PACF vs ACF: Key Differences
You'll learn: How PACF isolates direct lag relationships while ACF captures both direct and indirect correlations.
The Core Difference
When you calculate the ACF (Autocorrelation Function) at lag k, you measure the total correlation between a time series and its k-period-ago self. This includes both direct and indirect effects.
The PACF (Partial Autocorrelation Function), however, measures only the direct relationship at lag k, controlling for all intermediate lags. It "partials out" the influence of lags 1, 2, ..., k-1.
An Analogy
Imagine you're measuring family resemblance:
- ACF is like saying "I look like my grandmother" — this might be true, but partly because you look like your mother, who looks like her mother
- PACF is like saying "I inherited traits directly from my grandmother (beyond what came through my mother)" — the unique, direct connection
Why This Matters
The ACF at lag 3 might show strong correlation simply because:
- Lag 1 correlates with current values
- Lag 2 correlates with lag 1
- Lag 3 correlates with lag 2
This creates an indirect path. The PACF at lag 3 removes these intermediate effects, revealing whether lag 3 adds any new information beyond what lags 1 and 2 already explain.
When to Use Which
- ACF: Shows the overall correlation structure; useful for detecting moving average (MA) patterns
- PACF: Reveals direct dependencies; useful for detecting autoregressive (AR) patterns
Key Takeaway: ACF includes all direct and indirect correlations through intermediate lags, while PACF isolates only the direct relationship at each lag by controlling for all shorter lags.