Chapter 5 Vector autoregression models

Vector autoregression models

109

Chapter 5 Vector autoregression models

Introduction This chapter provides an overview of basic vector autoregression models (VARs) and structural vector autoregression models (SVARs). VARs are dynamic systems of equations that examine the inter-relationships between economic variables, using minimal assumptions about the underlying structure of the economy. They aim to provide good statistical representations of the past interactions between variables. SVARs represent an intermediate method between basic VARs and the more structural approach of traditional macroeconometric models: they combine the statistical methodology of basic VARs with a number of widely accepted restrictions derived from economic theory, and aim thereby to capitalise on the strengths of both approaches. The first section of this chapter outlines the origins of VAR models and their basic structure. Section 2 describes how they have been used at the Bank both for forecasting and as a tool to analyse the effects of various shocks. The Annex to the chapter details the different ways that VARs can be identified for such analysis. 1 The origins of VARs and their basic structure

The main development of VARs as a modelling tool was in the early 1980s, originating from concerns about the validity of some of the assumptions used in traditional macroeconometric models. In particular, Sims (1980)(1) argued that the restrictions used to identify the parameters in traditional modelswhich often took the form of excluding variables or their lags from equations, or assuming that a particular variable was exogenouswere incredible. He contended that theory was rarely sufficiently well defined to justify such exclusion restrictions or exogeneity assumptions,(2) and that such models were likely to be under-identified once these problems were taken into account. As a result, some of the economic interpretations drawn from such models were unlikely to be robust. The identification issue as it relates to VARs is discussed below in Section 2.2 and in the Annex. These concerns led to the development of VARs as an alternative modelling approach. VARs are dynamic systems of equations in which the current level of each variable in the system (eg GDP, unemployment and official interest rates) depends on past movements in that variable and in all the other variables in the system. In contrast with traditional models, such as the Banks macroeconometric model, basic VAR systems make few assumptions about the underlying structure of the economy and instead focus entirely on deriving a good statistical
(1) (2) Most of Sims concerns were not new. For instance, Liu (1960) had already raised many of them. But the poor performance of large-scale models in the wake of the two oil shocks increased the impact of Sims criticisms. For example, Sims argued that many of the identifying restrictions were based on partial-equilibrium analyses that did not hold when all the interactions were considered in a general-equilibrium framework. The identifying restrictions are further undermined when expectations are taken into consideration. The choice of lags included in the equations is also largely arbitrary.

110

Economic models at the Bank of England

representation of the past interactions between economic variables, letting the data determine the model. However, even VARs are not completely devoid of assumptions, since the choice of variables to include in the system and the length of lags allowed represent a type of restriction, which can have important implications. A VAR system can be expressed in the following form:(1) Zt = A1Zt-1 + A2Zt-2 + ...... + ApZt-p + t (1)

where Zt is a vector of endogenous variables at time t, Ai (i = 1, ..., p) are coefficient vectors, p is the number of lags included in the system, and t is a vector of residuals. This system of equations can be thought of as encompassing a number of structural models in an unrestricted way; the (arguably contentious) restrictions that distinguish the structural models from each other are not imposed. The residuals, t, represent the unexplained movements in the variables, reflecting the influence of exogenous shocks (ie shocks that arise outside the assumed model). The residuals represent a composite of the various exogenous shocks affecting the endogenous variables in the underlying structural model. It is not possible, therefore, to derive any economic interpretation from the residuals without transforming equation (1). This issue is discussed in more detail in Section 2 and in the Annex. Estimating a VAR involves choosing which variables to include in the system, and deciding on the number of lags. The results obtained can be sensitive to both of these choices. The number of lags is usually determined by statistical criteria,(2) and variable selection is generally informed by economic theory. These considerations highlight a few of the potential problems in estimating VARs. First, estimation problems increase as the number of variables and lags included in the system rises. More specifically, problems with degrees of freedom will occur if there are large numbers of parameters to be estimated. And the degree of correlation between the lagged variables is likely to reduce the precision of estimated coefficients. The application of economic theory to help determine which variables to include in the VAR is a type of restriction. This implies that VARs are not completely atheoretic. However, such concerns can be addressed by making the theory determining the choice of variables sufficiently general or uncontentious. Finally, it should be noted that if the restrictions imposed by more traditional macroeconometric models are valid, the parameter estimates derived from such models are likely to be more precise than those derived from the VAR. 2 Uses of VARs

The Bank has used VARs in two main ways: to produce forecasts of economic variables (especially inflation), and to examine the effects of economic shocks.

(1) (2)

Deterministic factors such as constants, dummy variables and trends can easily be incorporated. Equation (1) omits these elements to simplify the algebra. Such as sequential-likelihood ratio tests or Akaike/Schwartz information criteria.

Vector autoregression models

111

2.1

Forecasting

Deriving forecasts from VAR models is straightforward. The one period ahead forecast of ) is obtained by substituting past and current values the vector of endogenous variables (Z t+1 in periods t to t+1-p into equation (2): = A Z + A Z + ...... + A Z Z t+1 1 t 2 t-1 p t+1-p ) is obtained by Similarly, the two period ahead forecast for the vector of variables (Z t+2 substituting (Zt+1) and the observed values of the variables in periods t to t+2-p into equation (3): =A Z Z t+2 1 t+1 + A2Zt + ...... + ApZt+2-p (3) (2)

So VAR forecasts project forward the past statistical correlations between the variables in the system. The only assumption they embody is that these correlations will continue to hold in the forecast period. In contrast, forecasts derived from traditional structural models are normally conditioned on assumptions about the future profiles of variables assumed to be exogenous (such as fiscal policy). VAR forecasts avoid the need for such assumptions, by treating each of the variables as endogenous and letting the system generate their forecast profiles. Basic VAR forecasts are therefore unconditional. But if there are strong prior assumptions about the likely future paths of one or more of the variables, it is also possible to generate conditional VAR forecasts. VAR forecasts are useful both in their own right and to cross-check forecasts produced by traditional macroeconometric models. However, as with other techniques, VAR forecasts have several weaknesses. First, as outlined in Chapter 1, they are susceptible to the Lucas critique and more generally are likely to perform poorly following structural breaks (although some traditional macroeconometric models may also be affected by such problems).(1) Second, the inclusion of too many lags or too many variables (sometimes known as overfitting) can lead to VAR-derived forecasts performing poorly, even though the model fits the data well in the estimation period. This occurs if the lags of variables in the system pick up non-systematic relationships between the variables (ie spurious relationships or noise in the data), as well as the systematic (ie robust) relationships.(2) So unless the set of variables and lag lengths is kept relatively smallwhich would reflect implicitly theoretical considerationsforecasts derived from VARs are likely to be inaccurate. This is one reason why VARs are usually used only to produce relatively short-term forecasts; at longer horizons, the forecasting performance of VARs often deteriorates rapidly.

(1) (2)

Indeed, Clements and Hendry (1995) argue that VAR models are likely to be less affected by structural breaks than structural models, as they are often formulated in the first differences of the variables, which makes them less vulnerable to intercept-shift problems. This problem led Doan, Litterman and Sims (1984) to suggest the Bayesian VAR (BVAR) technique as a means of obtaining a more parsimonious representation of the data. As Holden (1995) explains, BVARs achieve this by making it easier for lagged dependent variables to enter an equation than for other lagged variables. Doan, Litterman and Sims (1984), Wallis et al (1987), and Todd (1990) provide more detailed discussions of BVARs.

