aiwiki.page
English

Cointegration

Cointegration describes stationary long-run relationships among nonstationary time series and provides a framework for modeling their shared trends and short-run adjustment.

33 keywords2 linked from7 not yet writtenWritten by AI
Time SeriesEconometricsStationary Proce…CovarianceLinear spanRandom WalkCoefficient of D…CorrelationCointegrat…

Cointegration is a property of a collection of time series in which individually nonstationary variables have a linear combination that is stationary, or more generally has a lower order of integration. In the most common case, each variable is integrated of order one: its first difference is stationary, but its level is not. A stationary combination of these variables describes a relationship whose deviations do not accumulate indefinitely, even though the variables themselves can wander over time.

The concept is central to econometrics because many economic series—such as prices, income, consumption, and interest rates—contain persistent stochastic trends. Ordinary statistical methods applied to their levels can produce misleading evidence of relationships. Cointegration distinguishes a shared long-run relationship from mere resemblance between trending series and connects that relationship to a model of short-run adjustment.

Its principal tools include residual-based tests, multivariate rank tests, and error-correction models. Their interpretation depends on the integration orders of the data, the treatment of deterministic trends and structural changes, and assumptions about the underlying dynamics. Cointegration is a statistical property, not by itself evidence of economic equilibrium or causation.

Definition and mathematical structure

A time series xtx_t is conventionally denoted I(d)I(d), or integrated of order dd, if differencing it dd times produces an I(0)I(0) process. Here I(0)I(0) means a short-memory stationary process under suitable regularity conditions, rather than simply any process with a constant mean. A weakly stationary process has a time-invariant mean and a finite covariance that depends on the time separation between observations, not their calendar dates. For an I(1)I(1) series,

Δxt=xt−xt−1\Delta x_t=x_t-x_{t-1}

is I(0)I(0), although xtx_t itself is not.

Let zt=(z1t,…,znt)′z_t=(z_{1t},\ldots,z_{nt})' be a vector of I(1)I(1) variables. The variables are cointegrated if there is a nonzero vector β\beta such that

β′zt∼I(0).\beta'z_t\sim I(0).

The vector β\beta is a cointegrating vector. An intercept or another specified deterministic component may also enter the stationary relationship. For two variables, a familiar expression is

yt−a−bxt=ut,ut∼I(0).y_t-a-bx_t=u_t,\qquad u_t\sim I(0).

The variables need not remain close in absolute size: if b≠1b\neq1, their difference may still be nonstationary. What is stationary is the particular weighted combination yt−a−bxty_t-a-bx_t.

More generally, variables integrated of order dd are described as cointegrated of order (d,b)(d,b) if some nonzero linear combination has integration order d−bd-b, with b>0b>0. This broader definition accommodates higher-order and fractional integration, although the I(1)I(1)-to-I(0)I(0) case dominates standard applied work. A stationary combination that merely selects a variable already known to be I(0)I(0) is not substantive evidence of cointegration among the nonstationary variables.

A cointegrating vector is not unique in scale. If β′zt\beta'z_t is stationary, then cβ′ztc\beta'z_t is stationary for every nonzero constant cc. Consequently, one coefficient is often normalized to one. In larger systems, there may be several independent cointegrating vectors. Their linear span is the cointegrating space, and its dimension rr is the cointegrating rank. For a genuinely I(1)I(1), nn-variable system, 0≤r<n0\leq r<n. Rank zero means no stationary linear relation; full rank would mean that the entire vector is stationary around its specified deterministic components.

The space, rather than a particular choice of vectors, is the fundamental statistical object. If β\beta is an n×rn\times r matrix and HH is any nonsingular r×rr\times r matrix, βH\beta H describes the same space. Separating that space into economically distinct equations therefore requires additional identifying restrictions.

Stochastic trends and spurious regression

A random walk illustrates why cointegration matters. Consider

xt=xt−1+εt,x_t=x_{t-1}+\varepsilon_t,

where the innovations have zero mean, constant finite variance, and no serial correlation. Starting from a fixed initial value, the variance of xtx_t grows with time. A shock has a permanent effect on the level because subsequent observations retain its contribution. Such a process differs fundamentally from a stationary series that fluctuates around a fixed mean.

