Bias–Variance Tradeoff
The short answer · 19 words
In short
Classical concept for balancing systematic model error due to too simple assumptions and sensitivity to fluctuations in training data.
Flow-alerta.cab (2026): Bias–Variance Tradeoff. https://flow-alerta.cab/en/glossar/bias-variance-tradeoff. Editorially reviewed on: 09 October 2026.
How much detail would you like?
Practical
Helps with model selection, regularization and interpretation of underfitting and overfitting. Modern, highly overparameterized models can deviate from the classical image.
Example scenario: A very simple regression approach systematically misses a pattern; an extremely flexible model, on the other hand, adjusts noise of the training set.
Observe training and validation errors separately, align model complexity and regularization, and test decisions on independent test data.
At a glance
- Core
- High bias models are often too inflexible and underfit; high variance models are highly responsive to specific training data and can overfit. In classically supervised learning problems, the bias-variance dilemma describes why model complexity must be chosen so that new data is generalized as well as possible.
- Use
- Helps with model selection, regularization and interpretation of underfitting and overfitting. Modern, highly overparameterized models can deviate from the classical image.
- Example
- Example scenario: A very simple regression approach systematically misses a pattern; an extremely flexible model, on the other hand, adjusts noise of the training set.
- Limits
- Confuse bias in the sense of fairness with statistical bias; consider the classic U-shaped image as a universal law; evaluate only training errors.
In context
Questions about this term
What does high bias mean?
Often systematic underfitting.
What does high variance mean?
Strong dependence on the specific training set.
Does the classic tradeoff always apply exactly?
No.
All details and background
High bias models are often too inflexible and underfit; high variance models are highly responsive to specific training data and can overfit. In classically supervised learning problems, the bias-variance dilemma describes why model complexity must be chosen so that new data is generalized as well as possible.
ELI5 – Explained simply
A model can be too simplistic and miss important patterns or too sensitive and memorize training details. We are looking for a good middle ground.
Why it matters in practice
Helps with model selection, regularization and interpretation of underfitting and overfitting. Modern, highly overparameterized models can deviate from the classical image.
Example
Example scenario: A very simple regression approach systematically misses a pattern; an extremely flexible model, on the other hand, adjusts noise of the training set.
Pitfalls and limits
- Confuse bias in the sense of fairness with statistical bias; consider the classic U-shaped image as a universal law; evaluate only training errors.
Origin & history
Bias variance decomposition is a classic concept of statistical learning. Recent research shows that modern overparameterized models can show more complex generalization gradients such as double descent.
Use in mid-sized companies
- Observe training and validation errors separately, align model complexity and regularization, and test decisions on independent test data.
Related terms
- Underfitting
- Overfitting
- Generalisation
- Regularisation
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: Diagnosis of underfitting and overfitting based on training and validation performance.Sources
- www.ibm.comIBM, Verzerrung-Varianz-Dilemma
- doi.orgBelkin et al., PNAS 2019
Quellenprüfung: 09.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.
Editorially reviewed on: