Concept Drift

An abrupt, discontinuous change in the data-generating process, often triggered by a singular external shock.

• Mechanism: The function f(x) changes instantly at a specific point in time.
• Supply Chain Example: A factory fire at a sole-source supplier immediately shifts lead time from 5 days to 'indefinite.'
• Detection Strategy: Control charts and change-point detection algorithms on the forecast error residuals.

A gradual, continuous evolution of the target concept over time, often due to natural market progression.

• Mechanism: The decision boundary slowly shifts position; the model becomes progressively more outdated.
• Supply Chain Example: A new product slowly gains market acceptance, causing a steady, month-over-month increase in baseline demand that a static model misses.
• Detection Strategy: Adaptive windowing (ADWIN) or monitoring the KL divergence between recent and reference data distributions.

Cyclical or seasonal changes where the statistical properties oscillate predictably between distinct states.

• Mechanism: The concept switches between two or more previously seen contexts, often tied to a temporal cycle.
• Supply Chain Example: Demand patterns for swimwear oscillate between a 'summer' state and a 'winter' state annually, or weekly grocery demand peaks every Saturday.
• Detection Strategy: Season-aware models like Bayesian Structural Time Series that explicitly model periodicity rather than treating it as drift.

A change in the distribution of the input features P(X) without a change in the conditional target distribution P(Y|X).

• Mechanism: The underlying relationship remains true, but the model sees data it was never trained on.
• Supply Chain Example: A forecasting model trained on pre-pandemic consumer mobility data encounters lockdown-era stay-at-home patterns. The relationship between mobility and demand might hold, but the input space has shifted.
• Detection Strategy: Monitoring the population stability index (PSI) of input features in the feature store.

A change in the conditional relationship P(Y|X) itself, meaning the same inputs now produce different outputs.

• Mechanism: The fundamental causal link between the predictor and the target breaks or alters.
• Supply Chain Example: A price drop previously guaranteed a 20% demand lift. Due to a competitor's aggressive promotion, the same price drop now yields only a 5% lift. The elasticity has changed.
• Detection Strategy: Directly monitoring the predictive performance metrics (e.g., CRPS) on a delayed ground-truth window.

The appearance, disappearance, or semantic change of input features over the model's lifecycle.

• Mechanism: The feature schema itself is non-stationary, breaking the inference pipeline.
• Supply Chain Example: A logistics provider adds a new 'carbon cost' feature to the API, or a legacy 'regional_code' field is deprecated and replaced with a 'geo_cluster' ID.
• Detection Strategy: Schema validation on incoming data streams and monitoring the fraction of missing values for critical features.

A specific type of dataset shift where the distribution of input features P(X) changes between training and production, while the conditional relationship P(Y|X) remains stable. In demand forecasting, this manifests when customer demographics or browsing patterns evolve but purchase behavior given those features stays constant. Unlike concept drift, covariate shift can often be corrected through importance weighting or domain adaptation without retraining the model's core logic.

A machine learning paradigm where models update continuously as new data streams arrive, processing one observation at a time rather than retraining on full batches. This approach is the primary defense against concept drift in high-velocity supply chains, enabling demand forecasts to adapt to shifting consumer behavior within minutes. Key algorithms include stochastic gradient descent with adaptive learning rates and Hoeffding trees for streaming classification.

The systematic evaluation of forecasting models by applying them to historical data and comparing predictions against actual outcomes using a rolling or expanding window. For concept drift detection, backtesting reveals when model performance degrades over specific time periods. Practitioners use time-series cross-validation with embargo periods to prevent data leakage, calculating metrics like RMSE and Continuous Ranked Probability Score across multiple temporal splits.

The reducible uncertainty arising from a lack of knowledge or insufficient training data. In the context of concept drift, high epistemic uncertainty signals that the model has encountered input regions poorly represented in its training distribution—a potential early warning indicator of drift. Unlike aleatoric uncertainty, which represents irreducible noise, epistemic uncertainty can be decreased by gathering more representative samples or employing Bayesian neural networks that explicitly model weight uncertainty.

A centralized repository for storing, versioning, and serving curated features for machine learning models. Feature stores are critical infrastructure for concept drift management because they maintain point-in-time correctness—ensuring that the exact feature values used during training are reproducible during inference. When drift is detected, data scientists can query historical feature distributions to identify which specific inputs have shifted, accelerating root cause analysis.

A technique that combines predictions from multiple diverse models to produce a single, more robust probabilistic forecast. Ensembles provide natural resilience against concept drift because different constituent models may respond differently to distributional shifts. Approaches include Bayesian model averaging, stacking, and dynamic model selection where the ensemble weights adapt based on recent performance, effectively creating a self-adjusting defense against gradual drift.