Model monitoring
What is model monitoring?
Model monitoring is the ongoing practice of tracking a deployed model's predictions, inputs, and performance to catch degradation from data drift, concept drift, or pipeline failures before it affects business outcomes.
Why model monitoring matters
A model's behavior reflects the world it was trained on. When production data drifts from that picture, the way a maze changes shape after a mouse has learned it, accuracy erodes even though nothing about the model's code changed. Under SR 26-2, monitoring cadence is tied to a model's materiality and change velocity rather than a fixed annual cycle, so an undetected drift event can become a compliance gap in addition to a performance one.
Types of model monitoring
Type
What it catches
Data (input) drift
Input statistics shifting away from training distribution
Concept drift
The input-output relationship changing even when inputs look normal
Feature monitoring
Individual feature distributions or missingness changing
Prediction monitoring
The model's own output distribution shifting
Fairness monitoring
Performance gaps emerging across protected groups
Operational monitoring
Uptime, latency, and throughput
Data drift and concept drift are the two most common causes of what's broadly called model drift.
How model monitoring works
- Collect predictions and inputs in production.
- Compare them against training-time baselines.
- Detect drift or degradation against a threshold.
- Alert the team responsible for the model.
- Retrain or roll back if the alert is confirmed.
For example, a lending model that starts seeing loan applications shift toward a demographic underrepresented in its training data would show up as input drift well before its approval rate visibly changes.
Industry use cases
Financial services
SR 26-2 names back-testing, outlier analysis, and prediction-versus-actual comparisons as accepted methods for the outcomes-analysis component of ongoing model validation, so a bank's monitoring setup has to produce evidence in that specific form, not just an aggregate accuracy dashboard.
Life sciences
GxP validation requires a documented revalidation trigger. A monitoring alert on a clinical or manufacturing model isn't just an engineering signal there, it's the event that formally kicks off a revalidation record.
Public sector
Agencies deploying models for benefits eligibility or fraud detection typically face public records and appeal requirements, which means monitoring has to produce an audit trail a caseworker or outside reviewer can actually read.
FAQ
What's the difference between model monitoring and model evaluation?
Model evaluation measures a model's performance at a point in time, usually before deployment. Model monitoring tracks that performance continuously after deployment to catch when it changes.
How often should a model be monitored and retrained?
Monitoring itself should run continuously where possible. How often you act on it depends on the model's materiality and how fast its data changes, not a fixed annual cycle. Retraining specifically gets triggered by sustained drift past an agreed threshold, not a calendar date.
What tools are used for model monitoring?
Purpose-built monitoring software such as Domino Model Monitor tracks prediction distributions, input drift, and concept drift automatically, and alerts when a model has degraded enough to warrant expert review, distinct from the uptime/response-time monitoring typical software already gets. Every statistically based monitoring system produces some false alarms (type-I errors) and misses some real issues (type-II errors), which is why tuning that balance is a shared responsibility between data science and IT, not something to automate away.
Is model monitoring part of MLOps?
Yes. It's the MLOps stage that closes the loop between deployment and retraining.