Feature Drift

A change in the distribution of the independent input variables P(X) while the conditional relationship P(Y|X) remains constant.

• Example: A credit model trained on applicants with incomes between $30k-$80k suddenly receives a surge of applications with incomes over $150k.
• Impact: The model is now extrapolating to an unfamiliar region of the feature space, leading to unreliable predictions.
• Detection: Monitored via univariate statistical distance metrics like the Population Stability Index (PSI) or Kullback-Leibler divergence on individual features.

A change in the distribution of the target variable P(Y) without a change in the likelihood P(X|Y).

• Example: A fraud detection model trained on a 2% fraud rate suddenly encounters a coordinated attack pushing the real-world fraud rate to 15%.
• Impact: A model calibrated for a specific base rate will produce miscalibrated probability estimates, breaking decision thresholds.
• Mitigation: Requires adjusting classification thresholds or retraining with inverse probability weighting to account for the new class balance.

A fundamental change in the relationship between the input features and the target variable, meaning P(Y|X) itself has changed.

• Example: A product recommendation model finds that the feature 'time_on_page' historically correlated with purchase intent. A new UI redesign makes users scroll longer without buying, breaking the historical pattern.
• Impact: This is the most dangerous form of drift as it directly invalidates the model's learned decision boundary.
• Detection: Requires analyzing model performance metrics (accuracy, F1-score) against a delayed ground truth source, often using tools like NannyML or Evidently AI.

A gradual or seasonal evolution of feature distributions strictly tied to the passage of time, often cyclical in nature.

• Example: A retail demand forecasting model sees 'purchase_frequency' spike predictably every November. If the model was trained only on summer data, it experiences severe drift in Q4.
• Characteristics: Unlike sudden covariate shift, temporal drift is often predictable and can be engineered away by including explicit time-based features (day of week, month) or Fourier transforms in the feature engineering stage.
• Strategy: Implementing a sliding window retraining cadence that captures full seasonal cycles prevents this degradation.

A sudden change in the statistical properties of features caused by upstream pipeline failures, not organic environmental changes.

• Example: A logging bug causes the 'session_duration' feature to be recorded as negative integers, or a broken sensor reports a constant null value.
• Impact: This breaks the feature serving contract and causes immediate prediction errors, often manifesting as a spike in missing values or out-of-range bounds.
• Detection: Automated feature validation suites that check for schema adherence, null percentage thresholds, and min/max range violations before data enters the online store.

A specific type of artificial drift caused by a discrepancy between the feature engineering code path used during training and the path used during online inference.

• Example: The training pipeline normalizes 'price' using the mean of the entire historical batch, but the online store normalizes it using a rolling 24-hour window, producing different values for the same raw input.
• Impact: The model receives feature vectors in production that are systematically different from those seen during training, even if the raw data is identical.
• Prevention: Enforced by using a shared Feature View and Serving API that guarantees point-in-time correctness and identical transformation logic in both environments.