Data Drift

ELI5 - The Vibe Check

Data drift is when the real world quietly stops looking like the world your model was trained on. A model trained on pre-pandemic shopping habits might keep running just fine technically, while silently getting worse and worse at predicting anything, because customer behavior shifted and nobody updated the model to notice.

Real Talk

Data drift occurs when the statistical properties of a model's input data change over time relative to the data it was trained on, causing prediction quality to degrade even though the model itself hasn't changed. It's typically detected by monitoring the distribution of incoming features against a training baseline and triggers retraining when the divergence crosses a threshold.

When You'll Hear This

"The model's been live for eight months with no retraining, we should check for data drift before trusting its output." / "Our monitoring flagged data drift on the income feature, average incomes in the input data have shifted a lot since training."