Loading field map
Time series econometrics is the branch of econometrics concerned with the statistical analysis of data observed sequentially over time. Its defining feature is that the order of observations carries meaning: the value of a variable today is often related to its own past values, to the past values of other variables, and to shocks that persist or dissipate over time. This dependence distinguishes time series data from cross-sectional data, where observations are assumed to be independent, and it requires a distinct set of concepts, models, and inferential procedures.
The central questions of the field are deceptively simple. How does a variable evolve over time? How can we characterize its dynamic behavior—its tendency to revert to a mean, to drift, to cycle, or to respond to external impulses? How do multiple series move together, and can we disentangle correlation from causation in a temporal setting? And, crucially, how can we use the past to say something reliable about the future, whether for forecasting, for policy evaluation, or for testing economic theories that make predictions about dynamics?
The foundational concept in time series econometrics is stationarity. A stationary series has statistical properties—mean, variance, autocorrelation—that do not change over time. Stationarity is the time series analogue of the assumption of identically distributed observations in cross-sectional work; it is what makes the past informative about the future in a stable way. Many classical results in statistics, such as the law of large numbers and the central limit theorem, apply to stationary series under mild conditions, allowing for valid inference.
The difficulty is that many economic series are not stationary. Gross domestic product, price levels, stock indices, and population figures tend to grow over time; their means are not constant. Early applied work often dealt with this by removing a deterministic trend—fitting a line through the data and analyzing the deviations. But a crucial conceptual shift occurred when researchers recognized that many economic series are better described as difference-stationary rather than trend-stationary. A difference-stationary series has no fixed mean to which it returns; a shock to the level is permanent, and the series wanders without any tendency to revert. Such a series is said to have a unit root, because its autoregressive representation has a root on the unit circle.
The distinction between trend-stationary and difference-stationary processes has profound implications. For a trend-stationary series, shocks are temporary; the series always returns to its deterministic path. For a difference-stationary series, shocks accumulate; the series is a random walk with drift, and its long-run level is unpredictable. This distinction matters for forecasting (a difference-stationary series has forecast intervals that widen without bound), for economic theory (many models predict that certain shocks should have only temporary effects), and for inference (regressions involving unit-root series can produce spurious results, where unrelated series appear strongly correlated simply because both drift upward over time).
The development of unit root tests in the 1980s, most prominently the Dickey–Fuller test and its augmented variants, gave researchers tools to distinguish between these two regimes. The tests are notoriously low-powered—they often fail to reject a unit root even when the series is actually stationary with a high degree of persistence—and the debate over whether particular series (real output, prices, exchange rates) contain unit roots has never been fully settled. What is not disputed is that the question itself is essential: the correct treatment of trends and persistence determines whether a researcher analyzes levels, differences, or deviations from a trend, and this choice can change the conclusions of an empirical study.
The simplest workhorse model for a single time series is the autoregressive moving average (ARMA) model, which describes the current value of a series as a function of its own past values (the autoregressive part) and past random shocks (the moving average part). ARMA models, systematized by George Box and Gwilym Jenkins in the 1970s, provided a flexible yet parsimonious way to model stationary series. The Box–Jenkins methodology—identify, estimate, diagnose—became the standard recipe for building univariate time series models, and it remains influential in forecasting practice.
For non-stationary series, the natural extension is the ARIMA model, where the letter I stands for "integrated": the series is differenced one or more times until it becomes stationary, and an ARMA model is then fitted to the differenced series. ARIMA models are purely statistical; they make no claim about economic structure. Their virtue is flexibility and a track record of reasonable forecasting performance, especially for short horizons. Their vice is that they offer no explanation of why a series behaves as it does, and they cannot easily incorporate information from other related series.
The multivariate extension of ARMA models is the vector autoregression (VAR), introduced into econometrics by Christopher Sims in 1980. A VAR models each variable in a system as a linear function of its own past values and the past values of every other variable in the system. It is a deliberately atheoretical approach: rather than imposing restrictions from economic theory, the researcher includes a set of variables that theory suggests are interrelated and lets the data speak. The VAR became the dominant framework for macroeconometric analysis, replacing the large simultaneous-equation models that had dominated earlier decades.
The key tools for interpreting VARs are the impulse response function and forecast error variance decomposition. An impulse response function traces the effect of a one-time shock to one variable on the subsequent path of all variables in the system. A variance decomposition asks what fraction of the forecast error variance of each variable is attributable to shocks to each of the other variables. These tools give economists a way to describe the dynamic causal structure of the system—but only after a crucial identification problem is solved.
The identification problem in VARs arises because the reduced-form errors—the unpredictable parts of each variable's movement—are generally correlated across equations. A shock to one variable is typically accompanied by contemporaneous movements in others, and the data alone cannot tell us which variable "caused" which. To compute impulse responses with a causal interpretation, the researcher must impose restrictions that disentangle the underlying structural shocks.
The earliest and most common approach is the Cholesky decomposition, which imposes a recursive ordering on the variables: the first variable is assumed to respond to no other variable's contemporaneous shock, the second responds to the first but not to later ones, and so on. This ordering is often justified by appealing to institutional knowledge—for example, that monetary policy reacts to output and inflation within the period, but output and inflation do not react to monetary policy within the period. The results can be sensitive to the chosen ordering, which is a serious limitation.
Subsequent developments offered more sophisticated identification strategies. Structural VARs (SVARs) allow the researcher to impose restrictions derived from economic theory, such as long-run neutrality constraints (e.g., a technology shock is the only shock that affects productivity in the long run) or sign restrictions (e.g., a contractionary monetary policy shock raises interest rates and lowers output). These approaches are more flexible than the recursive ordering, but they are also more demanding: the restrictions must be genuinely justified by theory, and the results can be sensitive to the exact form of the restrictions imposed.
