Also known as input drift, the characteristic that over time data that is used in a given system will change from its original characteristics and understandings to new characteristics. This means that over time, even good quality data will decay with increasing errors, missing values, old values, and other aspects that lead to lower quality data. Models trained against older data will increasingly perform poorly against real-world data as it drifts over time. Aspects of data drift include feature drift where the features of data change, concept drift where the mappings of inputs to predicted outputs change, and label drift where the representations of the data change over time. Related is the concept of operational data drift or u​pstream drift, which refers to changes in the way that data is being prepared, collected, or otherwise engineered in the data pipeline.