# If a time series is second order stationary, does this imply it is strictly stationary?

A process $X_t$ is strictly stationary if the joint distribution of $X_{t_1},X_{t_2},...,X_{t_m}$is the same as the joint distribution of $X_{t_1+k},X_{t_2+k},...,X_{t_m+k}$ for all $m$, for all $k$ and for all $t_1,t_2,...,t_m$.

A process is second order stationary if its mean is constant and its autocovariance function depends only on the lag.

Therefore does second order stationary imply strict stationary?

Also under second order stationary it says that no assumptions are made about higher moments than those of first and second order. First moment corresponds to the mean, does the second moment correspond to the autocovariance?

• See also this post for a related discussion. Oct 13 '14 at 9:08
• What you call (or your course calls) second-order stationary is often called weakly stationary or wide-sense-stationary (WSS) or stationary in the wide sense. WSS processes are not necessarily strictly stationary because the mean and autocovariance are not, in general, enough to determine the distribution. Of course, a WSS Gaussian or normal process (meaning all the $X_t$ are normal random variables) is strictly stationary because the mean and covariance matrix determine the joint distribution. Oct 29 '14 at 21:32
• See also Example of a process that is 2nd order stationary but not strictly stationary. The two are very close to being duplicates. This question also asks about whether second moment refers to autocovariance, but that is really a sub-question and is at any rate handled on the thread What is a second order stationary process? Sep 29 '15 at 23:19
• mathoverflow.net/questions/42141/… Oct 7 '15 at 12:38

Second order stationarity is weaker than strict stationarity. Second order stationarity requires that first and second order moments (mean, variance and covariances) are constant throughout time and, hence, do not depend on the time at which the process is observed. In particular, as you say, the covariance depends only on the lag order, $k$, but not on the time at which it is measured, $Cov(x_t, x_{t-k}) = Cov(x_{t+h}, x_{t+h-k})$ for all $t$.

In a strict stationarity process, the moments of all orders remain constant throughout time, i.e., as you say, the joint distribution of $X_{t1},X_{t2},...,X_{tm}$ is the same as the joint distribution of $X_{t1+k}+X_{t2+k}+...+X_{tm+k}$ for all $t1,t2,...,tm$ and $k$.

Therefore, strict stationarity involves second order stationarity but the converse is not true.

Edit (edited as answer to @whuber's comment)

The previous statement is the general understanding of weak and strong stationarity. Although the idea that stationarity in the weak sense does not imply stationarity in a stronger sense may agree with intuition, it may not be so straightforward to proof, as pointed out by whuber in the comment below. It can be helpful to illustrate the idea as suggested in that comment.

How could we define a process that is second-order stationary (mean, variance and covariance constant throughout time) but it is not stationary in strict sense (moments of higher order depend on time)?

As suggested by @whuber (if I understood correctly) we can concatenate batches of observations coming from different distributions. We just need to be careful that those distributions have the same mean and variance (at this point let's consider that they are sampled independently of each other). On one hand, We can for example generate observations from the Student's $t$-distribution with $5$ degrees of freedom. The mean is zero and the variance is $5/(5-2)=5/3$. On another hand, we can take the Gaussian distribution with zero mean and variance $5/3$.

Both distributions share the same mean (zero) and variance ($5/3$). Thus, the concatenation of random values from these distribution will be, at least, second-order stationary. However, the kurtosis at those points governed by the Gaussian distribution will be $3$, while at those time points where the data come from the Student's $t$-distribution it will be $3+6/(5-4)=9$. Therefore, the data generated in this way are not stationary in strict sense because moments of fourth order are not constant.

The covariances are also constant and equal to zero, since we considered independent observations. This may seem trivial, so we can create some dependence among observations according the following autoregressive model.

$$y_t = \phi y_{t-1} + \epsilon_t \,, \quad |\phi| < 1 \,, \quad t = 1,2,...,120$$ with \begin{eqnarray} \epsilon_t \sim \left\{ \begin{array}{ll} N(0, \sigma^2=5/3) \quad & \hbox{if} \; t \in [0,20], [41,60], [81,100] \\ t_5 \quad & \hbox{if} \; t \in [21,40], [61,80], [101,120] \,. \end{array} \right. \end{eqnarray}

$|\phi| < 1$ ensures that second-order stationarity is satisfied.

