A unit root is a property of a time series where the data has no natural tendency to return to a long-run average. When a series has a unit root, shocks or random changes accumulate over time rather than fading away, causing the series to wander unpredictably. This makes the data “non-stationary,” which breaks many of the standard tools used in statistics and forecasting.
The Random Walk Analogy
The simplest way to understand a unit root is through a random walk. Imagine someone standing on a number line who flips a coin each second: heads, step right; tails, step left. Their position at any moment is just their previous position plus or minus one step. There’s no force pulling them back to where they started. Over time, they can drift arbitrarily far in either direction.
In mathematical terms, a random walk looks like this: today’s value equals yesterday’s value plus some random noise. The coefficient on yesterday’s value is exactly 1, and that “1” is the unit root. If the coefficient were something smaller, say 0.8, each new value would partially revert toward zero. The series would be stationary, fluctuating around a stable average. But when the coefficient is 1, there is no reversion. Every random shock becomes a permanent part of the series.
Why It Makes Data Non-Stationary
A stationary time series has a constant mean and a stable variance over time. A unit root destroys both properties. Because shocks accumulate rather than decay, the average of the series drifts. The variance also grows without bound as time passes. If you tried to forecast far into the future, your uncertainty would eventually stretch from plus infinity to minus infinity, making the prediction useless.
This matters because most regression and forecasting methods assume the data is stationary. When you run a standard regression on two unrelated series that both have unit roots, you can get results that look statistically significant even though the relationship is completely meaningless. The t-statistics and R-squared values don’t behave normally. Instead of settling toward fixed distributions as you collect more data, they actually diverge, growing larger and more misleading the more data you add. This problem, known as spurious regression, was one of the major motivations for developing unit root tests in the first place.
Permanent Versus Temporary Shocks
One of the most important practical consequences of a unit root is how the series responds to shocks. In a stationary process, a sudden disturbance (an unexpected event, a policy change, a market surprise) gradually fades. The series returns to its long-run path. In a unit root process, the shock never fades. It shifts the entire future trajectory of the series permanently.
This distinction has real economic implications. Research from the Federal Reserve found that permanent shocks to housing values had a significantly greater effect on household spending than temporary ones. If home prices follow a unit root process, a downturn doesn’t just temporarily reduce wealth. It resets the baseline. The same logic applies to GDP, stock prices, exchange rates, and interest rates, all of which are commonly tested for unit roots in economic research.
How to Test for a Unit Root
Since you can’t tell by looking at a chart whether a series has a unit root or is just trending, formal statistical tests are essential. The two most widely used are the Augmented Dickey-Fuller (ADF) test and the KPSS test, and they approach the question from opposite directions.
The ADF test starts by assuming the series has a unit root (the null hypothesis) and asks whether the data provides enough evidence to reject that assumption. Specifically, it estimates a coefficient that captures how strongly the series reverts to a previous level. If that coefficient is statistically less than zero, you reject the unit root hypothesis and conclude the series is stationary. Importantly, the ADF test uses special critical values rather than the standard ones from a normal distribution, because the usual statistical tables don’t apply when a unit root might be present.
The KPSS test flips the logic. Its null hypothesis is that the series is stationary, and it looks for evidence of a unit root. Running both tests together gives a more complete picture. If the ADF rejects a unit root and the KPSS fails to reject stationarity, you have strong evidence the series is stationary. If both tests point the other way, a unit root is likely. Conflicting results suggest the data is ambiguous, perhaps nearly non-stationary or affected by structural breaks.
Choosing Lag Length
One practical wrinkle in ADF testing is deciding how many lagged terms to include. Too few lags can distort the test’s size, making it reject too often or not often enough. Researchers typically use information criteria like the AIC or BIC to select the number of lags, but these standard criteria tend to choose lag lengths that are too short when the data has certain error structures. Modified versions of these criteria, such as the Modified AIC, account for the bias that arises specifically in unit root testing and generally produce more reliable results.
Fixing the Problem With Differencing
If a series has a unit root, the standard remedy is differencing: instead of analyzing the raw values, you analyze the change from one period to the next. For a random walk, the first difference is just the random noise term, which is stationary. All the accumulated history cancels out.
A series that becomes stationary after taking one difference is called “integrated of order 1,” written as I(1). Occasionally, a series needs to be differenced twice (I(2)), though this is less common in practice. Once you’ve differenced the data to stationarity, you can apply the full toolkit of time series methods: autoregressive models, moving average models, standard regressions, and reliable forecast intervals.
Failing to difference when you should has two concrete consequences. First, regressions on the raw data tend to produce spuriously significant trends, making it look like there’s a meaningful pattern when there isn’t one. Second, long-run forecast intervals will be far too narrow, giving a false sense of certainty about where the series is headed. Differencing costs you one observation and removes the level of the series, but it protects against both of these problems.
Unit Root Versus Trend Stationarity
Not every series that trends upward has a unit root. A series can follow a deterministic trend, steadily rising over time, while still being stationary around that trend. The distinction matters because the correct treatment is different. A trend-stationary series should be de-trended by regressing on time and analyzing the residuals. A unit root series should be differenced. Applying the wrong procedure gives unreliable results in both directions.
The two types of series can look nearly identical on a chart, especially over short time spans. The key difference is under the surface: in a trend-stationary process, deviations from the trend line are temporary. In a unit root process, there is no fixed trend line to deviate from. The series is free to wander, and today’s random shock permanently alters every future value. This is why formal testing, rather than visual inspection, is necessary to tell them apart.