1 Definitions and core concepts
1.1 General meaning
Stationarity is a property of a process whose statistical behavior remains stable over time. In a stationary setting, quantities such as the average level, variability, and dependence across time do not systematically change as the process evolves. This makes the process easier to summarize, compare, and model.
The concept is used as an idealization in many fields. Real data often show trends, cycles, or abrupt shifts, but stationarity offers a useful baseline for describing systems that fluctuate around a stable pattern.
1.2 Stationary random processes
A stationary random process is a family of random variables indexed by time whose probabilistic structure is unchanged under shifts in time. The precise meaning depends on how much of the distribution is required to remain invariant. In practice, the term often refers to processes whose first and second moments are stable.
1.2.1 Strict stationarity
Strict stationarity requires that the joint distribution of any collection of observations be unchanged by a common shift in time. If one takes values at several time points and then moves all those points forward or backward by the same amount, the resulting joint distribution is identical. This is the strongest and most exact form of stationarity.
1.2.2 Weak stationarity
Weak stationarity, also called second-order stationarity in many contexts, requires only that the mean be constant over time and that the covariance between observations depend only on the time gap separating them, not on the absolute time. This weaker condition is often sufficient for many statistical methods and theoretical results.
1.2.3 Covariance stationarity
Covariance stationarity is a common synonym for weak stationarity, especially in time series analysis. It emphasizes the stability of the mean, variance, and autocovariance structure. Many standard forecasting and inference tools are built on this assumption.
1.3 Stationarity in time series
In time series analysis, stationarity describes sequences of observations recorded in time that maintain stable statistical properties. Analysts often distinguish between series that fluctuate around a fixed level and those that show changing mean or variance. Stationarity is important because many estimation procedures assume that past patterns provide reliable information about the future.
1.4 Stationarity in stochastic processes
For stochastic processes more generally, stationarity concerns invariance in time of the process’s distributional features. The idea extends beyond ordinary time series to spatial data, signals, and random fields. In each case, the key question is whether shifting the index changes the statistical description of the process.
2 Mathematical properties
2.1 Invariance under time shifts
The central mathematical feature of stationarity is invariance under translation in time. If the process is stationary, its statistical laws do not depend on the choice of origin. This symmetry means that only relative time differences matter, not the absolute time at which observations are taken.
2.2 Moments and distributional stability
Stationary processes typically have moments that are stable over time. The mean, and often the variance, remain constant. In stronger forms, the entire distribution is unchanged by time shifts. This stability allows analysts to estimate model parameters from observed data and use them consistently across time.
2.3 Autocovariance structure
A stationary process has an autocovariance function that depends only on lag. This gives a compact description of temporal dependence and is one of the most useful consequences of stationarity. The shape of the autocovariance function helps determine how quickly the influence of past values fades.
2.3.1 Lag dependence
Under stationarity, correlation between two observations depends on how far apart they are, not on when they occur. Observations separated by the same lag have the same covariance. This property supports the use of lag-based summaries in modeling and prediction.
2.3.2 Spectral implications
Stationarity is closely tied to spectral analysis. When a process is stationary, its dependence structure can often be represented in terms of frequency components. The spectral density provides an alternative view of the process, showing how variance is distributed across different cycles or oscillations.
2.4 Ergodicity and stationarity
Ergodicity concerns whether time averages computed from one long realization reflect ensemble averages across many realizations. Stationarity and ergodicity are related but distinct. A process may be stationary without being ergodic, although many practical methods rely on both properties. Together, they support statistical inference from a single observed series.
3 Types of stationarity
3.1 Strict stationarity
Strict stationarity requires full distributional invariance under time shifts. It is mathematically strong and conceptually clear, but it can be difficult to verify from data. Many models satisfy strict stationarity only under certain parameter conditions.
3.2 Second-order stationarity
Second-order stationarity requires constant mean and finite, time-invariant second moments. It is widely used because it is easier to work with than strict stationarity while still supporting covariance-based analysis. Most classical linear time series methods are formulated under this assumption.
3.3 Trend-stationarity
A trend-stationary process is not stationary in raw form but becomes stationary after removing a deterministic trend. The fluctuations around the trend are stable, while the long-term trajectory is governed by a predictable pattern. Such processes are often modeled by combining a deterministic component with stationary noise.
3.4 Difference-stationarity
A difference-stationary process becomes stationary after differencing once or more. Instead of oscillating around a fixed mean, the series may contain a stochastic trend. Differencing removes the accumulated movement and reveals a more stable underlying series.
3.4.1 Unit roots
A unit root is a feature of some time series models that causes shocks to have permanent effects. Processes with unit roots are typically difference-stationary rather than stationary in levels. Identifying a unit root is important because it changes how the series should be modeled and interpreted.
3.4.2 Differencing
Differencing subtracts each observation from a previous one, often the immediately preceding value. This operation can eliminate persistent trends and stabilize the mean. In practice, it is one of the simplest tools for transforming a non-stationary series into a stationary one.
4 Testing for stationarity
4.1 Visual diagnostics
Visual inspection is often the first step in assessing stationarity. Plots can reveal changing levels, variance shifts, trends, or seasonal patterns that suggest non-stationary behavior. Although informal, these methods provide useful initial guidance.
4.1.1 Time plots
A time plot shows the series against time and can reveal obvious departures from stationarity. A stable band around a constant level suggests stationarity, while visible drift or widening fluctuations may indicate otherwise. Time plots are especially helpful for identifying qualitative patterns.
