Loading field map
A central challenge in empirical economics is that most macroeconomic and financial data arrive as sequences of observations over time—GDP, inflation, stock prices—and these sequences are rarely independent. A shock today echoes into tomorrow, and trends can shift unpredictably. Early econometricians faced a practical pressure: how to model the dynamic structure of such data without imposing rigid theoretical assumptions that might be wrong. The answer came not from a single breakthrough but from a series of frameworks that each addressed a different facet of the problem—filtering, forecasting, multivariate interactions, nonstationarity, and volatility. These frameworks remain active today, often used in combination, but their origins reveal a field in constant dialogue with its own limitations.
The first major framework, State-Space and Kalman Filter Models (1960–Present), originated in control engineering. Rudolf Kalman’s filter provided a recursive algorithm to estimate the unobserved state of a system from noisy measurements, updating estimates in real time as new data arrived. Economists quickly recognized its potential: many economic variables (e.g., potential output, the natural rate of unemployment) are unobservable but can be inferred from observable series. The state-space representation also allowed for a clean separation between the underlying dynamics (the state equation) and the measurement process (the observation equation). This framework gave researchers a flexible way to incorporate structural assumptions—such as a Phillips curve relationship—while letting the data update those beliefs sequentially. It remains a workhorse for nowcasting, dynamic factor models, and DSGE estimation.
A decade later, Box-Jenkins ARIMA Methodology (1970–Present) took a deliberately different path. George Box and Gwilym Jenkins argued that before imposing any economic structure, the analyst should let the data “speak” through a systematic iterative process of identification, estimation, and diagnostic checking. Their Autoregressive Integrated Moving Average (ARIMA) model captured autocorrelation and nonstationarity by differencing the series and then fitting AR and MA terms. Unlike state-space models, which often required the researcher to specify a structural form, ARIMA was purely data-driven. It narrowed the focus to univariate forecasting, trading structural interpretability for simplicity and robustness. The two frameworks coexisted: state-space models were preferred when a theoretical structure was available, while ARIMA became the default for baseline forecasting and seasonal adjustment.
By the late 1970s, macroeconomists grew frustrated with the “incredible” restrictions imposed by large structural models. Christopher Sims’s Vector Autoregression (VAR) (1980–Present) directly challenged that tradition. Instead of specifying which variables were exogenous or which lags entered which equation, VAR treated every variable as endogenous and modeled each as a linear function of its own past and the past of all other variables. This was a deliberate narrowing: VAR abandoned structural interpretation in favor of a reduced-form description of the data’s joint dynamics. It emerged as a reaction to the limitations of both ARIMA (which was univariate) and state-space models (which often required strong theory). VAR allowed researchers to trace out impulse responses—how a shock to one variable propagates through the system—without prior theoretical commitments. However, the atheoretical nature of VAR also drew criticism: it could produce implausible results if the underlying data were nonstationary or if important variables were omitted.
That concern about nonstationarity led directly to Unit Root and Cointegration Econometrics (1981–Present). David Dickey and Wayne Fuller had already shown that standard t-tests break down when a series contains a unit root—a random walk that never reverts to a mean. Later, Peter Phillips and others developed robust inference for unit roots under general conditions. But the real breakthrough came when Clive Granger and Robert Engle introduced cointegration: even if individual series are nonstationary, a linear combination of them may be stationary, implying a long-run equilibrium relationship. This framework transformed how economists thought about spurious regression. It also complemented VAR by providing a way to model both short-run dynamics and long-run relationships through error-correction models. Where ARIMA and VAR had simply differenced away nonstationarity, cointegration preserved the economic content of levels. The unit root revolution thus absorbed the earlier frameworks’ concern with stationarity while adding a richer toolkit for handling trends.
A different kind of pattern—volatility clustering—had long been observed in financial returns: large changes tend to follow large changes, and small changes follow small changes. Standard time series models assumed constant variance, but that assumption was clearly violated for asset prices. Conditional Heteroskedasticity Models (1982–Present) addressed this gap. Robert Engle’s Autoregressive Conditional Heteroskedasticity (ARCH) model allowed the variance of the current error to depend on the squared errors of previous periods. Tim Bollerslev generalized this to GARCH, which also includes past variances. These models did not replace the earlier frameworks; they added a new dimension. A researcher could now combine an ARIMA or VAR specification for the conditional mean with a GARCH specification for the conditional variance. The framework became essential for risk management, option pricing, and any application where uncertainty itself evolves over time. It remains the standard tool for modeling financial volatility, with extensions (EGARCH, GJR-GARCH) that capture asymmetries and leverage effects.
All five frameworks remain active today, but they are no longer isolated. A modern time series analysis often begins with unit root tests to determine the order of integration, then uses a VAR or cointegrated VAR to model multivariate dynamics, and may add a GARCH component if volatility matters. State-space models are used for unobserved components and nowcasting, while ARIMA remains a benchmark for univariate forecasting. The leading frameworks agree on the fundamental importance of handling autocorrelation and nonstationarity properly. They disagree, however, on the role of economic theory. State-space models and cointegration often embed theoretical restrictions (e.g., long-run elasticities), while VAR and ARIMA are more agnostic. Conditional heteroskedasticity models are largely atheoretical, focusing on statistical patterns in variance. This tension between theory-driven and data-driven modeling is not resolved; it is a productive disagreement that shapes how each framework is applied. The field today is pluralistic: the best analysis uses the right tool for the right question, often combining several frameworks in a single empirical project.