Dataset Drift Detector
Compare a reference dataset against a current dataset to detect distribution shift in each feature. Uses the KS Test (statistical significance) and PSI (Population Stability Index) — the industry standard for monitoring model inputs.
How it works: numeric features are binned using the reference distribution. PSI and KS scores are computed per feature, then aggregated into an overall stability score. PSI < 0.10 = stable · 0.10–0.20 = moderate · ≥ 0.20 = significant drift.
Reference dataset (baseline / training)
Drop reference CSV or click
Historical / training data
Current dataset (new / production)
Drop current CSV or click
Recent / production data
More bins = finer granularity; requires larger datasets.
Analysing distributions…
Computing KS tests and PSI scores for each feature
Feature Drift Summary
Sorted by drift severity. Click any row to view its distribution chart.
| Feature | Type | PSI | PSI Severity | KS Stat | KS p-value | KS Drift |
|---|
Distribution Comparison
Reference Current
Numeric Feature Statistics
Mean, std, and median for reference vs current. Red = >20% shift, amber = >5% shift.
| Feature | Ref Mean | Cur Mean | Ref Std | Cur Std | Ref Median | Cur Median |
|---|
Continue your pre-training checklist
PSI thresholds (0.1 / 0.2) are industry conventions originating from credit risk monitoring. Appropriate thresholds vary by domain and sample size. KS p-values are sensitive to sample size — large datasets may show statistically significant but practically irrelevant drift. Always interpret results in context.
What you get
Stability score
A 0–100 score based on the proportion and severity of drifted features, with a colour-coded verdict.
PSI per feature
Population Stability Index for every column — numeric features use quantile bins, categoricals use frequency bins.
KS test results
Two-sample Kolmogorov-Smirnov statistic and p-value for numeric features. Flags shift at p < 0.05.
Distribution charts
Overlay histogram (numeric) or grouped bar chart (categorical) comparing reference vs current for each feature.