4.1.2 Rolling statistics
Rolling statistics compute summary measures over moving windows. By tracking the mean and variance across successive intervals, one can see whether these quantities remain stable. Large changes in rolling summaries often signal non-stationarity.
4.2 Statistical tests
Formal tests provide systematic procedures for evaluating stationarity or related hypotheses. These tests typically examine whether the data are consistent with a unit root or with stable fluctuations around a fixed or deterministic mean. Results are usually interpreted alongside graphical evidence.
4.2.1 Augmented Dickey-Fuller test
The Augmented Dickey-Fuller test examines whether a series contains a unit root. It is one of the most widely used tests in time series analysis. The test includes lagged differences to account for autocorrelation in the errors.
4.2.2 Phillips-Perron test
The Phillips-Perron test also addresses unit roots, but it uses a different approach to correcting for serial correlation and heteroskedasticity. It is often considered alongside the Augmented Dickey-Fuller test because both aim to detect non-stationary stochastic trends.
4.2.3 KPSS test
The KPSS test reverses the usual null hypothesis by treating stationarity as the default assumption. It is useful as a complement to unit root tests because it asks whether the data are consistent with stationarity rather than with a unit root. Using both types of tests can provide a more balanced assessment.
4.3 Limitations of stationarity tests
Stationarity tests are imperfect and sensitive to sample size, model specification, and structural changes. Short series can make it hard to distinguish stable behavior from drifting behavior. In addition, a series may appear stationary over one period and not another, so test results should be interpreted cautiously.
5 Applications
5.1 Time series forecasting
Forecasting methods often rely on stationarity because stable statistical structure makes future values more predictable from past patterns. Models such as autoregressive and moving average formulations are commonly developed under stationary assumptions. When data are non-stationary, special transformations or alternative models are often needed.
5.2 Signal processing
In signal processing, stationarity helps describe signals whose frequency content and statistical features remain constant over time. This assumption simplifies filtering, detection, and noise reduction. Many standard spectral methods work best when the signal can be treated as approximately stationary.
5.3 Econometrics
Econometrics uses stationarity to study economic and financial series, many of which exhibit trends, persistent shocks, or changing volatility. Determining whether a variable is stationary affects model choice, inference, and long-run interpretation. Stationarity is therefore a central diagnostic in empirical economic analysis.
5.4 Physics and dynamical systems
In physics and dynamical systems, stationarity can describe random fluctuations around equilibrium or a statistically steady regime. It is useful for characterizing noise, turbulence, and other systems where exact trajectories may vary but overall statistical behavior is stable. The concept also supports the study of long-run averages and equilibrium-like states.
6 Methods for achieving stationarity
6.1 Transformation techniques
Transformations can stabilize variance or remove systematic patterns in a series. These methods are often applied before fitting a model so that the transformed data better satisfy stationarity assumptions. The choice of transformation depends on the nature of the non-stationarity.
6.1.1 Logarithmic transformation
A logarithmic transformation is often used when variability increases with the level of the series. By compressing large values more than small ones, it can reduce heteroskedasticity and make fluctuations more uniform. This is common for positive-valued data such as prices or counts.
6.1.2 Detrending
Detrending removes a systematic upward or downward movement from a series. The trend may be estimated with a line, polynomial, or other smooth function, after which the residual component is analyzed. If the remaining variation is stable, the detrended series may be treated as stationary.
6.2 Differencing methods
Differencing is a direct way to reduce non-stationarity caused by persistent changes in level. First differences are the most common, though higher-order differencing may be used when necessary. Care is needed because excessive differencing can remove useful structure and increase noise.
6.3 Seasonal adjustment
Seasonal adjustment removes recurring patterns tied to calendar cycles. By eliminating predictable periodic changes, it can reveal a more stable underlying series. This is especially relevant for monthly or quarterly data with strong seasonal behavior.
7 Related concepts
7.1 Non-stationarity
Non-stationarity refers to the absence of stable statistical properties over time. It may arise from trends, seasonal effects, changing variance, structural breaks, or evolving dependence. Identifying non-stationarity is often the first step before choosing an appropriate model.
7.2 Homogeneity
Homogeneity is a broader idea of uniformity across a domain, and in some contexts it overlaps with stationarity. In stochastic settings, a homogeneous process may have properties that do not vary with time or location. The exact meaning depends on the discipline and model.
7.3 Invariance principles
Invariance principles formalize the idea that certain properties remain unchanged under specified transformations. Stationarity is one form of invariance, namely invariance under time shifts. Such principles are central in probability theory because they reduce complex structures to more manageable symmetries.
7.4 Markov properties
Markov properties concern whether the future depends only on the present state and not on the full past. They are distinct from stationarity, but the two concepts often appear together in stochastic modeling. A Markov process may be stationary under suitable conditions, and stationarity can simplify its long-run analysis.
8 Historical development
8.1 Early statistical theory
Early work on random processes and time-dependent data helped establish the need for concepts describing stable statistical behavior. As data collection expanded, researchers sought methods that could handle dependence over time without treating each observation as independent. Stationarity emerged as a foundational idea in this development.
8.2 Modern time series analysis
Modern time series analysis placed stationarity at the center of model building, especially in linear stochastic models. Theories of autoregression, moving averages, and spectral analysis all rely heavily on it. As statistical computing advanced, stationarity became a routine diagnostic and modeling criterion.
8.3 Influence on computational methods
Computational methods for data analysis have strengthened the practical importance of stationarity. Automated forecasting systems, simulation tools, and numerical estimation procedures often assume or test for stable behavior. At the same time, increased computational power has made it easier to detect and correct departures from stationarity.