Calculating Partial Autocorrelations
What you'll learn: You'll understand the two main computational methods for finding PACF values—using regression and the Yule-Walker equations.
The Core Idea
Unlike ACF (which captures all indirect relationships), PACF isolates the direct correlation between observations at different lags. But how do we actually compute these values? There are two primary approaches: the regression method and the Yule-Walker equations.
The Regression Approach
This is the more intuitive method. To find the PACF at lag k:
- Regress your current observation on all intermediate lags (1 through k)
- The coefficient on lag k in that regression is the partial autocorrelation
For example, PACF at lag 3 comes from regressing y_t on y_(t-1), y_(t-2), and y_(t-3). The coefficient on y_(t-3) gives you PACF(3).
Why this works: The regression "controls for" the intermediate lags, removing their influence and revealing only the direct relationship.
Yule-Walker Equations
This algebraic approach solves a system of equations based on the ACF values you already computed. It's computationally efficient because it calculates partial autocorrelations sequentially—each new lag builds on previous calculations.
The Yule-Walker method uses matrix algebra to relate ACF values to PACF values, making it faster for software implementations. You won't calculate these by hand, but statistical packages use this approach under the hood.
Which Method to Use?
In practice, Python and R handle this for you. The regression approach is conceptually clearer for understanding what PACF means, while Yule-Walker is computationally preferred for actual calculation.
Key Takeaway: PACF can be computed through regression (fitting models at each lag) or Yule-Walker equations (solving systems based on ACF values)—both isolate direct correlations by removing intermediate lag effects.