Decay Rates in ACF
What you'll learn: How the speed at which autocorrelation values drop off tells you whether your time series is stationary or not.
The Core Idea
When you plot an ACF, you're looking at how correlation changes as you increase the lag between observations. The rate of decay—how quickly those bars shrink toward zero—reveals crucial information about your series' behavior.
There are two main patterns to recognize:
Exponential (Fast) Decay
In a stationary time series, the ACF bars drop off quickly and exponentially. Picture a ball rolling down a steep hill—it loses height rapidly. By lag 5 or 10, most autocorrelations have fallen inside the significance bounds (those dotted lines), becoming statistically indistinguishable from zero.
This fast decay means past values lose their influence quickly. The series has no "memory" of the distant past, which is exactly what we expect from stationary data.
Slow Decay
In a non-stationary time series (especially one with a trend), the ACF bars decrease very slowly—if at all. Imagine a ball on a gentle slope: it takes forever to lose elevation. Even at lag 20 or 30, autocorrelations remain large and significant.
This slow decay signals that the series has a long memory. Values from many periods ago still strongly correlate with today's values, a telltale sign of non-stationarity.
Why This Matters
The decay rate is your visual shortcut for diagnosing stationarity. Before you see formal test results (like ADF), a glance at ACF decay often reveals the problem. Fast exponential decay? Likely stationary. Slow, lingering decay? You probably need differencing or detrending.
Key Takeaway: Exponential decay in the ACF indicates stationarity, while slow decay suggests non-stationarity and the presence of trends or long-term dependence.