Random Walks, Stochastic Trends, And Integration

Why Differencing Matters

Economic time series often move persistently: a shock to an exchange rate, price index, or output measure may continue to affect its level long after the shock occurs. A stationary fluctuation behaves differently: its distribution does not change over time, and shocks do not permanently shift the series. Differencing provides a precise way to distinguish these patterns.

The central example is the random walk. Its changes can be stationary even though its level is not. The level carries forward the cumulative history of its innovations, creating a stochastic trend. Whether the random walk has a nonzero drift changes its expected path, but does not remove the stochastic trend when its innovations have positive variance.

The difference operator is \(\Delta y_t=y_t-y_{t-1}\). Repeated differencing is defined recursively: \(\Delta^2 y_t=\Delta(\Delta y_t)\), and so on. Stationarity here refers to the process implied by a model, not to a visual impression from a short finite record.

Random Walks With and Without Drift

Let \(y_0\) be a fixed initial value, and let the innovations \(\varepsilon_t\) be independent and identically distributed with mean zero and variance \(\sigma^2>0\). The positive-variance condition is important: it ensures that random innovations actually accumulate. A random walk with drift \(\mu\) is

When \(\mu=0\), this is a random walk without drift. When \(\mu\neq0\), the series has an expected upward or downward movement of \(\mu\) units per period, in addition to its stochastic movement. Repeated substitution gives

The first difference is \(\Delta y_t=\mu+\varepsilon_t\). These differences are independent across time, have constant mean \(\mu\), and have variance \(\sigma^2\); therefore they are covariance stationary. The level, however, has expectation \(y_0+\mu t\) and variance \(t\sigma^2\). Its variance changes with time, so the level is nonstationary whether the drift is zero or nonzero. Consequently, under the stated assumptions, the random walk is I(1).

Why the Level Is Nonstationary

The nonstationarity follows directly from the moments of the process. Because the innovations have mean zero and are independent, their sum has mean zero and variance equal to the sum of their variances. Hence

If \(\mu\neq0\), the level mean changes with \(t\). If \(\mu=0\), the mean is constant, but the variance still increases as \(t\sigma^2\), since \(\sigma^2>0\). Either way, covariance stationarity fails. More generally, for \(s\leq t\), the innovations shared by \(y_s\) and \(y_t\) are \(\varepsilon_1,\ldots,\varepsilon_s\), so

This covariance depends on the earlier time \(s\), not only on the lag \(t-s\). It gives another way to see why the level is not covariance stationary. By contrast, the first differences have covariance zero at every nonzero lag because the innovations are independent.

The assumption \(\sigma^2>0\) is not cosmetic. If \(\mu=0\) and \(\sigma^2=0\), every innovation is zero almost surely, so \(y_t=y_0\) for all \(t\). That constant process is I(0), not a stochastic-trend I(1) process. This degenerate case is excluded from the nondegenerate random-walk claims above.

Drift, Stationary Fluctuations, and Integration

A random walk with drift combines a deterministic change in expectation with a stochastic trend. Its increments fluctuate around the constant mean \(\mu\), while its level includes the cumulative innovations. This is different from a stationary fluctuation around a stable level: in a stationary process, shocks may affect later observations, but they do not accumulate as a permanent sum of innovations.

The distinction matters for interpretation. A persistent rise in a series does not, by itself, tell us whether it comes from a deterministic trend, a stochastic trend, or a sequence of stationary fluctuations. The model and its differenced behavior matter. Later analysis may specify deterministic terms explicitly; here, the focus is on the random-walk mechanism and on classifying the stated processes by their differences.

Why the Minimum Number of Differences Matters

A process does not become I(1) merely because its first differences happen to be stationary. A stationary process can also have stationary first differences. Integration order records the minimum number of differences needed, so a process that is already stationary in levels is I(0), even if differencing it leaves a stationary process.

For an I(0) process \(y_t\), the difference \(\Delta y_t\) is a linear combination of two observations of a stationary process, and is itself stationary. For example, its mean is zero if \(y_t\) has a constant mean, and its variance is finite and constant because the autocovariances of \(y_t\) depend only on lag. This is why the minimum-differencing condition is essential: checking only whether \(\Delta y_t\) is stationary cannot distinguish I(0) from I(1).

A Practical Classification Procedure

For a specified stochastic process, classify integration order by checking the level first and then differencing. A concise workflow is:

A finite observed sequence can be differenced arithmetically, but it cannot by itself establish the probability structure of an unobserved process. Integration-order conclusions require assumptions or a statistical analysis of the data; a handful of increasing or decreasing observations is not sufficient.

Common Pitfalls

• Calling a zero-drift random walk stationary. A constant expected level does not imply stationarity. With \(\sigma^2>0\), its variance is \(t\sigma^2\), which changes over time.
• Ignoring the drift. A random walk with drift has differences with mean \(\mu\), not necessarily zero. Stationarity does not require a zero mean; it requires that the mean remain constant.
• Classifying from the first difference alone. The first difference of an I(0) series can also be stationary. Check the level first and use the minimum number of differences.
• Inferring integration from a short plot or path. A finite sequence may look persistent by chance. A model-based classification concerns the process-generating properties, not just the direction of a few observations.