112

Economic models at the Bank of England

The Bank has used VAR models to generate forecasts of inflation to compare with those produced using the macroeconometric model. VAR models using both monthly and quarterly data have been constructed for this purpose, and are often re-estimated and re-specified. The monthly models have the advantage of using more timely information, but the disadvantage of being estimated with more noisy data. The latest version of the quarterly model was estimated over the period 1974 Q1 to 1998 Q4 and contains six lags of seven variables: the annual growth rates of retail sales, GDP and industrial production, and the first difference of the annual growth rates of RPIX, M4, M0 and whole-economy unit labour costs. Chart 1 shows a series of two year ahead inflation forecasts derived from this model, beginning in 1997 Q1. As the chart shows, the forecasts within the sample period have generally been quite close to the actual outturn. Chart 1 Unconditional forecasts of RPIX inflation using a VAR model
Per cent 10 9 8 Actual RPIX inflation 7 6 5 4 3 2 1 0 1990 91 92 93 94 95 96 97 98 99 2000

2.2

Using VARs to examine the effects of economic shocks

VARs can also be used for purposes other than forecasting.(1) In particular, they are frequently used to investigate the effects of shocks on a system of variables. However, as with more traditional macroeconometric models, such exercises require the imposition of restrictions on the model. This section reviews this issue, known as identification, and discusses some examples of the use the Bank has made of VARs for investigating the effects of economic shocks. Identification: VARs and structural VARs If movements of the endogenous variables within a VAR system are viewed as reflecting the effects of exogenous shocks, the VAR can be used as a tool to examine these shocks. The
(1) Hendry and Mizon (1993) argue that one way of evaluating traditional macroeconometric models is by their ability to encompass VAR models. And VARs are the basis for the widely used Johansen (1988) cointegration test.

Vector autoregression models

113

error terms in a basic VAR are complex functions of the shocks hitting the system, and have no economic interpretation. However, the different shocks and their effects can be disentangled. This is achieved by placing identifying restrictions on the VAR to transform it into the structural moving-average representation shown in equation (4). The details of this transformation are given in the Annex. Zt = C0et + C1et-1 + C2et-2 + C3et-3 + ........ + Cmet-m (4)

In equation (4), Ck (k = 0, ..., m) represent vectors of parameters determining the current and lagged effects of the shocks, et-k (k = 0, ..., m), on the endogenous variables. The transformation used to obtain equation (4) involves imposing n2 restrictions on a system of n variables.(1) These restrictions deliver two desirable properties. First, the shocks (et) become orthogonal to each other. This means that they can be interpreted as representing independent economic phenomena. Second, the dynamic structure of the effects of the shocks, represented by the Ck vectors, can be given an economic interpretation. The benefit of imposing these identifying restrictions is that they allow the structural parameters of the system to be uncovered. However, as for structural macroeconometric models, the imposition of these restrictions may be contentious. For example, Cooley and LeRoy (1985) argued that the types of identifying restrictions proposed by Sims (1980) often implied implausible economic relationships. The need to formulate robust identifying restrictions led to the development of structural VARs (SVARs).(2) These models only incorporate restrictions that are widely accepted (ie common to a variety of theoretical models). SVARs represent an intermediate approach between the almost purely statistical method of basic VARs and the more structural approach of traditional econometric models. An important difference between traditional models and SVARs is that, whereas the former focus directly on the relationships among economic variables, SVARs try to obtain information about the shocks driving movements in the endogenous variables and then use this information to make inferences about the co-movements between them. SVARs are especially useful when the exogenous shocks are unobservable. For example, Bean (1994) argues that traditional investigations of changes in unemployment suffer from problems of obtaining reliable measures of their determinants. The SVAR approach helps to circumvent this problem by estimating the shocks from the reduced-form residuals and the movements in the endogenous variables. It then allows the relative importance of the different structural shocks in the movements of the endogenous variables to be determined. SVARs are also useful when theory suggests that different structural shocks should generate different co-movements between the endogenous variables, since they allow the dynamic response of each variable to each shock to be examined. Examples of both these techniques are given later in this chapter.

(1) (2)

See Hamilton (1994). The main types of identification procedures used in SVARs are discussed in more detail in the Annex.

114

Economic models at the Bank of England

However, there are several potential problems with SVARs. First, as Blanchard and Quah (1989) recognised, the economy may be hit by more shocks than are identified in the SVAR system. SVARs will therefore produce reliable results only if the most important types of shock are identified in the system. Second, the fact that most estimated SVAR systems are exactly identified means that many of the formal tests applied to traditional structural models cannot be implemented. Consequently, SVAR modellers are forced to use less formal and less robust tests to examine the economic plausibility of the structural shocks identified; the estimated dynamic responses of the variables to the shocks; and the estimated sources of endogenous variable co-movements. Examples of Bank work using VARs and SVARs (i) Money VARs

Research undertaken by Astley and Haldane (1995, 1997) used a series of bivariate VARs to examine the leading-indicator properties of monetary and credit aggregates for measures of activity and inflation. Statistical tests(1) were used to analyse whether the monetary aggregates contained significant information about future movements in activity (or inflation), over and above the information contained in past movements of activity (or inflation). The economic content of the statistically significant relationships was then assessed by examining impulse response functions. These functions trace out the dynamic effects of a particular exogenous shock on an endogenous variable. The results provided some stylised facts about the short-run correlations between monetary variables and activity indicators. For example, broad money (M4) holdings of industrial and commercial companies (ICCs)(2) were found to contain statistically significant leading-indicator information about fixed investment by companies. The associated impulse response functions were consistent with a plausible economic story. Chart 2 illustrates the dynamic paths of aggregate and manufacturing fixed investment following a 1 percentage point increase, relative to base, in ICCs M4 balances. The relationship is positive for around eight quarters, with a peak effect of almost 0.3 percentage points. Astley and Haldane rationalised these relationships as reflecting ICCs increasing their money balances ahead of planned but lumpy investment outlays.(3) The Bank has used the above relationships in two ways. First, they have helped to interpret the implications of developments in monetary and credit aggregates. For example, the May 1997 Inflation Report associated the strong growth in ICCs investment in 1996 with the growth in their money holdings in 1995. Second, the stylised relationships uncovered have motivated further research at the Bank, for example on the demand for money at a sectoral level.(4)

(1) (2) (3) (4)

These were tests for Granger causality. Unlike other VAR tools, Granger-causality tests are not sensitive to identification issues. Following the adoption of the ESA95 statistical classifications, ICCs are now included in the private non-financial corporations (PNFC) sector. This example illustrates clearly how Granger causality is not a test of true economic causality. See Thomas (1997a and 1997b).

Vector autoregression models

115

Chart 2 Impulse response of investment to shock to ICCs M4 balances

Percentage points 0.4

0.3

0.2

0.1 Fixed investment

+
0.0

_
0.1

0.2 Fixed investment in manufacturing 0.3 0 2 4 6 8 10 12 14 16 18 Quarters after shock Sources: Astley and Haldane (1995, 1997).

(ii)

Monetary policy transmission mechanism

A large number of papers have used VARs to examine aspects of the monetary policy transmission mechanism. Dale and Haldane (1995) analysed the speed with which monetary policy changes feed through to the corporate and personal sectors of the UK economy. Ganley and Salmon (1996) extended this work to examine the output responses to changes in monetary policy of 24 sectors of the UK economy. Chart 3 illustrates the estimated output effects on the four main sectors of the economy of an unexpected increase in official interest rates of 1 percentage point. The results of this research were used in the November 1997 Inflation Report, which noted that the tightening of monetary policy in 1997 was likely to affect the construction and production sectors sooner and more markedly than other sectors. Further research at the Bank has used an SVAR to analyse the effects of four different types of monetary disturbance. First, permanent monetary policy shocks were distinguished from temporary monetary policy shocks. The former were defined to reflect shifts in the authorities underlying inflation target and the latter to reflect either policy errors or transitory changes in the parameters of the authorities reaction function. Second, two types of velocity shock were identified: a financial intermediation shock and a money demand shock. The intermediation shock was defined to reflect changes in the provision of credit by the banking system and the degree of financial liberalisation, and the money demand shock was defined to reflect changes in liquidity preferences. This research highlighted the importance of understanding the causes of monetary disturbances. In particular, the impulse response results indicated that each of the four types of monetary shock discussed above generates different dynamic relationships between money, asset prices, output and inflation. To illustrate this, Charts 4 and 5 compare the

