Loading field map
Time series analysis is the branch of statistics concerned with data collected sequentially over time, where the order of observations carries meaning. Its central problem is that standard statistical tools, which assume observations are independent, fail when applied to data in which each value is correlated with those that came before. A time series is any sequence of measurements indexed by time—daily stock prices, monthly unemployment figures, hourly electricity demand, annual global temperatures, or the electrical activity of a single neuron recorded in milliseconds. The field develops models and methods that exploit, rather than ignore, the temporal dependence in such data.
The stakes of the field are practical and high. Forecasts of demand, disease spread, financial risk, and climate variables all depend on time series models. But the field also addresses a deeper scientific question: how can one infer the mechanisms that generate a sequence of observations when the only evidence is the sequence itself? This dual character—part engineering discipline for prediction, part scientific tool for understanding dynamic systems—shapes every major approach within the field.
Two features distinguish time series data from other data types. The first is serial dependence (or autocorrelation): the value at time t tends to be similar to, or predictable from, values at earlier times. The second is non-stationarity: the statistical properties of the series—its mean, variance, or correlation structure—may change over time. A series with a trend, a seasonal pattern, or a sudden regime shift is non-stationary.
These two features create the field's central difficulties. Ignoring serial dependence leads to misleading inferences: standard confidence intervals and hypothesis tests assume independent observations, and with autocorrelated data they can be wildly overconfident. Ignoring non-stationarity leads to models that describe the past but fail on the future, because the process generating the data has changed. Much of the history of time series analysis is a sequence of attempts to handle these two problems, and the major approaches differ chiefly in how they model dependence and how they treat change over time.
The foundational framework for time series analysis was laid in the first half of the twentieth century, drawing on work in probability theory and statistical mechanics. The key idea was to treat a time series not as a single sequence of numbers but as one realization of a stochastic process—a family of random variables indexed by time. The goal was to characterize the process that generated the observed data.
The classical theory focused on stationary processes, in which the joint distribution of any set of observations is unchanged by shifting time. For practical work, attention centered on weak stationarity, which requires only that the mean be constant and that the covariance between two observations depend only on their time separation, not on their absolute positions. For a weakly stationary process, the entire dependence structure is captured by the autocovariance function (or its normalized version, the autocorrelation function), which describes how strongly an observation at time t is correlated with an observation at time t + k.
The central mathematical result of this classical period was the spectral representation: a stationary process can be decomposed into a sum (or integral) of uncorrelated sinusoidal components at different frequencies. This insight connected time domain analysis—working directly with the sequence of values—to frequency domain analysis, in which the series is described by its spectrum, a function showing how much of the series' variance is contributed by oscillations at each frequency. The two perspectives are mathematically equivalent for stationary processes, but they emphasize different questions. Time domain methods ask how the present depends on the past; frequency domain methods ask which periodic cycles dominate the series.
The classical framework reached its mature form in the mid-twentieth century with the development of ARMA models (autoregressive moving average). An autoregressive model of order p, denoted AR(p), expresses each observation as a linear combination of the previous p observations plus a random shock. A moving average model of order q, denoted MA(q), expresses each observation as a linear combination of the current and previous q random shocks. The combination, ARMA(p, q), provides a flexible but parsimonious way to model the autocorrelation structure of a stationary series. The Box–Jenkins methodology, introduced in the 1970s, turned this into a practical workflow: identify an appropriate model from the data's autocorrelation patterns, estimate its parameters, and check whether the residuals resemble white noise.
The classical approach remains the backbone of the field. Its assumptions, however, are restrictive. Real-world series are rarely stationary: they have trends, seasonal cycles that change over time, and structural breaks. The classical response was to transform the data—taking differences to remove trends, for example—before applying stationary models. This worked reasonably well for forecasting but was unsatisfying as a description of the underlying process.
The most influential practical development in the time domain was the ARIMA model (autoregressive integrated moving average), which extended ARMA to handle non-stationarity through differencing. If a series must be differenced d times to become stationary, it is said to follow an ARIMA(p, d, q) process. The model is simple, computationally tractable, and remarkably effective for short-term forecasting of many economic and industrial series. The Box–Jenkins methodology made ARIMA the default tool for time series forecasting for several decades.
