? Last week Google introduced its Agentic Payments Protocol, supported by over 60+ payments companies. In the months leading up to that, we have seen the schemes, including Visa and Mastercard, and other prominent payment methods share their perspective on how they believe Agentic Commerce should work. At Silverflow we are on top of these exciting developments, to help our partners capture the value this will unlock. However, for many payments companies, the value they are looking for, doesn't sit outside of their organization, but inside. Most PSPs are sitting on a treasure chest of value while complaining about thin margins. They've got terabytes of transaction intelligence collecting digital dust in their databases. Because every payment carries signals beyond the obvious amount and merchant details. Time patterns, device fingerprints, velocity indicators, geographic clusters. This data is crucial in helping to predict customer behavior, fraud patterns, and market trends with high accuracy. Yet most PSPs treat this information like exhaust from their processing engines. They store it for compliance, maybe run basic reports, then ignore the strategic goldmine they're accumulating daily. The smart ones are waking up. They're using transaction intelligence to boost authorization rates by 3-5 percentage points. That improvement alone can increase a PSP's revenue by 15-20% without acquiring a single new merchant. But here's where it gets interesting. Better authorization rates and lower fraud costs don't just improve your P&L. They transform your company valuation multiple. Investors pay premium prices for PSPs with superior unit economics and data advantages. A PSP processing $1 billion annually with standard margins might be valued at 8-10x revenue. The same PSP with data-driven performance improvements can command 12-15x multiples. That's the difference between a $800 million company and a $1.5 billion one. The reinvestment cycle kicks in fast. Higher valuations mean cheaper capital for expansion. Better margins fund product development and talent acquisition. Data advantages compound as you process more transactions and refine your models. And in an Agentic world, understanding your data will be crucial. Legacy PSPs miss this because their infrastructure can't handle real-time analytics. They're stuck with overnight batch processing and static risk rules. Modern platforms turn every transaction into actionable intelligence within milliseconds. The winners aren't just processing payments anymore. They're monetizing the intelligence inside those payments. Your data is already worth more than your processing fees. The question is whether you're systems are smart enough to mine it. To learn more about how Silverflow...

Could your team answer this today, and how long would it take? "How many of the customers who... - Signed up through our holiday marketing campaign in Q4 - On the annual family plan - paying with a Chase-issued card Had their first renewal attempt declined for ‘insufficient funds’, ? And of those... - How many were successfully recovered within 7 days - Versus how many ended up churning within 30 days?" An organization that can answer these types of questions quickly and accurately can react quickly, which results in happier customers and satisfied executives. At its core, payments optimization is only as good as the join between internal business data and transaction data. PSPs, networks, and banks give you a rich stream of authorization, settlement, and dispute data, but merchants must bring their own customer and product context in order to produce actionable insights that are relevant to their business. To make this data usable, merchants need to normalize a few key dimensions: A. Customer-level data - Unique customer identifier that persists across systems (CRM, billing, payments, support). - Cohort tags: acquisition channel, geography, subscription tier, customer tenure, annual vs monthly - Payment Event stage: Card verification, free-to-paid trial conversion, non-trial conversion, renewal (1 to N) B. Subscription contract data - Plan/tier ID (standardized across billing and payments). - Start and end dates, renewal frequency, billing currency. - Status flags: active, paused, canceled (with standardized cancel reasons). C. Invoice/order-level data - Invoice ID (must be consistently mapped to payment transaction IDs). - Line items (tier, add-ons, discounts, tax). - Net and gross amounts, including refund/credit adjustments. D. Payment transaction data - Transaction ID (gateway/PSP ID). - Decline reason codes (normalized across acquirers/networks). - Payment method type, issuer BIN, network tokenization status. - Success/settlement status, fraud signals, retries, dispute lifecycle. E. Linking logic - A clean key structure: customer_id → subscription_id → invoice_id → transaction_id. - Consistent timestamping (UTC, ISO8601, normalized across systems). - Master data management (e.g., ensuring “Invoice 12345” in billing = “Payment 12345” in PSP). When structured this way, you can slice performance by cohort (e.g., “Trial-to-paid in LatAm declines more often on first attempt”), payment method (e.g., “BIN ranges in Southeast Asia fail more often at renewal”), or customer lifecycle (e.g., “Annual renewals have higher closed account declines than monthly”). Most merchants I meet with can't answer even a fraction of that first question. Not because the data doesn't exist, but because it isn't structured. If this sounds familiar, let's talk.

🚨 𝐇𝐨𝐰 𝐭𝐨 𝐔𝐧𝐥𝐨𝐜𝐤 𝐕𝐚𝐥𝐮𝐞 𝐟𝐫𝐨𝐦 𝐏𝐚𝐲𝐦𝐞𝐧𝐭𝐬 𝐃𝐚𝐭𝐚 — 𝐛𝐲 DEUNA 👇 As commerce evolves into an intelligent, agent-driven ecosystem, payments data is emerging as a critical strategic asset — but one that remains vastly underutilized. This post outlines the key limitations holding teams back — and how leading merchants are transforming data into a competitive advantage. — 𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 𝐏𝐚𝐲𝐦𝐞𝐧𝐭𝐬 𝐃𝐚𝐭𝐚 Today’s commerce stack generates rich and interconnected data from multiple systems: → Payments Data: transaction metadata, approval rates, methods, issuer response → Orders Data: cart value, SKUs, timestamps, discounts → Identity Data: device fingerprinting, customer behavior, fraud signals → Other Sources: CRM, logistics, loyalty, geography, and more — 𝟑 𝐒𝐲𝐬𝐭𝐞𝐦𝐢𝐜 𝐁𝐥𝐨𝐜𝐤𝐞𝐫𝐬 𝐓𝐨 𝐕𝐚𝐥𝐮𝐞 𝐑𝐞𝐚𝐥𝐢𝐳𝐚𝐭𝐢𝐨𝐧 1️⃣ Fragmented & Inconsistent Data → 80% of data teams’ time is spent cleaning instead of analyzing → Siloed PSPs, fraud tools, and BI platforms prevent unified decision-making 2️⃣ Context-less AI doesn't deliver  → 90% of AI projects fail to operationalize → Without business context, generic AI outputs remain disconnected from strategic goals 3️⃣ Constant Firefighting, Minimal Strategy → Operational teams are stuck responding to incidents instead of driving outcomes — 𝐇𝐨𝐰 𝐋𝐞𝐚𝐝𝐢𝐧𝐠 𝐓𝐞𝐚𝐦𝐬 𝐔𝐧𝐥𝐨𝐜𝐤 𝐕𝐚𝐥𝐮𝐞 The process to transform raw payments data into actionable intelligence: 1️⃣ Data Ingestion → Consolidating data streams from PSPs, checkout systems, CRMs, fraud tools, and commerce platforms into a central layer 2️⃣ Variable Definition & Structuring → Translating raw fields into business metrics (e.g., GMV, approval rate, ARPU) 3️⃣ Real-Time Monitoring & Alerts → Establishing live observability for key indicators (e.g., spikes in declines, fraud signals, latency issues) 4️⃣ Cleaning, Deduplication & Standardization → Removing duplicates, normalizing formats (e.g., date, currency, identifiers) 5️⃣ Industry-Specific Contextualization → Applying domain knowledge (e.g., travel refunds, retail loyalty...) to reclassify and enrich data 6️⃣ Addition of Exogenous Variables → Integrating external factors such as FX volatility, holidays, weather, or marketing campaigns to reveal hidden correlations and refine models. 7️⃣ AI Implementation → Using ML and intelligent agents to detect anomalies, identify high-impact opportunities, and automate routing, retries... in real time. — 𝐂𝐨𝐧𝐜𝐥𝐮𝐬𝐢𝐨𝐧: 𝟑 𝐏𝐫𝐢𝐧𝐜𝐢𝐩𝐥𝐞𝐬 𝐃𝐞𝐟𝐢𝐧𝐢𝐧𝐠 𝐓𝐡𝐞 𝐍𝐞𝐰 𝐒𝐭𝐚𝐧𝐝𝐚𝐫𝐝 1️⃣ Unified Data Foundation 2️⃣ Context-Aware Opportunity Detection 3️⃣ Real-Time Execution on Intelligence → Automate retry logic, fraud triggers, routing adjustments, and more — Source: DEUNA ► Subscribe to 𝐓𝐡𝐞 𝐏𝐚𝐲𝐦𝐞𝐧𝐭𝐬 𝐁𝐫𝐞𝐰𝐬: https://lnkd.in/g5cDhnjC ► Connecting the dots in payments... | Marcel van Oost

