Drift denotes the gradual change in the statistical properties of data after a model is deployed, causing its quality to decline over time even though the model itself has not changed at all. Two kinds are distinguished. Data drift (also covariate shift) means a change in the distribution of the input features – the customer mix changes, a different sensor type is adopted, a new market opens – while the relationship between features and target remains. Concept drift is more serious: the relationship between input and output itself changes, so the earlier regularity ceases to hold. Classic examples are customer behaviour after an economic shock, a change in the tactics of fraudsters reacting to detection, or a shift in the meaning of words. By speed, drift may be sudden, gradual, incremental or recurring seasonally. Detection proceeds either by monitoring error against ground truth as it becomes available, or – when that is unavailable – by statistically comparing distributions using tests and indices such as PSI. The response is retraining on current data, a sliding window or online learning, always with an eye on forgetting. Drift monitoring is therefore an integral part of operations, not an optional extra.
Imagine a map of a city you drew ten years ago and have trusted unquestioningly ever since. At first it works. Then a new district goes up – that is data drift, you are driving through places the map does not have. And one day the old centre introduces one-way streets running the other way – that is concept drift, because now the map actively leads you astray. It does not help that you drew it carefully. You simply have to go out and draw it again – and above all, you need someone watching for the city to change.