Skip to main content

Drift

Drift compares each feature's current window with its reference. The reference is set per model version (see Baselines); the length of the current window is set in Monitoring settings.

Data drift​

One row per feature:

Figure 1. Data drift on every feature, and prediction drift on the model output.

ColumnMeaning
FeatureThe feature name from the model's schema.
KindNumeric or Categorical.
PSIPopulation stability index. This is the value judged against your drift limits.
KS D and KS pKolmogorov-Smirnov test statistic and p-value. Numeric features only.
JS distanceJensen-Shannon distance. Categorical features only.
Values (ref / now)How many values the reference and current windows held.
StatusStable, Moderate or Severe.

Prediction drift​

The same tests over the model's prediction or predicted label, in one row named Prediction.

How the status is set​

Only PSI sets the status. With the default limits:

PSIStatus
Below 0.1Stable
0.1 or moreModerate
0.25 or moreSevere

Change the limits for a model with Moderate drift (PSI) and Severe drift (PSI) in Monitoring settings. Each drifting feature, and drifting predictions, open a finding on the model page.