Skip to content
    Flow-alerta.cab

    Backtesting

    AI fundamentals & modelsBack-testHistorical simulation
    Editorially reviewed on: 1 source
    Your reading view

    The short answer · 19 words

    In short

    Examination of a rule, strategy or forecasting method based on past data without incorporating future information into previous decisions.

    Flow-alerta.cab (2026): Backtesting. https://flow-alerta.cab/en/glossar/backtesting. Editorially reviewed on: 09 October 2026.

    How much detail would you like?

    Practical

    Important for time-dependent models and rules. A clean backtest reduces data leakage but does not guarantee future performance.

    Example scenario: A sales forecast is recalculated each month with all data available until then and tested on the following month.

    For forecasting, capacity planning or rule-based decisions, simulate timely data statuses and compare results over several time windows.

    At a glance

    Core
    In backtesting, a process is evaluated as if historical times were the present one after the other. For prognosis models, a rolling origin is often used: trained only with data that would have already been available at the time. Backtesting is particularly known from financial and time series applications.
    Use
    Important for time-dependent models and rules. A clean backtest reduces data leakage but does not guarantee future performance.
    Example
    Example scenario: A sales forecast is recalculated each month with all data available until then and tested on the following month.
    Limits
    Include future data in training or feature calculation; optimize parameters on the final test period; ignore transaction or operating costs; overlook structural changes.

    In context

    Data Leakage in Machine Learning

    Data leakage occurs when the training or test incorporates information that would not be available at the real prediction time.

    Questions about this term

    Is Backtesting the Same as Random Cross Validation?

    No.

    Does a good backtest prove future success?

    No.

    Why should future information not be allowed in?

    Otherwise, the historical performance will be artificially improved.

    All details and background

    In backtesting, a process is evaluated as if historical times were the present one after the other. For prognosis models, a rolling origin is often used: trained only with data that would have already been available at the time. Backtesting is particularly known from financial and time series applications.

    ELI5 – Explained simply

    You play the past step by step and check how your method would have cut back then, without looking into the future.

    Why it matters in practice

    Important for time-dependent models and rules. A clean backtest reduces data leakage but does not guarantee future performance.

    Example

    Example scenario: A sales forecast is recalculated each month with all data available until then and tested on the following month.

    Pitfalls and limits

    • Include future data in training or feature calculation; optimize parameters on the final test period; ignore transaction or operating costs; overlook structural changes.

    Origin & history

    In time series analysis, forecasts are often evaluated with a rolling historical origin. For each test time, only previously available data may be included in the training.

    Use in mid-sized companies

    • For forecasting, capacity planning or rule-based decisions, simulate timely data statuses and compare results over several time windows.

    Related terms

    Matching Flow-alerta service

    AI opportunity analysis & roadmapAI consulting: AI potential analysis & roadmap on tasks, data and prerequisites of a possible first AI use case.

    Matching tool / use case

    View use casesUse Case: Walk forward evaluation of a monthly sales forecast.

    Sources

    • otexts.comHyndman & Athanasopoulos, Forecasting: Principles and Practice – Time series cross-validation

    Quellenprüfung: 09.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases