Insights
Why Your Fraud Rate Keeps Changing Weeks After You Measure It
Stripe data shows only 41% of fraud disputes land by day 14 and 86% by day 60. Learn why early fraud rates mislead and what to ask before trusting a vendor test.
November 3, 2026
6 min readThe Number You're Trusting Is Incomplete
Your fraud rate from last month probably looks reasonable. It's also probably wrong.
Not because the math is bad, but because most of the fraud hasn't been reported yet. Chargebacks and fraud alerts trickle in over weeks and months after a transaction. When you pull your fraud rate at the two-week mark, you're looking at a partial picture and treating it as the full one.
How Fraud Labels Actually Arrive
When a cardholder's card is used fraudulently, it can take weeks before they notice, contact their bank, and trigger a dispute. The issuer then files a chargeback or sends a fraud notification (called a TC40 on Visa's network or a SAFE report on Mastercard's). Each step adds delay.
Stripe published data on this in an October 2026 engineering blog post titled "Scaling Radar: Fighting fraud with agent-assisted machine learning," by engineers Ryan Drapeau and Sai Samant. A cumulative distribution chart in that post shows the share of fraudulent disputes reported over time across the Stripe network: roughly 41% by day 14, 66% by day 30, and 86% by day 60 after the payment.
At the two-week mark, only 41% of the fraudulent disputes Stripe observed across its network had been reported. At 30 days, about a third of those disputes still hadn't arrived. And even at day 60, 14% of the total was still outstanding.
Your own curve will look different. Label lag varies by merchant category, geography, and fraud type. A subscription SaaS company with mostly domestic cards will typically see a different reporting shape than a cross-border luxury retailer, for instance. But the direction is universal: early numbers undercount fraud.
Why the Lag Is Structural
The delay isn't a quirk of any one processor. It's built into how card networks operate.
Visa and Mastercard both give cardholders up to 120 days from the transaction date to file most fraud-related chargebacks, according to industry publications like Chargeback Gurus and Clear.sale (summarizing network rules, not the official rulebooks).
Beyond formal chargebacks, issuer-side fraud reports add their own lag. Visa TC40 claims typically release within days of an initial fraud report, but the chargeback itself may still be weeks away. Mastercard issuers can wait up to 60 days after a chargeback to submit a SAFE report.
Then there's friendly fraud. A customer forgets about a subscription renewal, doesn't recognize a billing descriptor, or simply changes their mind. Cardholders have up to 120 days to file a dispute, and disputes frequently arrive weeks or months after the transaction. Visa estimates friendly fraud accounts for up to 75% of all chargebacks, as reported by Chargebacks911.
What Immature Labels Do to Model Evaluation
This isn't just an accounting problem. It distorts how you evaluate fraud tools.
Consider what happened inside Stripe's own fraud team. In the same blog post, Drapeau and Samant described a telling episode. Stripe removed downsampling from its Shield fraud model's training data, expanding the dataset from roughly 800 million to 11 billion transactions. The larger model looked better in offline tests and early production results.
Then, as more fraud reports arrived over subsequent weeks, the model's measured performance fell below the earlier model's. Stripe rolled it back.
The reason: removing downsampling added about 20 times more fraudulent transactions that were still labeled as legitimate (because their fraud reports hadn't arrived yet). The model learned that these patterns were safe. The most recent weeks of training data, which contained the freshest fraud signals, were also the weeks with the most immature labels.
This wasn't a failure of engineering. It was a careful team catching the problem and disclosing it publicly. But it illustrates the trap: a model trained or evaluated on immature labels will look better than it actually is, because the fraud it misses hasn't been counted yet.
A Worked Example (Illustrative Numbers)
To make this concrete, here's a hypothetical. These numbers are illustrative, not sourced from any specific merchant or platform.
Imagine you process 100,000 transactions in October. On November 1, your dashboard shows 200 chargebacks: a 0.20% fraud rate. You're comfortable. Maybe you're comparing two fraud vendors and Vendor A shows 0.18% while Vendor B shows 0.22%.
By January 1, an additional 340 chargebacks have arrived for those same October transactions. Your actual fraud rate is now 0.54%, more than double the November figure.
That 0.04% gap between Vendor A and Vendor B? It's noise. Both numbers were measured on a label set that was less than half complete. The vendor that looked "better" in November may simply have been measured at a slightly earlier point on the same incomplete curve.
This is why the VAMP timing angle matters too. If you're tracking your dispute ratio for Visa's VAMP program, the same lag applies: November's fraud can hit your Q1 ratio when your transaction volume has dropped. (For more on that timing mismatch, see Why Black Friday Disputes Hit Your VAMP Ratio in February.)
What to Ask Before Trusting a Fraud Rate
Whether you're evaluating your own numbers or a vendor's test results, five questions separate a reliable measurement from a misleading one:
What is the observation window? How many days after a transaction does the measurement wait before counting results? Anything under 30 days is suspect. Under 60 is incomplete.
What is the label freeze day? At what point does the evaluator stop accepting new fraud reports for a given cohort? If there's no freeze day, the number is a moving target.
How are the most recent weeks treated? The freshest data has the most immature labels. A good evaluation either excludes recent weeks or weights them accordingly.
What is the holdout method? A proper test uses a holdout set where both the old and new approaches are evaluated on the same transactions with the same label maturity. If the comparison uses different time periods, label maturity differences can explain the entire gap.
Are cohorts compared at the same label age? A fraud rate from 90-day-old transactions and a fraud rate from 14-day-old transactions aren't comparable, even if they cover the same merchant.
Measure What's Real, Not What's Fast
Fraud numbers feel authoritative because they come from your processor's dashboard with decimal-point precision. But precision isn't accuracy. A number can be exactly calculated and still be fundamentally incomplete.
The fix is patience and process: evaluate on mature cohorts, set a label freeze day, and compare apples to apples. When a vendor shows you results from the last two weeks, you now know what question to ask.
If you want to see how your own fraud labels mature over time, and where your current numbers might be understating reality, Book a Demo →.
Sources
Stripe, "Scaling Radar: Fighting fraud with agent-assisted machine learning" by Ryan Drapeau and Sai Samant (October 19, 2026): https://stripe.dev/blog/scaling-radar-fighting-fraud-with-agent-assisted-machine-learning
Chargeback Gurus, "A Merchant's Guide to Chargeback Time Limits" (July 28, 2026): https://www.chargebackgurus.com/blog/dont-run-out-the-clock-understanding-chargeback-time-limits
Clear.sale, "Chargeback time limits of major credit card issuers" (July 9, 2026): https://www.clear.sale/blog/what-is-the-time-limit-on-chargebacks
Chargeflow, "Friendly Fraud: Causes, Evidence and Merchant Prevention" (October 10, 2026): https://www.chargeflow.io/blog/everything-you-should-know-about-friendly-fraud
Chargebacks911, "Friendly Fraud: The 2026 Guide to Stop Chargeback Abuse" (July 21, 2026): https://chargebacks911.com/friendly-fraud/
ChargebackStop, "TC40s & SAFEs: Notifications of Fraud Explained for Merchants": https://www.chargebackstop.com/blog/tc40s-safes-fraud-notifications-merchants
Verifi, "What is TC40 data?" (May 11, 2016): https://www.verifi.com/chargebacks-disputes-faq/what-is-tc40-data.html
Clear.sale, "The Chargeback Lag: What It Is, and How It Impacts Your E-Commerce Business" (January 10, 2024): https://www.clear.sale/blog/the-chargeback-lag-what-it-is-and-how-it-impacts-your-e-commerce-business