The identification problem is not a technical nuisance; it is the central conceptual challenge of time series econometrics. Time series data are observational, not experimental. The econometrician sees the joint evolution of variables but cannot run controlled experiments to see how one variable would respond to an exogenous change in another. All causal claims in time series econometrics rest on assumptions that are, in principle, untestable from the data alone. The field's progress has largely consisted of making these assumptions more explicit, more flexible, and more closely tied to economic reasoning.
A major theoretical development came with the concept of cointegration, introduced by Clive Granger in the early 1980s. Cointegration addresses a puzzle: if two or more series are individually non-stationary (each contains a unit root), can a linear combination of them be stationary? If so, the series are said to be cointegrated, and the stationary combination represents a long-run equilibrium relationship among them.
The economic intuition is compelling. Consumption and income may each wander without a fixed mean, but the gap between them—the saving rate—may be stable in the long run. Short-term and long-term interest rates may drift, but the spread between them may revert to a constant. If a set of variables is cointegrated, then there exist forces—error correction—that pull the variables back toward their equilibrium relationship whenever they drift apart. This gives empirical content to the idea of economic equilibrium without assuming that the variables themselves are stationary.
The error correction model (ECM) formalizes this intuition. An ECM describes the short-run dynamics of the variables as a function of their own past changes plus a correction term that measures the deviation from the long-run equilibrium. The ECM is not merely a statistical convenience; it has a natural economic interpretation. Agents adjust their behavior to correct past disequilibria, and the speed of adjustment is an estimable parameter.
The econometric theory of cointegration was developed most fully by Søren Johansen, who provided a maximum-likelihood procedure for estimating the number of cointegrating relationships in a system and testing hypotheses about their form. The Johansen approach became the standard tool for applied work on long-run relationships, from purchasing power parity in exchange rates to the term structure of interest rates to money demand functions. Cointegration also resolved a practical problem: regressions involving non-stationary variables can be spurious, but if the variables are cointegrated, the regression has a meaningful interpretation as an estimate of the long-run relationship.
A separate strand of time series econometrics focuses not on the level of a series but on its variance. Financial returns, exchange rates, and other asset prices exhibit volatility clustering: periods of high volatility tend to be followed by high volatility, and calm periods by calm. The variance of these series is not constant; it evolves over time in a way that is itself predictable.
The autoregressive conditional heteroskedasticity (ARCH) model, introduced by Robert Engle in 1982, and its generalization, GARCH, developed by Tim Bollerslev, model the conditional variance as a function of past squared shocks and past variances. These models capture the persistence of volatility and allow researchers to forecast the entire distribution of future returns, not just the mean. They also provide a correction for inference: standard errors that ignore time-varying volatility can be badly misleading.
GARCH models spawned a large family of extensions—EGARCH for asymmetric responses to positive and negative shocks, GJR-GARCH for leverage effects, multivariate GARCH for systems of asset returns, and stochastic volatility models that treat volatility as a latent process. These models are central to empirical finance, where they are used for risk management (value-at-risk calculations), option pricing, and portfolio allocation. They also illustrate a general theme in time series econometrics: the same data can be modeled at multiple levels, and the second moment can be as interesting as the first.
Contemporary time series econometrics is characterized by several overlapping developments rather than a single dominant paradigm.
Bayesian methods have become increasingly prominent. Bayesian VARs, which impose prior distributions that shrink the model toward simpler specifications, can handle systems with many variables and produce forecasts that often outperform their frequentist counterparts. Bayesian inference also provides a natural framework for handling parameter uncertainty, model uncertainty, and the sequential updating of forecasts as new data arrive.
Structural break and regime-switching models address the possibility that the parameters of a time series model are not constant over time. The Markov-switching model, associated with James Hamilton, allows the process to shift between discrete regimes (e.g., expansion and recession) with transition probabilities estimated from the data. Tests for structural breaks, developed by researchers including Donald Andrews and Jushan Bai, ask whether the coefficients of a model changed at some unknown date. These tools matter because the stability of parameters is a maintained assumption of most time series models, and that assumption is often violated in practice.
Nonparametric and machine learning methods have entered the field, offering alternatives to the parametric models that dominate traditional practice. These methods are more flexible—they do not assume a particular functional form—but they require large samples and can be difficult to interpret. Their use in forecasting has grown, particularly for high-dimensional problems where the number of potential predictors is large relative to the sample size.
Panel time series methods combine cross-sectional and time series variation. When many units (countries, firms, individuals) are observed over time, the researcher can exploit both dimensions. Tests for unit roots and cointegration in panels have higher power than their time-series-only counterparts, and they allow for heterogeneity across units. This literature has been especially influential in empirical macroeconomics and international finance.
Throughout these developments, a persistent tension remains between theory-driven and data-driven approaches. Structural VARs and cointegration analysis impose economic structure to achieve identification. ARIMA and machine learning models make minimal assumptions and prioritize predictive accuracy. The field has not resolved this tension, and it is unlikely to do so: the choice depends on the research question. For forecasting, a model that fits well may be preferred regardless of its economic interpretability. For policy analysis, a model that can answer "what would happen if" questions requires structural assumptions, however imperfect.
The durable contribution of time series econometrics is not any single model or test but a set of habits of mind: a respect for the dependence structure of temporal data, a recognition that non-stationarity can invalidate standard inference, an insistence on making identification assumptions explicit, and a willingness to let the data challenge theoretical priors. These habits have diffused far beyond econometrics proper, into empirical finance, macroeconomics, political science, epidemiology, and climate science. The field's methods are now standard tools wherever data arrive over time and the past is expected to inform the future.