Why a Negative Lag-One ACF Deserves Attention
Differencing is used to remove nonstationarity, but applying it to a series that is already stationary can create unnecessary dependence. One recognizable consequence is a negative autocorrelation at lag one: adjacent differences share an observation with opposite signs. A pronounced negative first-lag spike can therefore be a useful warning when inspecting an ACF.
The warning is not a diagnosis by itself. A genuine moving-average process can also have a negative lag-one autocorrelation, and sample ACF values vary from sample to sample. The interpretation depends on the differencing operation, the remaining ACF pattern, and the context of the series. This tutorial develops the clue and its limitations, aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
What Differencing Does to Autocovariances
Suppose \(X_t\) is weakly stationary, with mean \(\mu\) and autocovariance \(\gamma_X(k)=\operatorname{Cov}(X_t,X_{t-k})\). Weak stationarity means that the mean is constant, the variance is finite and constant, and the covariance depends only on the time separation \(k\). Define \(Y_t=\Delta X_t=X_t-X_{t-1}\). Its mean is zero, and its autocovariance at lag \(k\) follows from expanding the covariance of two differences.
For \(k\geq 1\), the two same-direction covariance terms are each \(\gamma_X(k)\). The two cross terms are \(\gamma_X(k-1)\) and \(\gamma_X(k+1)\), and each is subtracted. At lag zero, the variance of a difference is the sum of the two variances minus twice their covariance. These identities require stationarity and finite second moments; they are not a formula for the ACF of an arbitrary nonstationary series.
The lag-one covariance is \(2\gamma_X(1)-\gamma_X(0)-\gamma_X(2)\). It is negative whenever the two adjacent-lag covariances together exceed twice the lag-one covariance. The fact that this can happen explains the diagnostic clue; it does not imply that every stationary series, when differenced, must have a negative lag-one ACF.
Worked Example: Differencing White Noise
Let \(X_t\) be white noise with mean zero and variance \(\sigma^2=9\). Thus \(\gamma_X(0)=9\), and \(\gamma_X(k)=0\) for every nonzero lag. Find the ACF of \(Y_t=X_t-X_{t-1}\), and explain why its lag-one value is a strong overdifferencing clue.
Method. Apply the differenced-series autocovariance formulas. White noise is stationary, so differencing it is unnecessary if the goal was simply to achieve stationarity.
At lag zero, \(\gamma_Y(0)=2\{9-0\}=18\). At lag one, \(\gamma_Y(1)=2(0)-9-0=-9\). For \(k\geq 2\), all three autocovariances of \(X_t\) in the formula are zero, so \(\gamma_Y(k)=0\).
Therefore the ACF is \(\rho_Y(1)=-9/18=-0.5\), with \(\rho_Y(k)=0\) for \(k\geq2\). The variance is positive, and the lag-one covariance is negative because \(Y_t\) contains \(X_{t-1}\) with a minus sign while \(Y_{t-1}\) contains it with a plus sign.
Check. The differenced series is \(Y_t=X_t-X_{t-1}\), a moving-average process with coefficient \(-1\) under the convention \(Y_t=\varepsilon_t+\theta\varepsilon_{t-1}\). The resulting lag-one ACF is \(-1/(1+1)=-0.5\), matching the covariance calculation. The common mistake is to describe this as evidence that the original white-noise series was nonstationary; it was stationary before differencing.
How the Original ACF Affects the Clue
The white-noise result is a useful benchmark, not a universal value. For example, suppose \(X_t\) is a stationary AR(1) process with mean zero, \(X_t=\phi X_{t-1}+\varepsilon_t\), where \(|\phi|<1\) and \(\varepsilon_t\) is white noise with positive finite variance. Its ACF is \(\rho_X(k)=\phi^k\) for \(k\geq0\). If \(\phi\) is positive, the original series has positive persistence; differencing it can still produce a negative lag-one ACF.
Worked Example: Differencing a Persistent AR(1) Series
Assume \(X_t\) is a stationary AR(1) process with \(\phi=0.6\), mean zero, and \(\gamma_X(0)=25\). Find the first two autocorrelations of \(Y_t=\Delta X_t\), then interpret the pattern.
Method. Use the AR(1) autocovariances \(\gamma_X(k)=25(0.6)^k\), followed by the difference formulas. The required values are \(\gamma_X(1)=25(0.6)=15\), \(\gamma_X(2)=25(0.36)=9\), and \(\gamma_X(3)=25(0.216)=5.4\).
The variance of the differences is \(\gamma_Y(0)=2(25-15)=20\). The lag-one covariance is \(\gamma_Y(1)=2(15)-25-9=30-34=-4\). Thus \(\rho_Y(1)=-4/20=-0.2\).
The lag-two covariance is \(\gamma_Y(2)=2(9)-15-5.4=18-20.4=-2.4\). Thus \(\rho_Y(2)=-2.4/20=-0.12\).
Check. The lag-two result is also \(\phi\rho_Y(1)=0.6(-0.2)=-0.12\). The negative lag-one value is less extreme than the white-noise benchmark because the original series was positively persistent. Here the population ACF remains negative at later lags and decays; overdifferencing need not produce a single isolated negative spike.
