📏Akaike Information Criterion

NEW

Also known as: AIC

A criterion balancing model fit and complexity to compare statistical models for selection.

DESCRIPTION

The Akaike Information Criterion (AIC) estimates out-of-sample prediction error by penalizing model complexity relative to goodness of fit. Lower AIC indicates a better trade-off between bias and variance among candidate models fitted to the same data. It is computed as AIC = 2k − 2ln(L), where k is the number of parameters and L the maximum likelihood. AIC is widely used for model selection in regression, time series, and probabilistic modeling.

ALTERNATIVE DEFINITIONS

Wikipedia

The Akaike information criterion (AIC) is an estimator of prediction error and thereby relative quality of statistical models for a given set of data.

Highlights prediction error estimation and relative model quality.

Source

Penn State STAT 501

AIC provides a means for model selection. It balances the goodness of fit of the model against the complexity of the model.

Educational framing focusing on the bias–variance trade-off via penalization.

Source

RELATED TERMS

Model Selection Model Overfitting Cross-Validation Likelihood Penalized Likelihood Information Criterion Kullback–Leibler Divergence

COMPETING TERMS

Bayesian Information Criterion

KNOWLEDGE AREAS

Mathematics Data Science AI Testing & Validation

TARGET AUDIENCE

Data Scientist Data Analyst AI Engineer MLOps Engineer

EXTERNAL SOURCES

• Wikipedia - Akaike information criterion
• AICcmodavg Vignette - Guide to AIC and model selection in R
• NIST/SEMATECH e-Handbook - Model selection criteria including AIC
• Rob J Hyndman - Clarifications on AIC usage and pitfalls

TOP QUESTIONS

How do I interpret AIC values across models?

Lower AIC is better. Compare only models fit on the same dataset with the same response; differences > 2 are often considered meaningful.

Category: Management

What is the difference between AIC and BIC?

BIC penalizes complexity more strongly and is consistent with a true model assumption; AIC targets predictive accuracy.

Category: Alternatives

When should I use AICc instead of AIC?

Use AICc for small samples where k/n is not negligible. It corrects AIC’s small-sample bias.

Category: Requirements

Can I compare non-nested models with AIC?

Yes, AIC supports comparing non-nested models as long as they are fit to the same data and likelihood is comparable.

Category: Example

What is a common mistake with AIC?

Comparing models trained on different datasets or responses. This invalidates AIC comparisons.

Category: Mistakes

Does lower AIC guarantee better real-world performance?

Not always. Validate with held-out data or cross-validation to confirm predictive gains.

Category: Risks

Is AIC appropriate for penalized models (e.g., Lasso)?

Use criteria adapted for effective degrees of freedom or alternative validation methods; naive k counts can mislead.

Category: Opinion

What is the theoretical basis for AIC?

It approximates out-of-sample Kullback–Leibler divergence between the true process and the candidate model.

Category: Example

How do I report AIC results?

Provide AIC for candidates, ΔAIC, and model weights. Include assumptions and data scope.

Category: Management