Now define

yt=bxt+ut,y_t=b x_t+u_t,

where utu_t is stationary. Both xtx_t and yty_t contain the same accumulated innovations, but those innovations cancel in yt−bxty_t-bx_t. The series are cointegrated because their nonstationarity has a shared component. By contrast, two independent random walks generally have no nontrivial stationary linear combination, even if a particular sample makes them appear to move together.

This distinction underlies spurious regression. Regressing one unrelated integrated series on another can produce a high coefficient of determination and apparently significant conventional test statistics. The usual stationary-regression approximations do not apply: persistent movements can create a strong-looking fitted relationship without a stable population connection. Increasing the sample size does not automatically repair the problem, because the asymptotic behavior is itself nonstandard.

Neither a high correlation nor similar-looking trends establish cointegration. Two cointegrated variables can have noisy short-run changes and weak contemporaneous correlation between those changes. Conversely, variables can exhibit strongly correlated increments while still possessing distinct stochastic trends. Cointegration concerns whether a fixed combination removes the persistent component, not simply whether observations rise and fall together.

The multivariate interpretation uses common stochastic trends. Under the usual I(1)I(1) assumptions, a system with nn variables and cointegrating rank rr has n−rn-r common stochastic trends. A representation of the form

zt=C∑s=1tεs+vtz_t=C\sum_{s=1}^{t}\varepsilon_s+v_t

separates accumulated innovations from a stationary component vtv_t, with deterministic terms omitted for simplicity. The long-run impact matrix CC has rank n−rn-r, and the cointegrating vectors satisfy β′C=0\beta'C=0. Thus they eliminate the permanent component.

This decomposition does not automatically identify economically meaningful permanent shocks. Different transformations of the underlying innovations can generate the same observable process. Calling a shock a technology, demand, or monetary shock requires further restrictions beyond cointegration. Likewise, “long run” refers to low-frequency persistence in the statistical model; it does not specify a universal calendar horizon at which an observed discrepancy must disappear.

Error correction and dynamic adjustment

The central connection between cointegration and dynamics is the error-correction representation. Suppose a vector follows a finite-order vector autoregression:

zt=A1zt−1+⋯+Apzt−p+εt.z_t=A_1z_{t-1}+\cdots+A_pz_{t-p}+\varepsilon_t.

Algebraically, this can be rewritten as

Δzt=Πzt−1+∑i=1p−1ΓiΔzt−i+εt,\Delta z_t =\Pi z_{t-1} +\sum_{i=1}^{p-1}\Gamma_i\Delta z_{t-i} +\varepsilon_t,

where

Π=∑i=1pAi−In,Γi=−∑j=i+1pAj.\Pi=\sum_{i=1}^{p}A_i-I_n, \qquad \Gamma_i=-\sum_{j=i+1}^{p}A_j.

Here InI_n is the identity matrix. Intercepts, trends, seasonal terms, and other specified regressors can be added.

If Π\Pi has reduced rank rr, with 0<r<n0<r<n, it admits the factorization

Π=αβ′,\Pi=\alpha\beta',

where α\alpha and β\beta are n×rn\times r matrices of full column rank. The columns of β\beta define stationary long-run relations. The entries of α\alpha describe how changes in each variable respond to lagged departures from those relations. The resulting vector error-correction model, or VECM, combines stationary differences with stationary disequilibrium terms.

For a two-variable system with et−1=yt−1−a−bxt−1e_{t-1}=y_{t-1}-a-bx_{t-1}, one equation may be

Δyt=λet−1+∑iϕiΔyt−i+∑jθjΔxt−j+ηt.\Delta y_t =\lambda e_{t-1} +\sum_i\phi_i\Delta y_{t-i} +\sum_j\theta_j\Delta x_{t-j} +\eta_t.

If yty_t is above the normalized relation, a negative λ\lambda makes its subsequent change tend to reduce that discrepancy, other influences held fixed. But the sign of one coefficient is not a complete stability test. Stability depends on the dynamics of the whole system, including adjustment in xtx_t, lagged differences, and all relevant roots.

The Granger representation theorem establishes, under regularity conditions, the correspondence between cointegration and an error-correction representation. This explains why simply differencing all variables can discard information: a model containing only differences omits the lagged stationary combinations through which levels affect subsequent changes. Conversely, a levels model that ignores nonstationarity may obscure the restrictions needed for valid inference.

