Skip to content
    Flow-alerta.cab

    Data Drift

    AI fundamentals & modelsData distribution driftFeature driftInput drift
    Editorially reviewed on: 1 source
    Your reading view

    The short answer · 16 words

    In short

    Data drift refers to a change in the distribution of input data versus a fixed reference.

    Flow-alerta.cab (2026): Data Drift. https://flow-alerta.cab/en/glossar/data-drift. Editorially reviewed on: 10 October 2026.

    How much detail would you like?

    Practical

    Drift monitoring can point to changing customers, processes, sensors or detection systems at an early stage. It helps to trigger exams and, if necessary, follow-up training in a risk-based manner.

    After a range change, the distribution of shopping cart values and product categories is shifted compared to the training data of a forecasting model.

    Organizations monitor business-critical characteristics, document reference periods, and define responses. SMEs should start with few explainable signals and a manual verification process.

    At a glance

    Core
    Data drift changes the statistical distribution of characteristics P(X) between reference and operational data. It can be distinguished from Concept Drift, i.e. a change of P(Y|X), from Label Drift and from Training-Serving Skew. Drift is a test indication and not automatically a drop in performance.
    Use
    Drift monitoring can point to changing customers, processes, sensors or detection systems at an early stage. It helps to trigger exams and, if necessary, follow-up training in a risk-based manner.
    Example
    After a range change, the distribution of shopping cart values and product categories is shifted compared to the training data of a forecasting model.
    Limits
    Treat any statistical deviation as errors, set unsuitable thresholds or observe only individual features. Seasonal changes, data quality errors and real concept drift must be assessed separately.

    In context

    Concept Drift

    Change in the relationship between input data and target variable, which can make a previously suitable model less reliable over time.

    Questions about this term

    Is Drift Equally Poor Model Quality?

    No.

    How do you react?

    Check cause and performance indicators, then adjust data, model or process.

    What is Training-Serving Skew?

    A difference between training and production data due to generation or pipeline deviations.

    All details and background

    Data drift changes the statistical distribution of characteristics P(X) between reference and operational data. It can be distinguished from Concept Drift, i.e. a change of P(Y|X), from Label Drift and from Training-Serving Skew. Drift is a test indication and not automatically a drop in performance.

    ELI5 – Explained simply

    One model was learned with a certain mix of cases. If new cases are significantly different, you should check whether it still works reliably.

    Why it matters in practice

    Drift monitoring can point to changing customers, processes, sensors or detection systems at an early stage. It helps to trigger exams and, if necessary, follow-up training in a risk-based manner.

    Example

    After a range change, the distribution of shopping cart values and product categories is shifted compared to the training data of a forecasting model.

    Pitfalls and limits

    • Treat any statistical deviation as errors, set unsuitable thresholds or observe only individual features. Seasonal changes, data quality errors and real concept drift must be assessed separately.

    Origin & history

    With the operation of ML models, model monitoring emerged as a discipline of its own. Methods compare reference and production distributions, for example over distance dimensions and threshold values.

    Use in mid-sized companies

    • Organizations monitor business-critical characteristics, document reference periods, and define responses. SMEs should start with few explainable signals and a manual verification process.

    Related terms

    • Concept drift
    • Model monitoring
    • Training Serving Skew
    • Data quality
    • MLOs

    Matching Flow-alerta service

    AI opportunity analysis & roadmapAI Quality Assurance: Set up reference data, drift indicators, warning thresholds and testable response processes for productive models.

    Matching tool / use case

    View use casesVertex AI Model Monitoring or BigQuery ML; Use Case: Monitoring input distributions of a demand or risk model.

    Sources

    Quellenprüfung: 10.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases