Stationarity and Unit Roots

Graduate Depth 74 in the knowledge graph I know this Set as goal
Unlocks 4 downstream topics
stationarity unit-root ADF-test I(1) random-walk

Core Idea

A time series is (weakly) stationary if its mean, variance, and autocovariances do not depend on time. Many economic series — GDP levels, price indices, exchange rates — are non-stationary: they have stochastic trends, drifting means, and growing variance. A random walk y_t = y_{t−1} + ε_t has a 'unit root' and is integrated of order 1 (I(1)); its first difference Δy_t = ε_t is stationary. The Augmented Dickey-Fuller (ADF) test formally tests for unit roots. Regressing one I(1) series on another without cointegration produces spurious results; the standard remedy is to work in differences.

How It's Best Learned

Apply the ADF test to GDP levels and then to GDP growth rates — levels typically fail the test (unit root not rejected) while growth rates pass. Simulate a random walk and AR(1) with ρ<1 to see the difference visually.

Common Misconceptions

Explainer

Your time series background gives you the tools to model how economic variables evolve over time. The next essential question is whether a series behaves consistently over time — whether its statistical properties are stable or drifting. A series is weakly stationary if its mean, variance, and autocovariances are all constant over time. Think of coin flip outcomes: no matter when you start recording, the mean hovers at 0.5 and the variance stays fixed. A stationary series has a stable "center of gravity" it keeps returning to after shocks. Many standard results in time series econometrics — the law of large numbers, the central limit theorem — require stationarity to hold. When stationarity fails, those results break down, and so do many standard regression techniques.

The contrast is a random walk: y_t = y_{t−1} + ε_t, where ε_t is white noise. Each period, the series moves by a random shock — and here is the key: the shock is permanent. There is no mean to return to. After a positive shock today, the series simply starts from a higher level and wanders from there. The variance of a random walk grows without bound as time passes (it equals σ²t after t periods), which violates the stationarity requirement of constant variance. This is what it means to be integrated of order 1, or I(1): one differencing operation is needed to produce a stationary series. The first difference Δy_t = y_t − y_{t−1} = ε_t is simply white noise — stationary. GDP levels, stock prices, exchange rates, and many price indices behave like random walks (or near-random walks). GDP growth rates, stock returns, and inflation rates tend to be stationary.

The practical danger of non-stationarity is spurious regression. If you regress one I(1) series on another unrelated I(1) series — say, U.S. GDP on the population of Iceland — you will typically find a high R² and a statistically significant slope coefficient, even though no true relationship exists. Both series are simply trending over time, and OLS interprets their shared trend as a relationship. Your probability theory and random variables background helps here: you know that the sampling distributions of OLS estimates change fundamentally when variables are I(1), invalidating the usual t- and F-test critical values. This is why checking for stationarity before running regressions is not optional.

The Augmented Dickey-Fuller (ADF) test formalizes this check. The null hypothesis is that the series has a unit root (is non-stationary); rejection of the null means stationarity. The ADF regression tests whether the autoregressive coefficient is equal to one by running the transformed regression Δy_t = α + ρ*y_{t−1} + lagged differences + ε_t and testing whether ρ = 0. Note the counterintuitive direction of the test: you need evidence *against* the null (unit root) to conclude stationarity, and failing to reject does not prove the series is non-stationary — it may just mean you have insufficient power. If a series is I(1), the standard remedy is to work in differences: first-differencing removes the stochastic trend, produces a stationary series, and restores the validity of standard inference. The tradeoff is that differencing also removes all long-run level information — the cointegration framework, covered next, recovers long-run relationships without discarding them.

Practice Questions 5 questions

Prerequisite Chain

Counting to 10Counting to 20Understanding ZeroThe Number ZeroCounting to FiveOne-to-One CorrespondenceCombining Small Groups Within 5Addition Within 10Addition Within 20Two-Digit Addition Without RegroupingTwo-Digit Addition with RegroupingAddition Within 100Repeated Addition as MultiplicationMultiplication Facts Within 100Division as Equal SharingDivision as Grouping (Measurement Division)Division: Grouping (Repeated Subtraction) ModelDivision: Fair Sharing ModelDivision as Equal SharingDivision as GroupingBasic Division FactsDivision Facts Within 100Two-Digit by One-Digit DivisionDivision with RemaindersRemainders and Quotients in DivisionDivision Word ProblemsIntroduction to Long DivisionFactors and MultiplesPrime and Composite NumbersEquivalent FractionsRelating Fractions and DecimalsDecimal Place ValueReading and Writing DecimalsComparing and Ordering DecimalsAdding and Subtracting DecimalsMultiplying DecimalsDividing DecimalsDividing FractionsMixed Number ArithmeticOrder of OperationsInteger Order of OperationsVariable ExpressionsCombining Like TermsOne-Step EquationsTwo-Step EquationsSolving Multi-Step EquationsEquations with Variables on Both SidesAngle Pairs: Complementary, Supplementary, and VerticalParallel Lines and TransversalsCorresponding AnglesAlternate Interior AnglesTriangle Angle Sum TheoremExterior Angle TheoremTriangle Inequality TheoremSimilar Triangles: AA SimilaritySimilar Triangles: SSS and SAS SimilarityProportions in Similar TrianglesRight Triangle Trigonometry IntroductionTrigonometric Ratios ReviewRadian MeasureConverting Between Degrees and RadiansThe Unit CircleGraphing Sine and CosineGraphing Tangent and Reciprocal Trigonometric FunctionsDerivatives of Trigonometric FunctionsAntiderivativesIndefinite IntegralsBasic Integration RulesRiemann SumsDefinite Integral DefinitionFundamental Theorem of Calculus Part 1Fundamental Theorem of Calculus Part 2U-SubstitutionSeparable Equations (Intro)Stationarity and Unit Roots

Longest path: 75 steps · 403 total prerequisite topics

Prerequisites (6)

Leads To (1)