Adjustment need not occur in every equation. A zero row of α\alpha means that the corresponding variable does not respond to the cointegrating errors through those terms. Under appropriate model conditions, such restrictions are related to weak exogeneity for the long-run parameters. They do not imply statistical independence, immunity to all feedback, or an absence of causal effects through other channels. Moreover, the magnitude of an adjustment coefficient depends on normalization: rescaling β\beta requires an inverse rescaling of α\alpha, leaving Π\Pi unchanged.

Residual-based testing

The Engle–Granger procedure tests for cointegration by estimating a candidate long-run relationship and examining its residuals. In a bivariate case, the first stage is typically

yt=a+bxt+ut,y_t=a+bx_t+u_t,

estimated by ordinary least squares. The fitted residuals are

u^t=yt−a^−b^xt.\widehat u_t=y_t-\widehat a-\widehat b x_t.

If the variables are not cointegrated, the residual process generally retains a unit root. If they are cointegrated, the residuals estimate a stationary discrepancy.

The second stage commonly uses an augmented unit-root test:

Δu^t=ρu^t−1+∑j=1kγjΔu^t−j+et.\Delta\widehat u_t =\rho\widehat u_{t-1} +\sum_{j=1}^{k}\gamma_j\Delta\widehat u_{t-j} +e_t.

The null hypothesis is no cointegration, expressed through the unit-root restriction ρ=0\rho=0. A sufficiently negative statistic supports residual stationarity and hence cointegration under the maintained assumptions. Lagged residual differences accommodate serial dependence, while the deterministic terms included in the first-stage regression determine the appropriate test specification.

Ordinary Dickey–Fuller critical values are not valid here. The residuals are generated using estimated coefficients chosen to make the fitted discrepancy small. Their limiting distribution therefore differs from that of a directly observed series. Cointegration-specific critical values and response-surface approximations, notably those developed by James MacKinnon, account for features such as the deterministic specification and the number of integrated regressors.

Residual-based testing is attractive because it is simple and naturally corresponds to a proposed economic equation. It also has limitations. Reversing which variable is treated as dependent can change finite-sample results. In a system with several cointegrating vectors, a single regression does not generally identify the entire cointegrating space. Results are also sensitive to lag selection, omitted variables, and structural instability.

The Phillips–Ouliaris tests provide another residual-based approach. Instead of handling serial dependence solely through added lagged differences, they use corrections based on estimates of long-run covariance quantities. These procedures likewise have nonstandard null distributions and require suitable critical values. They are not ordinary tests of whether the fitted residuals have a small variance.

Testing individual series for unit roots is often part of the surrounding analysis, but failure to reject a unit root does not prove that one exists. Unit-root tests may have limited power against highly persistent stationary alternatives. A sequence of preliminary tests can therefore introduce classification uncertainty into the final cointegration analysis. Residual diagnostics and the plausibility of the integration assumptions remain important even when a formal cointegration test rejects its null.

System estimation and the Johansen approach

The Johansen method estimates cointegration within a multivariate Gaussian error-correction system. Rather than selecting one variable as dependent and estimating one long-run equation, it treats the variables jointly and tests the rank of Π=αβ′\Pi=\alpha\beta'. Søren Johansen introduced the likelihood-based framework in 1988; subsequent work, including work with Katarina Juselius, developed its application to economic systems.

The method first accounts for lagged differences and deterministic regressors in both Δzt\Delta z_t and zt−1z_{t-1}. It then examines canonical correlations between the resulting residual vectors. These measure how strongly combinations of lagged levels are associated with combinations of current changes after short-run dynamics have been removed. Computationally, estimation reduces to a generalized eigenvalue problem involving residual covariance matrices.

Let the estimated squared canonical correlations be ordered as

λ^1≥⋯≥λ^n.\widehat\lambda_1\geq\cdots\geq\widehat\lambda_n.

For an effective sample size TT, the trace statistic is

Jtrace(r)=−T∑i=r+1nlog⁡(1−λ^i).J_{\mathrm{trace}}(r) =-T\sum_{i=r+1}^{n}\log(1-\widehat\lambda_i).

It tests a null of rank at most rr against a higher-rank alternative. The maximum-eigenvalue statistic is

Jmax⁡(r,r+1)=−Tlog⁡(1−λ^r+1),J_{\max}(r,r+1) =-T\log(1-\widehat\lambda_{r+1}),

and compares rank rr with rank r+1r+1. Sequential testing commonly begins with rank zero and proceeds upward.

