Course contentsShow
← Data Science
Lesson 727 of 2,14519. Time Series FoundationsFree lesson

What is Partial Autocorrelation?

Definition of PACF as correlation between observations at different lags after removing intermediate correlations.

What is Partial Autocorrelation?

You'll learn what partial autocorrelation is and how it differs from regular autocorrelation by isolating direct relationships between observations.

The Core Idea

When you learned about autocorrelation (ACF), you saw how a time series correlates with its own past values at different lags. But there's a catch: the correlation at lag 3 might just be the indirect result of correlations at lags 1 and 2 passing through.

Partial autocorrelation (PACF) solves this by measuring the direct correlation between observations separated by k time steps, after removing the influence of all the intermediate lags.

A Simple Analogy

Imagine you want to know if your great-grandmother's height directly influences yours. Regular autocorrelation would say "yes, there's correlation" — but that's misleading! Your height correlates with your parent's height, which correlates with your grandparent's height, and so on. The correlation with your great-grandmother is mostly indirect, passing through the generations.

PACF asks: "What's the direct relationship with my great-grandmother's height, after accounting for everyone in between?" Often, you'll find it's near zero — the connection was just inherited through the intermediate generations.

Why It Matters

In time series, PACF helps you identify:

  • How many direct lag relationships actually exist
  • Which past values have genuine predictive power
  • The structure of autoregressive (AR) processes

Together with ACF, PACF is essential for model identification — helping you choose which past values to include when forecasting.

Key Takeaway: Partial autocorrelation isolates the direct correlation between observations at a specific lag by removing the confounding effects of all intermediate lags, revealing true temporal dependencies.