A crucial extension addressed seasonality. Many series—retail sales, airline passenger counts, energy demand—exhibit regular periodic patterns. The seasonal ARIMA model (SARIMA) adds seasonal autoregressive and moving average terms at the seasonal lag, allowing the model to capture both short-term dependence and longer seasonal cycles. This remains one of the most widely used forecasting tools in practice.
The time domain tradition also developed important diagnostic and inferential tools. The Augmented Dickey–Fuller test and related unit root tests ask whether a series is stationary or whether it contains a stochastic trend—a trend that evolves randomly over time rather than following a fixed deterministic path. This distinction matters enormously in economics: a shock to a stationary series decays over time, while a shock to a unit root series has a permanent effect. The debate over whether economic variables like GDP or unemployment contain unit roots was one of the most active research areas in econometrics in the 1980s and 1990s.
The limits of the ARIMA framework are significant. It is linear, so it cannot capture threshold effects, regime switches, or other nonlinear dynamics. It assumes the model structure is constant over time, so it handles structural breaks poorly. And it is univariate: it models a single series in isolation, ignoring the possibility that other variables might help predict it.
Running parallel to the time domain approach, and often in productive tension with it, is the frequency domain tradition. Spectral analysis asks not "how does the present depend on the past?" but "which oscillations at which frequencies make up the series?" The spectrum of a stationary process decomposes its variance by frequency, and estimating the spectrum from data became a major practical problem.
The early methods were crude: the periodogram, computed by taking the Fourier transform of the data, is an inconsistent estimator of the spectrum—it does not improve as the sample size grows. The solution, developed in the mid-twentieth century, was smoothing: averaging the periodogram over neighboring frequencies, or weighting the autocovariance function before transforming it. These techniques, associated with the work of Norbert Wiener, Maurice Kendall, and later Emanuel Parzen and others, produced reliable spectral estimates.
Spectral analysis proved especially valuable in geophysics, engineering, and the physical sciences, where the data often have natural periodic components. It is the standard tool for detecting hidden periodicities—cycles that are not obvious from the raw data—and for characterizing the response of a system to inputs at different frequencies. The cross-spectrum extends the idea to pairs of series, measuring how much of the variance in one series at a given frequency is related to the other series at the same frequency.
The frequency domain approach has a distinctive limitation: it assumes stationarity, and its estimates require long stretches of data to achieve reasonable precision. For short or non-stationary series, spectral estimates are unreliable. This limitation motivated later developments that combine time and frequency information, such as the wavelet transform, which decomposes a series into components localized in both time and frequency, allowing the analysis of series whose periodic behavior changes over time.
A third major tradition, with roots in control engineering and systems theory, treats a time series as a noisy observation of an underlying state that evolves over time according to its own dynamics. The state might be the "true" level of a variable, a latent trend, or a set of unobserved components such as trend, seasonal, and irregular terms. The observation equation links the state to the data, and the state equation describes how the state evolves.
The key computational tool is the Kalman filter, an algorithm that recursively updates estimates of the state as new observations arrive. It provides optimal linear estimates under Gaussian assumptions and is computationally efficient, making it suitable for real-time applications. The Kalman filter was developed in the late 1950s and early 1960s for spacecraft navigation, but it quickly became a general tool for time series analysis.
State space methods offer several advantages over the ARIMA framework. They handle missing data naturally, because the filter can skip observations. They accommodate time-varying parameters, because the state can include coefficients that evolve over time. And they provide a unified framework for a wide range of models: ARIMA models can be written in state space form, as can structural time series models, in which the series is explicitly decomposed into trend, seasonal, and irregular components. The structural time series approach, associated with Andrew Harvey and others, is a direct rival to ARIMA for economic forecasting, with the advantage that its components have direct substantive interpretations.
The main limitation of state space methods is that the model must be specified in advance, and the results depend on that specification. The Kalman filter also assumes linear dynamics and Gaussian errors, although extensions exist for nonlinear and non-Gaussian cases.
Since the 1980s, the field has fragmented into a diverse set of approaches, none of which has achieved the dominance that ARIMA once held. Several developments define the modern landscape.
