Incident management · Production

Stripe uses ML and time-series anomaly detection to monitor payment performance across 16,000+ slice dimensions

The problem

Aggregate payment monitoring masked degradations affecting specific traffic segments—a spike in failures for a particular card type or region might not move global metrics even as individual businesses felt acute impact.

First attempt

Standard time-series anomaly detection was insufficient because payment metrics lack a stable baseline—customer onboarding, fraud trends, and business behavior changes create underlying variation that would cause false positives.

Workflow diagram · grounded in source
1
Continuous slice monitoring
trigger
“We monitor payments in a high-dimensional space characterized by over 16,000 payment-related variables”
2
ML probability estimation
ai_action
“we leverage ML models to estimate the probability of success for every transaction in our monitoring dataset (i.e., the expected outcome). These models are trained on Stripe's vast transaction-level datasets”
3
Time-series anomaly detection
ai_action
“we conduct near real-time, time-series anomaly detection, adjusting for the underlying probability of success”
4
Loss threshold gating
validation
“we use a finite state machine that aggregates losses over time, only triggering alerts when loss thresholds from sustained events are breached”
5
Alert classification and routing
routing
“Alerts are classified based on urgency—derived from the rate of volume loss—and inferred root cause, streamlining the routing to the appropriate team for investigation and remediation”
6
Expert investigation
human_review
“it automatically alerts the relevant experts at Stripe to investigate and resolve the underlying issue”
Reported outcome

The slice monitoring platform identifies real payment performance degradations each day with precision exceeding 90%, achieving excellent coverage without generating unsustainable operational burden from false positives.

Reported metrics
Slice monitoring detection precisionexceeding 90%
Payment-related monitoring dimensionsover 16,000
payment volume over Black Friday and Cyber Monday 2024more than $31 billion
Reported stack
ML modelstime series algorithmsfinite state machine
Source
https://stripe.com/blog/using-ml-to-detect-and-respond-to-performance-degradations-in-slices-of-stripe-payments
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The slice monitoring platform identifies real payment performance degradations each day with precision exceeding 90%, achieving excellent coverage without generating unsustainable operational burden from false positives.

What tools did this team use?

ML models, time series algorithms, finite state machine.

What results were reported?

Slice monitoring detection precision: exceeding 90%; Payment-related monitoring dimensions: over 16,000; payment volume over Black Friday and Cyber Monday 2024: more than $31 billion (source-reported, not independently verified).

What failed first in this deployment?

Standard time-series anomaly detection was insufficient because payment metrics lack a stable baseline—customer onboarding, fraud trends, and business behavior changes create underlying variation that would cause fals…

How is this incident management AI workflow structured?

Continuous slice monitoring → ML probability estimation → Time-series anomaly detection → Loss threshold gating → Alert classification and routing → Expert investigation.

Related incident management cases