PACF for AR Process Identification
You'll learn how to use the PACF plot's cutoff pattern to determine the correct order of an autoregressive model.
The Core Idea
The PACF (Partial Autocorrelation Function) is your primary diagnostic tool for identifying AR (autoregressive) processes and determining their order. Here's the key insight: for a pure AR(p) process, the PACF will show significant spikes up to lag p, then cut off sharply to near-zero beyond that lag.
This cutoff behavior is the PACF's superpower—it tells you exactly how many AR terms you need.
The AR Signature Pattern
When you look at a PACF plot:
- AR(1): Only the first lag is significant; all others fall within the confidence bounds
- AR(2): First two lags are significant; lag 3 onward drops off
- AR(p): First p lags are significant, then cutoff
This sharp cutoff happens because partial autocorrelation removes the indirect effects of earlier lags. Once you've accounted for all the true AR dependencies, there's nothing left to explain.
Distinguishing AR from MA
The PACF cutoff pattern is also how you distinguish AR from MA processes:
- AR process: PACF cuts off sharply; ACF decays gradually
- MA process: ACF cuts off sharply; PACF decays gradually
Think of it as mirror images: each process shows clean cutoff in one function and messy decay in the other.
Practical Use
When identifying a time series model, you examine the PACF plot alongside the ACF. If you see a clear cutoff at lag 2 in the PACF (and gradual decay in the ACF), you've identified an AR(2) process. The number of significant PACF spikes before cutoff is your AR order.
Key Takeaway: The PACF's sharp cutoff at lag p is the signature of an AR(p) process—count the significant spikes before cutoff to determine your AR model order, and use the cutoff vs. decay pattern to distinguish AR from MA processes.