Nonlinear models relax the linearity assumption of the classical framework. Threshold autoregressive models allow the dynamics to change when the series crosses a threshold value. Markov switching models, introduced by James Hamilton, allow the series to move between distinct regimes, each with its own dynamics, with the transitions governed by an unobserved Markov chain. GARCH models (generalized autoregressive conditional heteroskedasticity), developed by Robert Engle, address a different kind of nonlinearity: the variance of the series changes over time, with periods of high volatility clustering together. GARCH and its many variants became the standard tool in financial econometrics for modeling volatility, which is not directly observed but must be estimated from the squared or absolute returns of asset prices.
Cointegration, developed by Engle and Clive Granger, addresses a problem that arises with non-stationary series. Two or more non-stationary series may move together in the long run, even though each individually wanders unpredictably. If a linear combination of them is stationary, they are said to be cointegrated. This concept transformed the analysis of economic time series, because it provided a rigorous way to model long-run equilibrium relationships among variables—such as consumption and income, or spot and futures prices—while allowing for short-run deviations. The error correction model combines the long-run relationship with short-run dynamics, describing how the system adjusts toward equilibrium.
Vector autoregressions (VARs), introduced by Christopher Sims, extended the univariate AR model to multiple series. A VAR models each variable as a linear function of its own past and the past of all other variables in the system. VARs became the standard tool for macroeconomic empirical work, used for forecasting, for analyzing the dynamic response of the economy to shocks, and for testing economic theories. The related concept of Granger causality—a variable X is said to Granger-cause Y if past values of X improve predictions of Y beyond what past values of Y alone achieve—provides a statistical notion of predictive influence that is widely used, though it does not establish true causation.
Machine learning methods entered the field in the 1990s and have grown steadily in influence. Neural networks, random forests, gradient boosting, and more recently deep learning architectures such as recurrent neural networks and transformers have been applied to time series forecasting, often with strong empirical results on large datasets. These methods are nonparametric: they do not assume a specific functional form for the dependence structure, but instead learn it from data. Their strengths are flexibility and scalability; their weaknesses are interpretability, the need for large amounts of data, and the difficulty of quantifying uncertainty in predictions. The relationship between classical statistical time series models and machine learning methods remains an active area of debate, with some researchers arguing for hybrid approaches that combine the interpretability of statistical models with the flexibility of machine learning.
Long memory and fractional integration address series in which the autocorrelation decays more slowly than the exponential decay of ARMA models. Such series exhibit dependence between observations that are far apart in time, and they are modeled using fractional differencing, in which the differencing parameter is allowed to be non-integer. Long memory has been found in hydrology, finance, and geophysics, though its presence in economic series remains contested.
Bayesian methods have become increasingly prominent. Bayesian time series analysis treats model parameters as random variables and updates prior distributions as data arrive. This approach naturally handles parameter uncertainty, allows the incorporation of prior information, and provides a coherent framework for model comparison and averaging. Dynamic linear models, a Bayesian version of state space models, have been developed extensively by Mike West and Jeff Harrison and are widely used in forecasting and monitoring.
Despite the diversity of approaches, the field is unified by a common set of concerns. Every method must address the same questions: How should dependence be modeled? How should non-stationarity be handled? How should uncertainty be quantified? How should models be compared and validated? The different traditions answer these questions differently, but they are engaged in a common enterprise.
The classical stationary process theory provides the mathematical foundation on which nearly everything else builds. The time domain and frequency domain approaches are complementary perspectives on the same underlying processes, and modern methods often combine them. State space methods provide a general framework that encompasses many specific models. The modern nonlinear, multivariate, and machine learning approaches extend the classical framework to handle the complexity of real data.
The field's history is not a simple succession of one approach replacing another. ARIMA models are still widely used, spectral analysis remains a standard tool in the physical sciences, and state space methods have become a general framework for many applications. The modern landscape is best described as a set of coexisting approaches, each with its own strengths and limitations, chosen according to the nature of the data and the questions being asked. The central challenge of the field—making reliable inferences and predictions from data that are neither independent nor stationary—remains as central today as it was a century ago.