More generally, for this AR(1) case the formulas give \(\rho_Y(k)=-(1-\phi)\phi^{k-1}/2\) for \(k\geq1\), when \(|\phi|<1\). This result follows by substituting \(\gamma_X(k)=\gamma_X(0)\phi^k\) into the covariance formula and dividing by \(2\gamma_X(0)(1-\phi)\). It illustrates why the pattern depends on the original series: the white-noise benchmark is only one possible pattern.
Reading a Sample ACF Without Overclaiming
An observed ACF is calculated from a finite sample, so it will not generally equal the population ACF. For a series \(y_1,\ldots,y_n\) with sample mean \(\bar y\), one common sample autocorrelation at lag \(k\) is \(r_k=\sum_{t=k+1}^{n}(y_t-\bar y)(y_{t-k}-\bar y)\big/\sum_{t=1}^{n}(y_t-\bar y)^2\). For a rough reference under a white-noise null, individual sample autocorrelations are often compared with \(\pm1.96/\sqrt{n}\). This is an approximate reference, not a proof or a universal test; model fitting, dependence, and multiple lags affect its interpretation.
Worked Example: A Negative Spike in a Sample ACF
A hypothetical monthly series has been differenced, leaving \(n=64\) observations. Its sample ACF is \(r_1=-0.46\), \(r_2=0.08\), \(r_3=-0.04\), and \(r_4=0.12\). Use the approximate white-noise reference bounds to assess the pattern and state what can and cannot be concluded.
Method. Calculate the reference bound, then compare the lag-one value with both the lower and upper bounds. The bound is \(1.96/\sqrt{64}=1.96/8=0.245\), so the approximate interval is from \(-0.245\) to \(0.245\).
Since \(-0.46<-0.245\), the lag-one sample autocorrelation is more negative than the approximate lower bound. The other supplied values satisfy \(|0.08|<0.245\), \(|-0.04|<0.245\), and \(|0.12|<0.245\). This pattern—a pronounced negative first lag with later sample autocorrelations closer to zero—is compatible with overdifferencing.
Check. The comparison is against the negative bound as well as the positive one. A common mistake is to check whether \(-0.46>0.245\), conclude that it is not unusual, and ignore its magnitude and sign. The corrected comparison shows that it falls below the lower bound. Even so, the calculation does not establish overdifferencing: the bounds are approximate, and a different process can produce a similar ACF.
Why the Clue Is Not Conclusive
A negative lag-one ACF is not unique to overdifferencing. To see this, consider a stationary MA(1) model \(Z_t=\varepsilon_t+\theta\varepsilon_{t-1}\), where \(\varepsilon_t\) is white noise with variance \(\sigma^2>0\). Its variance is \((1+\theta^2)\sigma^2\), its lag-one covariance is \(\theta\sigma^2\), and its autocovariances are zero at lags greater than one. Therefore \(\rho_Z(1)=\theta/(1+\theta^2)\). With \(\theta=-0.5\), this gives \(-0.5/1.25=-0.4\), despite the model being a stationary MA(1), not necessarily an unnecessarily differenced series.
The exact white-noise-difference benchmark, \(\rho(1)=-0.5\) and zero at higher population lags, is also a particular boundary case of moving-average behavior. A sample ACF will rarely display that exact pattern, and other stationary processes can have negative lag-one autocorrelation. These facts prevent a rule such as “a negative lag-one spike proves overdifferencing.”
Common Interpretation Errors
- Ignoring the sign. A negative value must be compared with the lower reference bound, not only with the positive upper bound.
- Calling every negative spike overdifferencing. A stationary MA process, among other possibilities, can also have negative lag-one autocorrelation.
- Expecting exactly \(-0.5\) every time. That is the population lag-one value when white noise is differenced, not a universal value for differenced stationary processes or finite samples.
- Treating later-lag sample values as exact zeros. Even if the population ACF cuts off, sample autocorrelations fluctuate.
- Calling an approximate bound a conclusive test. The usual \(\pm1.96/\sqrt{n}\) reference is a rough guide under a white-noise null, not a general guarantee after model selection or repeated inspection of many lags.
Check Your Understanding
For each problem, distinguish the ACF calculation from the conclusion the evidence supports.
- A white-noise series has variance \(4\). It is differenced once. Calculate the variance of the difference and its lag-one ACF. (Answer: Variance \(=8\); lag-one ACF \(=-0.5\).)
- A stationary AR(1) series has \(\phi=0.2\) and \(\gamma_X(0)=10\). Calculate the lag-one ACF after first differencing. (Answer: \(-0.4\).)
- A differenced series has \(n=100\) observations and sample \(r_1=-0.31\). Using \(\pm1.96/\sqrt{n}\), calculate the approximate bounds and say whether the spike is beyond the lower bound. (Answer: Bounds are \(\pm0.196\); yes, \(-0.31<-0.196\), so it is beyond the lower bound and is a possible overdifferencing clue, not proof.)
- A stationary MA(1) model has \(\theta=-1\) under the convention \(Z_t=\varepsilon_t+\theta\varepsilon_{t-1}\). Calculate its lag-one ACF and explain why that value alone cannot prove overdifferencing. (Answer: \(-1/(1+1)=-0.5\); a negative lag-one ACF can arise from moving-average behavior, so it is not unique evidence of overdifferencing.)