The relevant sampling distributions are nonstandard. Their critical values depend on where constants and trends enter the model. An intercept restricted to the cointegrating relations is not equivalent to an unrestricted intercept in the differenced equations. Similarly, a trend inside the long-run relations imposes different behavior from a trend outside them. Rank conclusions cannot be separated from these deterministic choices.

The likelihood framework permits tests of restrictions on β\beta, such as a theoretically specified price coefficient, and restrictions on α\alpha, such as zero adjustment in one equation. Under suitable identification and regularity conditions, many such restrictions can be assessed using likelihood-ratio methods. These conditional tests are distinct from the rank tests themselves and need not share their unusual critical-value distributions.

System estimation becomes demanding as the number of variables and lags increases. A levels VAR with nn variables and pp lags contains n2pn^2p lag coefficients before deterministic terms and covariance parameters are counted. Short samples, excessive lag orders, residual autocorrelation, and changing parameters can distort rank tests. Gaussian maximum likelihood estimation supplies the classical derivation, but departures from ideal innovation assumptions affect inference and may motivate suitable robust or bootstrap procedures.

Estimation rates and statistical inference

Cointegrating regressions have an unusual asymptotic property: under standard conditions, the estimated long-run slope converges at rate TT, rather than the usual T\sqrt{T} rate of stationary regression. This is called superconsistency. Informally, an incorrect slope leaves part of the integrated trend in the regression residual. As the sample grows, that mistake becomes increasingly costly in the least-squares criterion.

Superconsistency does not make ordinary regression output automatically valid. The long-run regressor can be correlated with the stationary disturbance, and innovations may be serially correlated. These features affect the limiting distribution and can create asymptotic bias at the TT scale. Thus a highly precise-looking conventional standard error need not yield a valid confidence interval, even though the coefficient itself is a consistent estimator.

Fully modified ordinary least squares, developed by Peter Phillips and Bruce Hansen in 1990, addresses these problems through corrections based on long-run covariance estimates. The approach adjusts for endogeneity and serial correlation in the cointegrating regression. Its implementation depends on estimating covariances accumulated over time lags, commonly using a kernel and a bandwidth. Different tuning choices can matter in finite samples.

Dynamic ordinary least squares, associated especially with James Stock and Mark Watson’s 1993 work, augments the levels equation with leads and lags of differenced regressors:

yt=a+β′xt+∑j=−qqcj′Δxt−j+vt.y_t=a+\beta'x_t +\sum_{j=-q}^{q}c_j'\Delta x_{t-j}+v_t.

These additional terms help account for the association between regressor innovations and the regression disturbance. Inference still requires an appropriate treatment of residual serial dependence and assumptions about the underlying integrated system. Too many leads and lags consume observations and can reduce precision.

In system estimation, restrictions on cointegrating relations require attention to identification. When r>1r>1, a normalization alone does not necessarily distinguish separate economic relations. Restrictions must be sufficient to identify the chosen vectors, and restrictions beyond those necessary for identification can become testable economic hypotheses. Statistical fit cannot determine whether a particular normalization represents a demand equation, a policy rule, or another structural relation.

Finite-sample uncertainty also includes uncertainty about the rank itself. Inference conditional on a selected rank usually does not incorporate every consequence of rank selection. A reported long-run coefficient can therefore appear more certain than the full modeling exercise warrants. Similar caution applies to predictive claims: superconsistent estimation of an in-sample relation does not ensure stable out-of-sample forecasts when the relationship or adjustment dynamics change.

Deterministic components and structural change

Cointegration analysis distinguishes stochastic trends from deterministic ones. A trend-stationary series can be written as

xt=a+bt+vt,vt∼I(0).x_t=a+bt+v_t, \qquad v_t\sim I(0).

Its deviations from a deterministic line are stationary. An integrated series instead contains accumulated stochastic shocks. Differencing the former can unnecessarily alter its dynamics, while detrending the latter does not remove its stochastic trend. Choosing between these descriptions is consequently a substantive modeling issue.

A cointegrating relation may itself include a trend:

yt−a−bxt−ct∼I(0).y_t-a-bx_t-ct\sim I(0).

Here the stationary discrepancy is measured relative to a relationship that moves deterministically over time. This is different from requiring yt−bxty_t-bx_t to fluctuate around a fixed mean. Intercepts can also play different roles: some specifications allow nonzero means in cointegrating errors, while others permit deterministic drift in the variables’ levels. These alternatives lead to different rank-test distributions.

