What Correlation Can—and Cannot—Tell You
Autocorrelation is useful because it summarizes how a process is related to its own past through covariance. But it only measures linear, second-order dependence. A time series can have zero autocorrelation at every nonzero lag and still exhibit strong dependence in its volatility, its extreme events, or other nonlinear features. Even a rapidly decaying autocorrelation function need not imply that dependence between the distant past and future becomes small in a broader sense.
This distinction matters in econometrics. If large price changes tend to be followed by large changes, whether positive or negative, the relationship may be important for risk even when the signed changes have little linear association. The autocorrelation function alone cannot rule out that pattern. To describe broader dependence, one can examine transformations such as squared observations or use dependence coefficients that consider events in entire past and future information sets.
Zero covariance means that the product of centered observations has expectation zero. Independence is much stronger: it requires that all events determined by one observation be independent of all events determined by the other. A zero covariance does not generally establish independence. Likewise, a sequence of small autocorrelations says little by itself about how well separated collections of observations are independent.
A Nonlinear Relationship with Zero Correlation
A simple pair of random variables makes the limitation concrete. Let \(U\) take each of \(-2,-1,1,2\) with probability \(1/4\), and define \(X=U\) and \(Y=U^2-2.5\). The relationship between \(X\) and \(Y\) is deterministic and nonlinear, but their correlation is zero.
Worked Example: Zero Correlation, Perfect Nonlinear Dependence
First calculate the means. By symmetry, \(E[X]=E[U]=0\). Also,
Since \(X=U\), its variance is \(E[U^2]=2.5\). The values of \(Y\) are \(-1.5,-1.5,1.5,1.5\), so \(\operatorname{Var}(Y)=2.25\). The covariance is
Thus the correlation is \(0/\sqrt{2.5(2.25)}=0\). Yet \(Y\) is determined exactly by \(X\). For example, the event \(\{|X|=2\}\) is exactly the event \(\{Y=1.5\}\). Each event has probability \(1/2\), while their intersection has probability \(1/2\), not the \(1/4\) that independence would require. The zero correlation has missed a clear event-based dependence.
This example is not itself a time series, but it isolates the key issue: a measure built from a product of the original variables can miss a relationship that becomes obvious after a nonlinear transformation or event definition. The same issue arises across time. In particular, \(X_t^2\), \(|X_t|\), or an indicator for an extreme observation may have a very different dependence pattern from \(X_t\).
Zero Autocorrelation with Persistent Volatility Dependence
Consider a stationary two-state process \(S_t\) that takes the values 1 and 2. At each period it stays in its current state with probability \(q=0.95\) and switches states with probability \(0.05\). Its stationary probabilities are \(1/2\) for each state. Let \(\varepsilon_t\) be independent random signs, each equal to \(1\) or \(-1\) with probability \(1/2\), independent of the entire \(S_t\) process. Define \(X_t=S_t\varepsilon_t\). These assumptions give a stationary process with mean zero and finite variance.
Worked Example: The Signed Series and Its Squares
For any positive lag \(h\), independence of the random signs gives
because the signs at distinct dates are independent and each has mean zero. The variance is \(E[X_t^2]=E[S_t^2]=(1+4)/2=2.5\). Therefore, \(\rho(h)=0\) at every nonzero lag.
Now examine the squares. Since \(X_t^2=S_t^2\), define \(W_t=X_t^2-2.5\). It takes the values \(-1.5\) and \(1.5\), each with probability \(1/2\), so \(\operatorname{Var}(W_t)=2.25\). The two-state transition rule implies
To see the one-period step, conditional on the current state, \(W\) stays at its current value with probability \(q\) and changes sign with probability \(1-q\). Its conditional expectation is therefore \((q-(1-q))W_t=(2q-1)W_t\); iterating gives the displayed \(h\)-period result. Consequently,
The autocorrelation of the squared observations is thus \(0.9^h\). For example, at lag 5 it is \(0.9^5=0.5905\), even though the autocorrelation of the signed observations is exactly zero at that lag. Large absolute values tend to occur in the same volatility state. Looking only at the signed autocorrelation would conceal this substantial, slowly diminishing dependence in the squares.
The example illustrates why a useful diagnostic may examine transformed series as well as the original series. ACFs of squares or absolute values can reveal dependence in volatility that the ACF of signed observations cannot. They are still summaries, not universal tests for independence: other nonlinear or event-based patterns may remain undetected.
