Course contentsShow
← Data Science
Lesson 776 of 2,14520. Time Series ModelsFree lesson

Identifying AR Order (p) Using PACF

Use partial autocorrelation function plots to determine the appropriate AR order.

Identifying AR Order (p) Using PACF

What you'll learn: How to use the Partial Autocorrelation Function (PACF) plot to determine how many lagged terms you need in an Autoregressive (AR) model.

The Core Idea

When building an AR model, you need to decide the order p—that is, how many past values influence the current value. Should you use just yesterday's value? The last two days? Five days? The PACF plot gives you the answer by showing which lags have a direct relationship with the current value, filtered from indirect effects.

How PACF Reveals AR Order

Remember that PACF measures the correlation between a time point and a lag, after removing the influence of all shorter lags. For a pure AR(p) process, the PACF will:

  • Show significant spikes at lags 1, 2, ..., p
  • Cut off sharply to near-zero after lag p
  • Stay within the significance bounds for all lags beyond p

This cutoff pattern is your signal. If you see significant PACF values at lags 1 and 2, but everything after lag 2 is insignificant, you likely need an AR(2) model.

The Pattern Recognition

Think of PACF like tracing a family tree of influence. An AR(3) process means today's value is directly influenced by the past three days. The PACF will light up at lags 1, 2, and 3, then go dark—those are the only direct ancestors that matter. Everything beyond lag 3 is just indirect influence that's already captured.

The significance bounds (usually dashed lines on the plot) help you distinguish real patterns from random noise. Spikes outside these bounds indicate genuine relationships worth modeling.

Key Takeaway: The PACF plot's cutoff point tells you the AR order (p)—count the number of significant spikes before the PACF drops to insignificance and stays there.