Course contentsShow
← Data Science
Lesson 791 of 2,14520. Time Series ModelsFree lesson

Comparing Nested and Non-Nested Models

Using likelihood ratio tests for nested models and information criteria for non-nested model comparisons.

Comparing Nested and Non-Nested Models

What you'll learn: How to choose between competing time series models using different statistical tests depending on whether your models are nested or not.

Understanding Nested vs Non-Nested Models

When you've built multiple ARIMA models, you need a fair way to compare them. The comparison method depends on the relationship between your models.

Nested models occur when one model is a simplified version of another. For example, ARIMA(1,1,1) contains ARIMA(1,1,0) as a special case—just set the MA coefficient to zero. Think of nested models like Russian dolls: the smaller fits perfectly inside the larger.

Non-nested models are competitors that can't be simplified into one another. ARIMA(2,1,0) and ARIMA(0,1,2) are non-nested—neither is a special case of the other. They're more like apples and oranges: both fruit, but fundamentally different.

Choosing Your Comparison Tool

For Nested Models: Likelihood Ratio Test

When models are nested, use the Likelihood Ratio Test (which you learned earlier). This test statistically determines whether the extra complexity in the larger model significantly improves fit. It directly tests whether those additional parameters are worth keeping.

For Non-Nested Models: Information Criteria

When models aren't nested, you can't use likelihood ratio tests. Instead, rely on AIC and BIC (information criteria you've already learned). These metrics balance model fit against complexity, letting you compare any models—even if they're structurally different. Lower AIC/BIC values indicate better models.

Think of it like shopping: nested model comparison asks "is the deluxe version worth the upgrade?" while non-nested comparison asks "which product is the better overall value?"

Key Takeaway: Use likelihood ratio tests when comparing nested models (one is a special case of another), and use AIC/BIC when comparing non-nested models (structurally different alternatives).