Data Drift

The most common form of data drift, where the distribution of input features P(X) changes, but the conditional relationship P(Y|X) remains stable.

• Example: A volatility forecasting model trained on a low-VIX regime suddenly receives data from a high-volatility crisis period.
• Impact: The model extrapolates into regions of the feature space where it has no statistical support, producing high-variance predictions.
• Detection: Use the Kolmogorov-Smirnov test or Population Stability Index (PSI) to compare training and production feature distributions.

Occurs when the distribution of the target variable P(Y) changes independently of the input features. This is common in classification problems with evolving base rates.

• Example: A corporate default prediction model encounters a sudden recession, shifting the default rate from 2% to 15%.
• Nuance: The model's conditional probabilities P(Y|X) may remain perfectly calibrated, but the raw output distribution becomes mismatched with reality.
• Mitigation: Apply prevalence adjustment or recalibrate decision thresholds against recent label data.

A subtle degradation where the relative importance of input features shifts over time, even if univariate distributions appear stable.

• Mechanism: Correlations between features change, altering the geometry of the input manifold that the model learned during training.
• Example: In a multi-factor equity model, the correlation between momentum and value factors inverts during a market regime change, breaking the model's linear assumptions.
• Detection: Monitor SHAP value stability and track the rank-order of feature importance scores across time windows.

Drift that manifests as a function of time itself, often due to gradual structural changes in the data-generating process rather than sudden shocks.

• Characteristics: Slow, monotonic trends in feature means or variances that accumulate over weeks or months.
• Example: The average daily trading volume of an asset class steadily increases over years due to market electronification, causing a volume-normalization layer to become miscalibrated.
• Detection: Deploy CUSUM (Cumulative Sum) control charts to identify small, persistent shifts that t-tests might miss.

A pipeline-specific phenomenon where drift is introduced not by the external world, but by changes in upstream data engineering logic.

• Root Cause: Modifications to feature definitions, aggregation windows, or imputation strategies in the feature store create a distributional mismatch with the training set.
• Example: A data engineer updates a missing-value imputation from median to mean, shifting the entire feature distribution without any real-world change.
• Prevention: Implement data lineage tracking and schema enforcement with automated distribution checks on every feature pipeline commit.

A deliberate, malicious form of data drift where an external actor manipulates input data to degrade model performance or force specific erroneous outputs.

• Mechanism: Attackers exploit knowledge of the model's training distribution to craft inputs that appear benign but lie in blind spots of the feature space.
• Example: A market-making algorithm is fed a sequence of spoofed orders designed to shift the observed order book imbalance distribution, tricking the model into quoting unfavorable prices.
• Defense: Employ robust statistical monitoring with outlier-resistant metrics like the Median Absolute Deviation (MAD) and adversarial input detection layers.

A specific type of data drift caused by discrepancies between the data pipeline used during model training and the pipeline serving predictions. Common causes include:

• Feature engineering code that behaves differently in batch vs. real-time contexts
• Data preprocessing steps applied in training but omitted in production
• Aggregation windows that differ between offline and online computation Mitigation requires rigorous pipeline testing and serving infrastructure that guarantees identical transformations, often enforced through shared feature store logic.

A formal statistical term describing the scenario where the distribution of input variables P(X) changes between training and deployment, while the conditional distribution P(Y|X) remains stable. This is the most common form of data drift. For instance, a housing price model trained on pre-pandemic data encounters covariate shift when deployed during a market where square footage distributions have shifted toward suburban properties. Detection typically uses two-sample hypothesis tests like the Kolmogorov-Smirnov test or maximum mean discrepancy (MMD) to compare training and production feature distributions.