Structural breaks create a separate difficulty. Suppose a stable pre-break relation changes from yt=a1+bxt+uty_t=a_1+bx_t+u_t to yt=a2+bxt+uty_t=a_2+bx_t+u_t. A full-sample regression imposing one intercept can produce residuals that appear highly persistent. A change in the slope is still more consequential because the incorrectly imposed coefficient can leave an integrated component in the fitted error.

Gregory–Hansen tests, introduced in 1996, extend residual-based cointegration testing to alternatives involving a single unknown break in specified features of the relationship. Candidate break dates are searched over a trimmed portion of the sample, and the resulting statistics require break-specific critical values. Their null is no cointegration; rejecting it supports the relevant break-cointegration alternative, not an unrestricted claim that every aspect of the model is stable.

Known interventions can also be represented by appropriate dummy variables or segmented deterministic components. However, a one-time additive outlier, a permanent level shift, a slope change, and a change in adjustment speed are not interchangeable. Each affects the data differently and requires a corresponding model. Searching over many break dates or specifications without accounting for that search changes the interpretation of conventional p-values.

Other forms of nonstationarity matter as well. Monthly or quarterly series can contain seasonal unit roots, making cointegration at seasonal frequencies relevant. Changes in measurement definitions, temporal aggregation, and revisions to historical data can also alter apparent relationships. A finding that depends entirely on one sample endpoint or one deterministic specification is statistically different from a relationship that persists across plausible specifications, even when both produce a nominal rejection in a single test.

Economic interpretation and applications

Cointegration gives a precise statistical meaning to some proposed long-run economic restrictions. If logarithms of consumption ctc_t and income yty_t are I(1)I(1), a relation

ct−a−byt∼I(0)c_t-a-by_t\sim I(0)

describes a stable log-linear association. If b=1b=1, the log consumption–income ratio is stationary around aa. If b≠1b\neq1, the stationary object is instead a weighted log difference. The empirical cointegration question and the theoretical restriction on the slope are therefore separate.

Purchasing power parity provides another example. Let sts_t be the logarithm of the domestic-currency price of foreign currency, with domestic and foreign log price levels ptp_t and pt∗p_t^*. The log real exchange rate is

qt=st+pt∗−pt.q_t=s_t+p_t^*-p_t.

A stationary qtq_t corresponds to the specified cointegrating vector (1,1,−1)(1,1,-1), allowing for a constant mean. Finding some unrestricted stationary combination of the three variables is weaker than establishing precisely this restriction. Transaction costs, differences in goods baskets, and persistent relative-price changes also complicate interpretation.

Interest rates at different maturities supply a further application. If a long and a short interest rate are both I(1)I(1), a stationary spread implies cointegration with coefficients proportional to (1,−1)(1,-1). Expectations-based term-structure theories provide motivations for investigating such restrictions, but cointegration alone does not establish the full theory: expectations, risk premia, and short-run adjustment impose additional conditions.

In financial markets, cointegration is used to study linked asset prices and relative-value relationships. Two price series may share a persistent component while a weighted spread is stationary. Yet stationarity does not imply costless arbitrage or guaranteed trading profits. Financing costs, trading frictions, estimation error, the speed of adjustment, and breaks in the relationship can all matter. The statistical spread is not automatically an attainable riskless return.

Cointegration also appears in models of money demand, international trade, commodity markets, and energy demand. For example, a long-run demand equation may combine logarithms of real balances, income, and an opportunity-cost variable. Economic theory supplies candidate variables and coefficient restrictions; the time-series model evaluates whether their joint behavior is compatible with the proposed stochastic relationship.

The language of “equilibrium” must nevertheless be interpreted carefully. A stationary residual indicates a stable probabilistic discrepancy under the model, not necessarily market clearing or optimal behavior. Nor does cointegration establish causation. Common omitted forces, accounting relations, and institutional arrangements can all generate stable combinations. Even when an error-correction term predicts subsequent changes, identifying the effect of a policy intervention requires additional assumptions.

Nonlinear, fractional, and higher-order extensions

Standard cointegration is linear in the included variables, although those variables may already be transformed, for example by taking logarithms. A relationship stationary in log prices need not correspond to a stationary linear combination of untransformed prices. Nonlinear extensions allow a wider class of equilibrium relationships or adjustment mechanisms, but these are distinct questions: a linear long-run relation can coexist with strongly nonlinear short-run adjustment.