💻 Using Python to Detect Duplicate Payments — A Simple but Powerful Audit Tool 💡 As internal auditors and risk professionals, we often review large volumes of payment data for anomalies. One common (and costly) control gap? Duplicate payments. (Its a nightmare for inhouse internal audit team.) This can be automated using Python — and the results were immediate. 🔍 How it can be done Write a Python script to: ✅ Read ledger/exported payment data (from SAP, Oracle, Tally, etc.) ✅ Clean and normalize key fields (Invoice No., Vendor, Amount, Date) ✅ Flag records with identical or near-identical fields ✅ Highlight suspicious duplicates (same invoice paid twice, or minor reference differences with same amount/vendor/date) 📈 Results A list of suspected duplicate payments across multiple entities Identified control gaps in manual validation and workflow bypass This will help in recovering overpaid amounts and tighten the review process 🧠 Why This Matters ✅ Manual review can’t scale ✅ ERPs may not always block duplicates — especially if entries are slightly altered ✅ Python is flexible, fast, and easy to integrate into monthly audits or exception reports Its equally important to note that in a good environment, the ratio of confirmed to suspected duplicate payment can be as high as 1:1000 (1 transaction out of 1000 suspected transaction.) 👁️ Audit isn’t just about hindsight — it’s about real-time insight. Are you using analytics to enhance your internal controls? Let’s talk automation, fraud detection, and data-driven auditing. #InternalAudit #DataAnalytics #Python #DuplicatePayments #ControlTesting #RiskManagement #FinanceAudit #FraudPrevention #Automation #AccountsPayable #AuditTech

When our customers add more PSPs, migrate to new one, adjust their routing logic, or just reconfigure their setup, they trust Pagos to take on the data aggregation, research, and monitoring work. With Pagos in their corner, this once difficult work is now simple and repeatable! What we've learnt from assisting them is that once the data is harmonized, we see the following steps as critical to any migration journey: Step 1: Determine what metrics you're most concerned about for each processor. This obviously includes cost and approval rate, but can also include contractual obligations, SLA/uptime, fraud tools (and their efficency), chargeback dispute tools and so much more. . Step 2: Establish baseline metrics over time for each processor. With Peacock by Pagos, we ingest your payments data from every processor and normalize it, thereby allowing you to compare processors with different settings and functionality side by side. Step 3: If one processor demonstrates lagging performance, reassess your relationship with them - for parts of your business or all! It's entirely possible you set up a new PSP using configuration settings typical of your more established PSP; since each processor has their own unique way of handling payments, there may be opportunities for improvement purely in your setup. Step 4: Now that you have an easy way to compare key performance indicators for multiple PSPs in Peacock, you can monitor data over time for any unexpected changes which is also important in order to be able to hold your partners accountable. You can even run A/B tests for migrating specific customer segments to different PSPs. Step 5: Payments optimization! They could do all this without having to rip out their current infrastructure as Pagos does not require you to change your current processing flow as we can help maximize your performance without having to actually process your transactions..With our no-code integrations to most major PSPs, you could be live in minutes. Having said all that, we have also found that a lot of the opportunities to do better sits on the merchants side. What is Your performance? Have you optimize that part of the equitation? With detailed data metrics? If not, how do you know if your current partners are your problem and how would you determine success with a new partner..? It should also be noted that other types of partners like payment orchestration platforms (or gateways), fraud providers, subscription management platforms also impact your performance with your PSPs.

Stripe’s real-time billing analytics powered by Apache Pinot! Fascinating blog from the Stripe team about how they built a powerful billing analytics tool that enables customers to accurately track key metrics such as growth, churn, and trial conversion in near real time. This means Stripe users can react quickly to evolving trends with a personalized dashboard that’s always fresh — a game-changer for planning and decision-making. At the heart of this solution is Apache Pinot, which provides: * A highly concurrent multi-stage query engine supporting complex joins and window functions. * Seamless ingestion from Kafka (real-time) and batch data, with automatic query federation and supports ease of data backfill/correction. * A powerful execution engine that delivers predictable SLAs at scale. Read the full blog here: https://lnkd.in/gZsjA8gQ Pinot's multi-stage engine has come a long way thanks to the tireless efforts of the Pinot community including Xiaotian (Jackie) Jiang Gonzalo Ortiz Jaureguizar Xiang Fu Yash Mayya Neha Pawar Kishore Gopalakrishna Huge congratulations to the Stripe team for building such an elegant, customer-focused solution! 🚀