Augmented Dickey-Fuller Test: Is Your Spread Stationary?

Key Takeaways

• The ADF test's null hypothesis is that a series has a unit root (non-stationary); rejecting it means the series is stationary and mean-reverting, which is the prerequisite for pairs and spread trading.
• A semiconductor pairs spread over 750 days produced an ADF statistic of minus 3.86, rejecting the unit-root null at 1 percent and statistically supporting a mean-reverting pairs trade on that spread.
• The most common mistake is choosing the wrong deterministic specification, including a trend term when none exists costs test power, while omitting a genuine trend biases the test toward non-rejection.
• When ADF is applied to Engle-Granger cointegration residuals, standard critical values do not apply; use MacKinnon's tabulated critical values for cointegration tests, which are more negative.

What It Is

The augmented Dickey-Fuller unit root test extends the original Dickey-Fuller test from 1979 to handle autocorrelated errors. Its null hypothesis is that the series has a unit root (non-stationary). The alternative is that the series is stationary. You compute a test statistic, compare it to tabulated critical values, and reject or fail to reject the null.

The test comes in three forms: no constant and no trend, constant only, and constant plus linear trend. The right choice depends on what you think the deterministic part of the series looks like.

The Intuition

A random walk has no memory of its past level. Shocks are permanent. A stationary series does have a memory: it drifts back toward its mean after a shock. Most trading strategies that bet on reversion (pairs trades, spread trades, mean-reverting indicators) require stationarity to work.

The ADF test gives you a disciplined answer rather than an eyeball judgment. A price series of a single stock almost always fails to reject the unit-root null. A spread between two cointegrated stocks should reject. If it does not, you do not have a mean-reverting relationship worth trading.

How It Works

Start from an AR(p) representation of the series y_t. The ADF regression is:

dy_t = alpha + beta * t + gamma * y_{t-1} + sum_{i=1..p-1} phi_i * dy_{t-i} + e_t

where dy_t = y_t - y_{t-1}. The key quantity is the coefficient gamma. If the series is stationary, gamma < 0 and the test statistic has a non-standard distribution tabulated by Dickey, Fuller, and later MacKinnon. The ADF statistic is:

ADF_stat = gamma_hat / SE(gamma_hat)

You reject the unit-root null if ADF_stat is more negative than the critical value at your chosen significance level. Typical 5 percent critical values are roughly -2.86 (constant only) or -3.41 (constant plus trend). The lag order p is chosen to whiten residuals, commonly by the Akaike or Bayesian information criterion.

Worked Example

You have the daily log spread between two semiconductor stocks over 750 trading days. The spread oscillates between -0.08 and +0.06 with an apparent mean near 0. Fit the ADF regression with a constant and no trend. Information criteria pick p = 3 lags.

Suppose the fitted gamma_hat is -0.085 with standard error 0.022. Then:

ADF_stat = -0.085 / 0.022 = -3.86

The 5 percent critical value with constant only is -2.86 and the 1 percent value is -3.43. Since -3.86 is below both, you reject the unit-root null at 1 percent. The spread is stationary and mean-reverting by this test. A pairs trade built on this spread has statistical support.

By contrast, running ADF on either stock's log price series alone would typically yield a statistic near -1 to -2, far above the critical value, so you fail to reject and treat each price as a random walk.

Common Mistakes

1. Choosing the wrong deterministic terms. Including a trend when none exists reduces power. Omitting a trend when one exists biases the test toward non-rejection. Plot the series first and pick the specification that matches what you see.

2. Under-lagging. If residuals of the ADF regression are still autocorrelated, the distribution of the test statistic is wrong. Use an information criterion or inspect residual autocorrelations to pick p.

3. Ignoring structural breaks. A series that is stationary around two different means in two sub-periods will often fail to reject the unit-root null under ADF because the break looks like a permanent shock. Perron's modified tests or a Bai-Perron break analysis are needed when a break is plausible.

4. Treating failure to reject as proof of a unit root. Not rejecting the null is not the same as accepting it. The test can have low power, especially in small samples. Combining ADF with KPSS (null is stationarity) gives a more balanced read.

5. Forgetting that ADF critical values differ for cointegration residuals. When you apply ADF to an estimated regression residual (Engle-Granger step two), the critical values are more extreme than the standard ADF table. Use MacKinnon's tabulated values for cointegration tests.

Frequently Asked Questions

Q: What is the augmented Dickey-Fuller unit root test in simple terms? The ADF test asks a single yes-or-no question: does this time series behave like a random walk, with shocks that are permanent, or does it mean-revert? If the test rejects the unit-root null, the series is stationary and eligible for mean-reversion trading strategies.

Q: How does the ADF test affect investment decisions? It provides the statistical gatekeeper for spread and pairs trades. No matter how visually appealing a spread looks, the ADF test must reject the unit-root null before committing capital. A spread that fails the test can trend against you indefinitely, making position sizing and stop placement impossible to calibrate.

Q: What is a real-world example of the ADF test in trading? A semiconductor pairs spread is tested with the constant-only ADF specification, picking three lags by information criteria. The ADF statistic comes in at minus 3.86, below both the 5 percent critical value of minus 2.86 and the 1 percent value of minus 3.43. The spread is stationary at 1 percent confidence and eligible as a pairs trade. Running the same test on either stock's log price alone produces a statistic near minus 1.5, far from rejection, confirming each price is a random walk.

Q: How can investors choose the right ADF specification? Plot the series before testing and choose the deterministic terms that match what you see. If the series has no clear trend, use a constant only. If it trends upward or downward, include a linear trend term. Using the wrong specification changes both the critical values and the test's power.

Q: How is the ADF test different from the KPSS test? The ADF test has stationarity as the alternative hypothesis and non-stationarity (unit root) as the null; failing to reject means you cannot rule out a random walk. The KPSS test reverses this: its null is stationarity. Using both together gives a more complete picture, if ADF fails to reject and KPSS also fails to reject, the evidence is genuinely ambiguous about the series' true nature.

Sources

1. Dickey, D.A. and Fuller, W.A. (1979). "Distribution of the Estimators for Autoregressive Time Series with a Unit Root." JASA 74(366), 427-431. https://www.tandfonline.com/doi/abs/10.1080/01621459.1979.10482531
2. Said, S.E. and Dickey, D.A. (1984). "Testing for Unit Roots in Autoregressive-Moving Average Models of Unknown Order." Biometrika 71(3), 599-607. https://academic.oup.com/biomet/article-abstract/71/3/599/256788
3. MacKinnon, J.G. (2010). "Critical Values for Cointegration Tests." Queen's Economics Department Working Paper No. 1227. https://www.econ.queensu.ca/sites/econ.queensu.ca/files/qed_wp_1227.pdf
4. Hamilton, J.D. (1994). Time Series Analysis. Princeton University Press, Chapter 17. https://press.princeton.edu/books/hardcover/9780691042893/time-series-analysis

Disclaimer

This article is educational content only and is not financial advice. Nothing here is a recommendation to buy, sell, or hold any security. Consult a licensed advisor before making investment decisions.