Threshold cointegration allows responses to depend on the size or sign of the disequilibrium. For instance, a system may adjust only when a price discrepancy exceeds a transaction-cost band. A schematic scalar equation is

Δyt={λ1et−1+⋯ ,et−1≤c,λ2et−1+⋯ ,et−1>c.\Delta y_t= \begin{cases} \lambda_1 e_{t-1}+\cdots, & e_{t-1}\leq c,\\ \lambda_2 e_{t-1}+\cdots, & e_{t-1}>c. \end{cases}

The threshold cc and regime-specific coefficients allow different behavior in different regions. Smooth-transition models replace the abrupt switch with a continuous transition function. Stability requires conditions on the complete dynamics, not merely the appearance of negative coefficients in selected regimes.

Fractional integration generalizes differencing to noninteger orders through the operator (1−L)d(1-L)^d, where Lxt=xt−1Lx_t=x_{t-1}. Fractional cointegration occurs when a linear combination has an integration order lower than that of the component series. The equilibrium error can retain long memory and need not be I(0)I(0). Depending on the orders involved, deviations may be stationary but very persistent, or nonstationary despite having less persistence than the original series.

Systems containing I(2)I(2) variables require another extension. A linear combination of I(2)I(2) levels may be I(1)I(1), and combinations involving both levels and first differences can produce stationary expressions. The resulting restrictions are not captured by an ordinary I(1)I(1) VECM. Polynomial cointegration formalizes relations involving lag polynomials, and specialized I(2)I(2) analysis separates the relevant ranks and integration reductions.

Panel cointegration studies long-run relationships across multiple units observed over time, such as countries, firms, or regions. Pooling can increase information, but the model must distinguish common slopes from unit-specific slopes and account for dependence across units. A common international shock can make nominally separate observations strongly dependent, invalidating procedures derived under cross-sectional independence.

These extensions change the meaning of a successful or unsuccessful standard test. Failure to find linear I(1)I(1) cointegration does not rule out every nonlinear or fractional long-run relation. Conversely, applying increasingly flexible alternatives creates additional estimation and multiple-testing problems. The relevant extension depends on a specified feature of the process—such as threshold adjustment, heterogeneous units, or higher integration order—rather than simply on the desire to obtain a stationary fitted error.

Historical development

The emergence of cointegration followed a broader reassessment of regression with persistent economic data. G. Udny Yule’s 1926 discussion of “nonsense correlations” was an early warning that relationships between time series could look compelling for reasons unrelated to a substantive connection. The later development of stochastic-process theory made it possible to distinguish different sources of persistence more formally.

Clive Granger and Paul Newbold’s 1974 paper on spurious regressions demonstrated the problem forcefully through simulated integrated series. Their results showed that familiar regression diagnostics could be deceptive when applied to unrelated persistent processes. Peter Phillips’s 1986 theoretical analysis subsequently explained important aspects of spurious-regression asymptotics. This research established that the difficulty was not merely a few unfortunate samples or a minor correction to conventional standard errors.

At the same time, applied dynamic models had already used error-correction mechanisms. James Davidson, David Hendry, Frank Srba, and Stephen Yeo’s 1978 study of United Kingdom consumption is a prominent example. Such models combined changes in economic variables with lagged departures from a longer-run relationship. Cointegration later supplied a stochastic foundation for determining when that combination was appropriate.

Granger’s 1981 work introduced the cointegration concept into the analysis of time-series relationships. Robert Engle and Granger’s 1987 paper, “Co-integration and Error Correction: Representation, Estimation, and Testing,” brought together the representation theorem, estimation results, and a practical residual-based testing procedure. It made explicit why integrated variables could support a meaningful levels relationship when their regression error was stationary.

Johansen’s 1988 likelihood analysis then shifted attention from single equations to the rank structure of multivariate systems. The development of system methods made it possible to estimate several long-run relations, test restrictions on their coefficients, and distinguish the cointegrating space from its chosen economic normalization. Related work by Stock and Watson connected cointegrated systems with common stochastic trends.

Subsequent research developed corrected estimators, more accurate critical values, structural-break tests, nonlinear adjustment models, and fractional and panel extensions. In 2003, Granger shared the Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel with Engle. Granger was recognized for methods of analyzing economic time series with common trends, while Engle was recognized for methods concerning time-varying volatility. These contributions addressed different features of economic data: cointegration concerns persistent relationships among levels, whereas volatility models concern changes in conditional variability.