Autocorrelation measures how a series correlates with its lagged values.
ACF (Autocorrelation Function): Correlation at each lag. Includes indirect effects through intermediate lags.
PACF (Partial ACF): Direct correlation at each lag, controlling for shorter lags.
Using ACF/PACF:
- ACF decays slowly, PACF cuts off → AR model
- ACF cuts off, PACF decays → MA model
- Both decay → ARMA model
Interview question: "What does a spike at lag in ACF mean?"
Weekly seasonality. Today correlates with same day last week.