Rapidly Decaying Correlation Without Decaying Past–Future Dependence
Autocorrelation can also decay quickly while a dependence measure based on entire past and future information sets fails to decay at all. To state this precisely, let \(\mathcal{F}_{-\infty}^{0}\) be the information generated by \(X_t\) for \(t\leq 0\), and let \(\mathcal{F}_{k}^{\infty}\) be the information generated by \(X_t\) for \(t\geq k\). The alpha-mixing coefficient at separation \(k\) is
The coefficient considers all such past and future events, not just the covariance of two individual observations. Alpha mixing requires \(\alpha(k)\to 0\) as \(k\to\infty\). This is one standard, precise notion of weakening dependence; “weak dependence” can refer to other notions as well. The earlier material on Alpha-Mixing and Other Dependence Coefficients introduces this broader approach.
Worked Example: Geometric ACF but a Nonvanishing Alpha Coefficient
Let \(Y_t\) be a stationary AR(1) process with coefficient \(0.4\) and innovation variance \(0.84\). Its variance is \(0.84/(1-0.4^2)=1\). Let \(Z\) be a single random variable, shared across all dates, that equals 1 or 4 with probability \(1/2\) each. Let \(\varepsilon_t\) be independent standard normal variables. Assume \(Z\), the entire \(\varepsilon_t\) sequence, and \(Y_t\) are mutually independent, and set \(X_t=Y_t+\sqrt{Z}\varepsilon_t\). This process is stationary, has mean zero, and has finite variance.
For positive lags, the independent noise terms have zero covariance with each other and with \(Y_t\). Thus \(\gamma(h)=\operatorname{Cov}(Y_t,Y_{t-h})=0.4^h\). At lag zero, \(\gamma(0)=1+E[Z]=3.5\). The autocorrelation is therefore
For example, \(\rho(1)=0.1143\) and \(\rho(5)=0.002926\), so the autocorrelation decays geometrically. But the shared random scale \(Z\) remains in observations at every date. Conditional on \(Z\), the process is stationary and ergodic; its long-run average of squared observations converges to \(1+Z\). The earlier material on Ergodicity and When Time Averages Reveal Population Moments gives the relevant interpretation of such time averages. If \(Z=1\), the limit is 2; if \(Z=4\), it is 5.
For any fixed \(k\), define \(A\) as the past event that the average of \(X_t^2\) over \(t=-m,\ldots,0\) converges to a value greater than 3. Define \(B_k\) as the corresponding future event using \(t=k,\ldots,k+m\). Both events identify \(Z=4\), up to probability-zero exceptions. Hence \(P(A)=P(B_k)=1/2\) and \(P(A\cap B_k)=1/2\), so
Because these events belong to the separated past and future information sets, \(\alpha(k)\geq 0.25\) for every \(k\). Thus the autocorrelation decays geometrically, but the process is not alpha-mixing. Distant observations remain linked through the common scale, which can be inferred from an entire stretch of observations.
Why the Distinction Matters
Correlation decay is informative about linear persistence, but it is not a general certificate that dependence has become negligible. In applications, dependence in volatility or tail events can affect uncertainty assessments and the behavior of procedures that rely on observations being weakly dependent. The appropriate diagnostic depends on the feature of the data that matters: signed values, magnitudes, threshold events, or dependence between whole information sets.
A useful way to keep the concepts separate is:
| Quantity | What it assesses | What it cannot establish alone |
|---|---|---|
| Autocorrelation \(\rho(h)\) | Linear association between observations \(h\) periods apart | Independence or decay of all nonlinear and event-based dependence |
| Autocorrelation of \(X_t^2\) or \(|X_t|\) | Linear association in a transformed series, often related to volatility patterns | Absence of every other form of dependence |
| Alpha-mixing coefficient \(\alpha(k)\) | The largest event-based departure from independence between separated past and future information sets | Every possible definition of weak dependence; it is one specific dependence condition |
Common pitfall: “The ACF is zero, or decays rapidly, so observations far apart are nearly independent.” This conclusion does not follow. An ACF describes pairwise linear association, whereas independence concerns all relevant events; the common-scale example even has a geometrically decaying ACF and a nonvanishing alpha coefficient. A more careful conclusion is: “The observed autocorrelations indicate little linear association at these lags; other forms of dependence require separate consideration.”
Check Your Understanding
Use the definitions and examples above to answer these questions.
- In the first worked example, why are \(X\) and \(Y\) dependent even though their correlation is zero?
- For the stochastic-volatility process, what is the autocorrelation of \(X_t\) at a positive lag, and what is the autocorrelation of \(X_t^2\) at lag 5?
- In the common-scale example, why can both a distant past event and a distant future event identify the same value of \(Z\)?
- What does alpha mixing require about \(\alpha(k)\) as the separation \(k\) grows?
- Give one careful interpretation of a rapidly decaying sample ACF that does not claim independence.