We can simulate some of these series in the R software and check whether the sample mean, variance, first order covariance and kurtosis remain constant across batches of $20$ observations (the code below uses $\phi=0.8$ and sample size $n=240$, the Figure displays one of the simulated series):

# this function is required below
kurtosis <- function(x)
{
n <- length(x)
m1 <- sum(x)/n
m2 <- sum((x - m1)^2)/n
m3 <- sum((x - m1)^3)/n
m4 <- sum((x - m1)^4)/n
b1 <- (m3/m2^(3/2))^2
(m4/m2^2)
}
# begin simulation
set.seed(123)
n <- 240
Mmeans <- Mvars <- Mcovs <- Mkurts <- matrix(nrow = 1000, ncol = n/20)
for (i in seq(nrow(Mmeans)))
{
eps1 <- rnorm(n = n/2, sd = sqrt(5/3))
eps2 <- rt(n = n/2, df = 5)
eps <- c(eps1[1:20], eps2[1:20], eps1[21:40], eps2[21:40], eps1[41:60], eps2[41:60],
eps1[61:80], eps2[61:80], eps1[81:100], eps2[81:100], eps1[101:120], eps2[101:120])
y <- arima.sim(n = n, model = list(order = c(1,0,0), ar = 0.8), innov = eps)

ly <- split(y, gl(n/20, 20))
Mmeans[i,] <- unlist(lapply(ly, mean))
Mvars[i,] <- unlist(lapply(ly, var))
Mcovs[i,] <- unlist(lapply(ly, function(x)
acf(x, lag.max = 1, type = "cov", plot = FALSE)$acf[2,,1])) Mkurts[i,] <- unlist(lapply(ly, kurtosis)) } The results are not what I expected: round(colMeans(Mmeans), 4) #  0.0549 -0.0102 -0.0077 -0.0624 -0.0355 -0.0120 0.0191 0.0094 -0.0384 #  0.0390 -0.0056 -0.0236 round(colMeans(Mvars), 4) #  3.0430 3.0769 3.1963 3.1102 3.1551 3.2853 3.1344 3.2351 3.2053 3.1714 #  3.1115 3.2148 round(colMeans(Mcovs), 4) #  1.8417 1.8675 1.9571 1.8940 1.9175 2.0123 1.8905 1.9863 1.9653 1.9313 #  1.8820 1.9491 round(colMeans(Mkurts), 4) #  2.4603 2.5800 2.4576 2.5927 2.5048 2.6269 2.5251 2.5340 2.4762 2.5731 #  2.5001 2.6279  The mean, variance and covariance are relatively constant across batches as expected for a second-order stationary process. However, the kurtosis remains relatively constant as well. We could have expected higher values of the kurtosis at those batches related to draws from the Student's$t$-distribution. Maybe$20$observations is not enough to capture changes in kurtosis. If we didn't know the data generating process of these series and we looked at rolling statistics, we would probably conclude that the series is stationary at least up to order fourth. Either I didn't take the right example or some features of the series get masked for this sample size. • Although you are correct, you have not adequately demonstrated the final conclusion. (You seem to be presuming that the higher moments of a second-order stationary process can be prescribed independently of its first two moments, but that--although partly true--is not obvious.) The strongest way to demonstrate your conclusion would be to exhibit a process which is second-order stationary but not stationary. Although that is easy to do with a suitable sequence of independent random variables, it would be of interest to provide an example with non-vanishing correlations at all lags. – whuber Oct 13 '14 at 16:40 • @whuber I edited my answer. I thought I understood your point but my attempt to follow your idea was not fully satisfactory. Oct 13 '14 at 22:02 • Maybe this will help. Let$U_i,i=0,1$be independent Bernoulli variables with parameters$p\ne 1/2$and$1-p$, respectively, and let$(X_i)$be a sequence of iid Normal variables,$i\in\mathbb{Z}$. Define$Y_i=U_{[i]}-p_{[i]}+X_i$where$[i]=0$when$i$is even and$[i]=1\$ otherwise. Serial correlation is high, it is second-order stationary, but it is not stationary and not ergodic. You can generate one realization in R with code like n <- 300; p <- 1/4; x <- rnorm(n, (rbinom(2,1,c(p,1-p))-c(p,1-p)), 1/8). Running and plotting several such simulations is instructive.
– whuber
Oct 13 '14 at 23:13
• I wouldn't order strict stationarty and covariance- stationarity (although the use of the term "weak" also for the latter unfortunately points towards such ordering). The reason is that strict stationarity does not imply covariance-stationarity: the process may be strictly stationary but distribution moments may not exist or be infinite, in which case this strictly stationary process is not covariance-stationary. Oct 15 '14 at 15:48
• We cannot directly simulate non-existence of moments. Create a Cauchy strictly stationary process, to take the trivial example. The graph will look perfectly "stationary", because the behavior of the process is repetitive, a behavior that depends on the moments only when they exist. If they do not exist, then the behavior is described and depends on other characteristics of the distribution. Oct 15 '14 at 16:03

Since I can't comment, and I have a worthwhile caveat to @javlacalle's answer, I am forced to include this is a separate answer:

@javlacalle wrote that

strict stationarity involves second order stationarity but the converse is not true.

However, strong stationarity does not imply weak stationarity. The reason is that strong stationarity does not mean the process necessarily has a finite second moment. For example, an iid process with standard Cauchy distribution is strictly stationary but has no finite second moment. Indeed, having a finite second moment is a necessary and sufficient condition for the weak stationarity of a strongly stationary process.

Reference: Myers, D.E., 1989. To be or not to be . . . stationary? That is the question. Math. Geol. 21, 347–362.