Dernière mise à jour :
The change in the statistical properties of the target variable, which the model is trying to predict, over time, leading to model degradation.
Occurs when the underlying relationship between inputs and outputs shifts - e.g., seasonal trends, market shifts, or adversarial behaviors. Detecting drift requires monitoring input and output distributions, and governance workflows must define drift thresholds, retraining cadences, and human-in-the-loop checks before automatic redeployment.
Exemple concret :
A retailer’s demand-forecasting model sees accuracy fall sharply during a pandemic as buying patterns shift. Drift detectors flag the deviation, and the data-science team retrains the model on recent sales data with adjusted seasonality factors before the next planning cycle.
