Course contentsShow
← Data Science
Lesson 629 of 2,14516. Multiple Linear RegressionFree lesson

Akaike Information Criterion (AIC)

Use AIC to compare non-nested models by balancing goodness-of-fit with model complexity penalties.

Akaike Information Criterion (AIC)

You'll learn how to compare non-nested models using a metric that balances fit quality against model complexity.

What Is AIC?

The Akaike Information Criterion (AIC) is a tool for comparing different models—even when they're non-nested (meaning one model isn't simply a reduced version of the other). Unlike the F-test or likelihood ratio test, which require nested models, AIC lets you evaluate completely different model structures side-by-side.

AIC rewards models that fit the data well (low residual error) but penalizes models that use too many predictors. This penalty prevents overfitting—when a model becomes too complex and learns noise instead of true patterns.

How AIC Works

The formula includes two components:

  • Goodness-of-fit: How well the model explains the data (measured via likelihood)
  • Complexity penalty: A term that increases with the number of parameters (k)

Lower AIC values indicate better models. The penalty term means that adding predictors only helps if they genuinely improve fit enough to justify the added complexity.

When to Use AIC

Use AIC when:

  • Comparing non-nested models (e.g., predicting house price using [bedrooms, age] vs. [square_feet, location])
  • You want a simple number to rank multiple candidate models
  • You need an automated criterion for model selection

Think of AIC like judging a recipe: the best dish isn't always the one with the most ingredients—it's the one that tastes great (fits well) without unnecessary complexity.

Key Takeaway: AIC balances model fit against complexity, letting you compare any models by selecting the one with the lowest AIC value—even when those models aren't nested versions of each other.