Monitoring command center
Mission setup
Customer Income Shift
A marketing campaign attracted a wealthier customer segment.
Score
0
Detection quality
Time
03:00
Mission clock
Budget
100
Analysis credits
Alerts
0
Current warnings
Model health
96%
Estimated stability
Period
1
Production window
Live visual simulation
Feature streams and detector stations
Detector controls
Thresholds and windows
SensitiveConservative
PSI compares how observations move between distribution bins. Larger values indicate stronger population shift.
Custom drift generator
Shape the production distribution
Feature investigation
Ranked monitoring table
| Flag | Feature | Type | Train mean / mode | Production mean / mode | Missing | Drift score | Importance | Severity | Priority |
|---|
Distribution comparison
Training versus production
Risk prioritization
Feature drift scoreboard
Historical monitoring
Drift and model performance trend
Feature relationships
Correlation change heatmap
Root-cause investigation
What most likely happened?
Corrective response
Choose the next action
Actions consume budget differently. Strong responses reduce model risk, while unnecessary retraining can lower efficiency.
Learning feedback
Analyst briefing
Analyze the current window, flag suspected features, then diagnose the cause and choose a response.
Mission completion0%
Session analytics
Detection performance
Precision
—
Recall
—
False alerts
0
Best score
0
Achievements
Analyst milestones
Educational reference
Metric guide and response principles
Drift does not always mean failure
A feature may move without damaging predictions. Combine distribution evidence with feature importance, model metrics, and operational context.
Small important shifts can matter
A modest change in a critical feature can outrank a large change in an unused feature. Risk weighting prevents noisy prioritization.
Investigate before retraining
Pipeline defects, schema changes, and missing values often require data repairs. Retraining on corrupted data can make the model worse.