Bias-Variance Trade-Off Calculator

Calculation results

Waiting for analysis

Step-by-step calculation

Copyable calculation report

Choose an analysis mode

Squared bias

Variance

Irreducible noise variance

Model label

Complexity level

Training error

What-if sensitivity sliders

Squared bias: 4

Variance: 1.5

Noise: 0.5

True target value

Noise variance

Variance method

Repeated predictions

Use commas, spaces, semicolons, or one value per line.

Optional weights

Weights must match the number of predictions.

Optional multiple target points

Format each line as: target | predictions | noise variance.

Compare bias, variance, noise, errors, regularization, sample size, and cross-validation results.

Model	Complexity	Bias	Variance	Noise	Train	Validation	Test	Regularization	Sample size	CV score	Remove

Enter training and validation errors across increasing complexity levels.

Complexity	Training error	Validation error	Test error	Bias²	Variance	Remove

Actual class

Decision threshold

Optional noise variance

Repeated predicted probabilities

Every probability must be between zero and one.

The classic bias-variance decomposition is most direct under squared-error regression. Probability-based classification results are diagnostic approximations.

Advanced settings

Decimal places

Number format

Diagnosis sensitivity

Missing-value policy

Confidence level

Auto-download report

Import CSV data

Import behavior follows the active analysis mode.

Formula used

Bias = E[f̂(x)] − f(x)

Bias² = (E[f̂(x)] − f(x))²

Variance = E[(f̂(x) − E[f̂(x)])²]

Reducible Error = Bias² + Variance

Expected Error = Bias² + Variance + Irreducible Noise

How to use this calculator

Select the analysis mode matching your available data.
Enter error components, predictions, models, curves, or probabilities.
Choose variance, formatting, missing-value, and confidence settings.
Calculate and review the diagnosis, charts, table, and steps.
Copy, print, or download the report as CSV or PDF.

Example model comparison

Model	Complexity	Bias	Bias²	Variance	Noise	Total error	Diagnosis
Model A	1	2.00	4.00	0.50	1.00	5.50	Underfitting
Model B	3	0.80	0.64	1.10	1.00	2.74	Balanced
Model C	8	0.20	0.04	4.20	1.00	5.24	Overfitting

Understanding the trade-off

Bias measures systematic error caused by restrictive assumptions. High bias often produces underfitting and weak patterns. More flexibility can reduce this source of error.

Variance measures prediction changes across different training samples. High variance often indicates sensitivity and overfitting. Regularization, more data, or ensembles can reduce variance.

Irreducible noise reflects randomness unavailable to the model. It cannot be removed through model complexity alone. Better measurements may sometimes reduce observed noise levels.

Cross-validation estimates generalization across held-out data subsets. Training error alone usually favors excessive model complexity. Select complexity near the lowest stable validation error.

Frequently asked questions

1. What is the bias-variance trade-off?

It describes tension between systematic error and sampling sensitivity. Simpler models often have higher bias. Complex models often have higher variance.

2. What indicates underfitting?

High bias and low variance commonly indicate underfitting. Training error may also remain high. Adding useful complexity can improve the model.

3. What indicates overfitting?

Low bias with high variance commonly indicates overfitting. Training error may look deceptively low. Validation performance usually becomes less stable.

4. Why include irreducible noise?

Noise represents unpredictable variation in the outcome. Models cannot learn truly random changes. Total expected error therefore includes this component.

5. Should I use sample or population variance?

Use population variance when predictions represent the full experiment. Use sample variance for an estimated larger process. The selected denominator changes the numeric result.

6. How does regularization affect the trade-off?

Regularization usually increases bias while reducing variance. Moderate regularization can improve generalization. Excessive regularization can create underfitting.

7. Can ensembles reduce variance?

Bagging and averaging can reduce unstable prediction fluctuations. Diverse models are especially helpful. Correlated model errors limit the improvement.

8. Is the decomposition exact for classification?

The classic form is clearest for squared-error regression. Classification decompositions depend on loss definitions. This tool reports probability-based diagnostics carefully.

9. How should I select model complexity?

Compare validation error across increasing complexity levels. Prefer a stable minimum over a fragile dip. Confirm the choice using independent test data.