A gradual change in the statistical properties of a dataset over time, which can affect data quality and the performance of models trained on that data.
Data drift occurs when the real-world patterns that a dataset captures change over time. The data collection process may remain the same, but the underlying population, behaviour, or environment shifts. This can make historical data less representative of current conditions and cause predictive models to become less accurate.
Example: A model trained to predict property values uses historical sales data. Over five years, a neighbourhood undergoes significant development. The model's predictions become less accurate because the statistical patterns in the training data no longer reflect current market conditions — the data has drifted.
Monitoring for data drift is important in any system that relies on data collected over time, particularly machine learning applications and long-running analytics dashboards.