Monitoring payment performance beyond global averages

Stripe handles billions of dollars in payment volume daily, with spikes during major shopping events—businesses processed more than $31 billion on Stripe over Black Friday and Cyber Monday in 2024. Tracking aggregate performance across all payments would provide a broad overview, but it would likely hide degradations affecting specific traffic segments. For example, if a UK card issuer starts declining recurring payments on prepaid cards at high rates, that spike might not move global metrics even though businesses relying heavily on prepaid cards would feel an immediate impact.

To address this, Stripe built a system that provides near real-time visibility into the performance of subsets, or "slices," of payment traffic. The system combines machine learning (ML) and time series algorithms to detect performance degradations across metrics such as payment success rates, authentication rates, costs, and fraud. When a degradation is detected, the system automatically alerts the relevant experts to investigate. Building this capability required solving three problems: defining a slice of payment traffic, detecting when a slice is degrading, and triggering an effective response.

Defining slices in a high-dimensional space

Payment monitoring operates in a space of over 16,000 variables, including more than 10,000 issuing banks, hundreds of currencies, countries, card products, and payment features such as Apple Pay, mail or telephone orders, and account funding transfers. Performance issues can arise from unique combinations of these factors—for instance, a spike in failures on digital wallet payments from French debit cards.

Slice definitions must balance two competing needs: narrow enough to isolate the specific shape of a degradation, yet data-rich enough for high statistical confidence. Stripe refines slice definitions by:

  • Integrating new functionalities (such as advanced card features like multicapture or partial authorization) into the monitoring system as they become supported.
  • Incorporating insights from historical incidents—if a past degradation went undetected due to monitoring limitations, a new slice is established to track the relevant variables going forward.
  • Basing slices on an understanding of how API call information propagates through the broader financial environment, which requires collaboration with the engineering teams responsible for financial infrastructure.

Slice monitoring is built as a platform that lets engineering teams develop their own metrics and slices while reusing core monitoring algorithms and operational tooling.

Detection without a stable baseline

Once slices are defined, the next challenge is accurately identifying degradation within them. Payment metrics fluctuate dramatically for legitimate reasons, making simple anomaly detection insufficient. Standard time-series approaches compare current values against a historical baseline, but payment slices lack a stable one. Customer onboarding, fraud trends, and changes in business behavior all cause underlying variation. Launching a large free trial or a new product line can alter customer composition and therefore payment success rates. An algorithm that ignores these factors would trigger false positives and undermine the system's usefulness.

Stripe controls for underlying changes in transaction composition by combining its ML models with time-series analysis. First, ML models estimate the probability of success for every transaction in the monitoring dataset—that is, the expected outcome. These models are trained on transaction-level datasets. Then, near real-time time-series anomaly detection is performed, adjusted for the underlying probability of success.

Deciding when to act

Accurately detecting degradations is only the first step. Even with highly precise anomaly detection, monitoring tens of thousands of slices inherently produces false positives from random fluctuation. The system needs to enable rapid detection while avoiding unnecessary alarms from transient performance dips.

Stripe uses a finite state machine that aggregates losses over time, triggering alerts only when loss thresholds from sustained events are breached. Alerts are classified by urgency—derived from the rate of volume loss—and by inferred root cause, which streamlines routing to the appropriate team for investigation and remediation.

Extending the platform to users

The slice monitoring platform identifies real degradations in payment performance each day with precision exceeding 90%, allowing strong coverage without an unsustainable operational burden from false positives.

Later this year, Stripe plans to make slice monitoring alerts available directly to select users. All users will gain full visibility into payment success rates in the Payments analytics page of the Stripe Dashboard, along with recommendations for performance optimization. Stripe will also discuss this work at Stripe Sessions 2025.