Unit Roots and the ADF Test

A unit root is the precise statement of what makes a price series non-stationary: it wanders with no tendency to return, its variance grows without bound, and shocks to it are permanent. Distinguishing a unit-root (random-walk) series from a stationary or trend-stationary one is not academic hair-splitting, it determines whether you difference or detrend, whether a regression between two series is meaningful or nonsense, and whether a "mean-reverting spread" is real. The augmented Dickey–Fuller test is the standard instrument, and the spurious-regression phenomenon is the expensive lesson behind why it matters.

The random walk and the unit root

Consider the AR(1) $y_t = \rho\,y_{t-1} + \varepsilon_t$. The reciprocal characteristic root is $\rho$, and stationarity requires $|\rho| < 1$. The boundary case $\rho = 1$,

$y_t = y_{t-1} + \varepsilon_t \quad\Longleftrightarrow\quad \Delta y_t = \varepsilon_t,$

is the random walk, a unit root. Its properties are qualitatively different from a stationary AR(1):

• Permanent shocks. Iterating, $y_t = y_0 + \sum_{s=1}^t \varepsilon_s$: every shock is fully absorbed into the level forever; there is no mean reversion. (A stationary AR(1) decays shocks at rate $\rho^k$.)
• Variance grows with time. $\operatorname{Var}(y_t) = t\sigma^2 \to \infty$. The series has no fixed variance to estimate, weak stationarity fails.
• No unconditional mean to revert to. The best forecast of $y_{t+h}$ is just $y_t$ (a martingale), whatever the horizon.

Such a series is integrated of order one, $I(1)$: its first difference $\Delta y_t = \varepsilon_t$ is stationary $I(0)$. This is why prices ($I(1)$) are modeled through returns ($I(0)$). See Stationarity.

Stochastic versus deterministic trend

Both a random walk with drift, $y_t = \delta + y_{t-1} + \varepsilon_t$, and a trend-stationary process, $y_t = \delta t + u_t$ with stationary $u_t$, look like upward-drifting lines. But they demand opposite remedies:

• Difference-stationary (unit root): the trend is stochastic; the right transform is differencing. Detrending by regression leaves the series non-stationary.
• Trend-stationary: the trend is deterministic; the right transform is regressing out the trend. Differencing over-differences it, inducing a non-invertible MA unit root.

Choosing wrong leaves non-stationarity in or destroys real structure, which is exactly what a unit-root test is for.

The Dickey–Fuller test

Test $H_0: \rho = 1$ (unit root) against $H_1: |\rho| < 1$ (stationary). Rewrite the AR(1) by subtracting $y_{t-1}$:

$\Delta y_t = (\rho - 1)\,y_{t-1} + \varepsilon_t = \gamma\,y_{t-1} + \varepsilon_t,$

so the unit-root null becomes $H_0: \gamma = 0$ against $H_1: \gamma < 0$. You run this regression and look at the $t$-statistic on $\gamma$. The crucial subtlety: under the null the regressor $y_{t-1}$ is itself non-stationary, so the $t$-statistic does not follow a Student-$t$ or normal distribution. Its distribution is non-standard (a functional of Brownian motion), so you must compare against the special Dickey–Fuller critical values, which are more negative than the usual ones (about $-2.86$ at 5% with a constant, versus $-1.65$). Using the normal table here is a classic error that massively over-rejects. The augmented Dickey–Fuller (ADF) test adds lagged differences $\sum_i \delta_i \Delta y_{t-i}$ to soak up serial correlation in $\varepsilon_t$ so the residuals are white:

$\Delta y_t = \alpha + \beta t + \gamma\,y_{t-1} + \sum_{i=1}^k \delta_i\,\Delta y_{t-i} + \varepsilon_t.$

You choose whether to include the constant $\alpha$ and trend $\beta t$ based on the series; the critical values differ for each specification. The KPSS test flips the null (stationary under $H_0$) and is often reported alongside ADF for a confirmatory cross-check, since ADF has notoriously low power.

The spurious-regression trap

This is why unit roots matter to a trader. Granger and Newbold showed that regressing one independent random walk on another unrelated random walk produces, with high probability, a large $R^2$ and a hugely significant $t$-statistic on the slope, a completely fake relationship. Two series that share nothing but a stochastic trend will look strongly "related" because both wander persistently. The tell is a very low Durbin–Watson (strongly autocorrelated residuals) alongside the high $R^2$. Regressing price levels on price levels is exactly this trap: correlations of levels are near-meaningless. The disciplines that avoid it: regress returns (differenced, $I(0)$ series) on returns, or, when the levels genuinely move together, test for Cointegration, the one legitimate way to regress $I(1)$ on $I(1)$.

Worked example

Take two independent simulated random walks $x_t, y_t$ (no relationship whatsoever). Regress $y_t$ on $x_t$ in levels: you routinely get $R^2 > 0.5$ and $|t| > 4$ on the slope, spuriously "significant." Now first-difference both and regress $\Delta y_t$ on $\Delta x_t$: the slope is insignificant and $R^2 \approx 0$, correctly revealing no relationship. Same data, opposite conclusions, the levels regression was reading shared wandering as association. Running ADF on each level fails to reject the unit root (as it should), warning you not to regress the levels in the first place.

Failure modes in financial data

• Low power. ADF struggles to distinguish a true unit root from a stationary process with $\rho$ close to 1 (e.g. $\rho = 0.98$), exactly the persistent-but-mean-reverting case a pairs trader cares about. Short samples make this worse; pair ADF with KPSS and economic judgment.
• Structural breaks masquerade as unit roots. A stationary series with a level shift is often mistaken by ADF for $I(1)$; break-robust tests (Zivot–Andrews) are needed.
• Over-differencing. Differencing a trend-stationary or already-stationary series introduces a non-invertible MA and inflates variance, the mirror-image mistake.
• Fat tails and volatility clustering distort the finite-sample null distribution of the ADF statistic; the tabulated critical values assume homoskedastic Gaussian errors.

In interviews

Define a unit root via $\rho = 1$ in an AR(1) and list the three consequences, permanent shocks, variance growing as $t\sigma^2$, and the martingale forecast $\hat y_{t+h} = y_t$. Set up the ADF regression as $\Delta y_t = \gamma y_{t-1} + \dots$ testing $\gamma = 0$, and, the point that separates the prepared candidate, stress that the test statistic has a non-standard (Dickey–Fuller) distribution, not Student-$t$, because $y_{t-1}$ is non-stationary under the null, so you must use special critical values. Explain the difference-stationary vs trend-stationary remedies (difference vs detrend). The spurious-regression story is the applied payoff: regressing $I(1)$ on $I(1)$ manufactures fake significance, which is why we model returns and test for cointegration before ever regressing price levels.