Two-sample KS test
Compares the reference and current distributions for numerical features.
Feature-level statistical monitoring
Detect Data Drift Before It Affects Your Models.
DataDrift Monitor compares reference and current datasets with statistical tests to identify feature-level distribution changes and help investigate what changed.
- Reference
- Baseline dataset
- Current
- New observations
- Output
- Investigable findings
Distribution comparison
Statistical viewReference dataCurrent data
Compare distributions, then investigate the feature-level result.
Workflow
A clear path from comparison to investigation.
Each step keeps the comparison grounded in the uploaded datasets and documented statistical methods.
- 01
Upload datasets
Provide a reference CSV and a current CSV with matching features.
- 02
Run statistical tests
Numerical and categorical columns are analyzed with their respective tests.
- 03
Review findings
Identify results that remain significant after Bonferroni correction.
- 04
Investigate change
Inspect severity, distribution changes, categories, and relationships.
- 05
Export results
Download the full CSV results or a self-contained HTML report.
What data drift looks like
A change in the observed distribution.
Drift is assessed by comparing the feature distributions in a reference dataset and a current dataset. A visual shift is informative, but the dashboard uses statistical tests to evaluate the comparison.
ReferenceCurrentShift
Illustrative only. The dashboard reports the corresponding feature-level test result.
Statistical methods
Tests chosen for the feature type.
Chi-square test
Compares category frequencies for categorical features.
Cramér's V
Reports the categorical effect size alongside the chi-square comparison.
Bonferroni correction
Adjusts p-values across the set of tested common features.
Phase 2 investigation
Move beyond a test result.
Use the existing dashboard to focus attention on observed changes while keeping statistical magnitude separate from business or model impact.
Drift severity
Describes the magnitude of a statistical distribution difference. It is not business impact, model risk, financial impact, or predicted model performance.
What Changed?
Summarizes numerical median movement or the largest categorical proportion change.
Category changes
Shows new and removed categorical values alongside their proportion changes.
Relationship analysis
Reports substantial Pearson-correlation changes among numerical features. Association is not causation.
Portable results
Export feature-level findings as CSV or a self-contained HTML report.
Supplied sample dataset
A real local example, not a product metric.
The repository's supplied reference.csv and current.csv produce four significant-drift results and one stable result at the default alpha setting.
- 4
- Significant drift
- 1
- Stable
- 0.05
- Default alpha
Dashboard preview
Views from the working dashboard.
Reference, investigation, and feature-level views from DataDrift Monitor.
DataDrift MonitorDashboard views
Methodology
Interpret results in context.
DataDrift Monitor compares common features between a reference and current dataset. Numerical missing values are excluded for the KS test and their percentages are reported. Categorical missing values are handled as an explicit [MISSING] category.
P-values indicate evidence against the corresponding null hypothesis, not the cause or practical consequences of a change. Bonferroni-adjusted p-values support multiple feature comparisons.
Ready to investigate
Explore your data.
Run a reference-versus-current comparison in the existing dashboard.