116

Economic models at the Bank of England

Chart 3 Output responses of the major industrial groups to an unexpected monetary tightening
Per cent 1.0 0.5

Agriculture, forestry and fishing

+
0.0

_
0.5 Service industries 1.0 1.5 Construction Production industries 2.5 3.0 0 3 6 9 12 15 18 21 24 27 Quarters after shock Source: Ganley and Salmon (1996). 2.0

effects of permanent and temporary expansionary monetary policy shocks (of magnitude one standard deviation) on official short-term interest rates, real money balances, output and inflation. Both types of shock lead to an initial fall in official interest rates. This is associated with a rise in real money balances, output and inflation in the short to medium term. But there are also noticeable differences between the effects of the permanent and temporary shocks. In particular, the permanent shock has a long-run effect on the steady-state rate of inflation and of real money balances, since higher inflation leads to higher nominal interest rates and a higher opportunity cost of holding real money balances. Chart 4 Responses to a permanent monetary policy shock
Per cent Inflation (a) Interest rates (a) Output 1.2 0.8 0.4

+ _
0.0 0.4 0.8 1.2

Real money balances 1.6 2.0 2.4

12

15 18 21 24 Quarters after shock

27

30

33

36

39

(a) Percentage points.

Vector autoregression models

117

In contrast, the temporary shock has only transitory effects on the growth rate of nominal variables, leaving real money balances unaffected in the long run. Chart 5 Responses to a temporary monetary policy shock
Per cent 1.0

Real money balances

0.8 0.6

0.4 Inflation (a) 0.2 Output

+
0.0

_
Interest rates (a) 0.2

12

15 18 21 24 Quarters after shock

27

30

33

36

39

0.4

(a) Percentage points.

(iii)

Sterling exchange rate movements

The Bank has also used SVARs to examine sterling exchange rate movements. In particular, Astley and Garratt (1996, 1998)(1) used forecast-error variance decompositions (FEVD) to determine the relative importance of three structural shocks for the movements in sterling bilateral exchange rates, UK consumer prices (relative to their foreign equivalents) and UK GDP (relative to its foreign equivalents). Four foreign countries were considered in this analysis: France, Germany, Japan, and the United States. Table A shows the results for the DM/ real exchange rate. These suggest that a high proportion of sterlings movements against the Deutsche Mark over the period 197394 reflected the impact of real goods market shocks, such as shifts in relative consumption or investment, and possibly the effects on aggregate demand of unexpected changes in monetary policy. In contrast, movements in UK consumer prices relative to Germanys were largely associated with shifts in the money demand and supply schedules. Astley and Garratt also examined the sources of sterling exchange movements during sub-samples of the estimation period, using historical decompositions.(2) Chart 6 shows the decompositions of DM/ real exchange rate movements between 1986 and 1991. The relative importance of each type of shock is estimated by examining how closely the
(1) (2) This work was based upon the Clarida and Gali (1994) analysis of US data. However, other techniques for examining exchange rate movements, such as the uncovered interest parity condition, are also used at the Banksee Box D on page 19. See Burbidge and Harrison (1985) for details.

118

Economic models at the Bank of England

Table A Decomposition of DM/ real exchange rate movements

(95% confidence intervals in parentheses) No of quarters after shock 1 4 8 16 40 60 Percentage of exchange rate variation accounted for by: aggregate supply goods market nominal (money) shock shock shock 9.3 13.6 15.7 17.1 17.9 18.1 (1.317.3) (2.929.4) (2.928.5) (2.831.4) (2.932.9) (3.133.1) 72.0 78.2 80.3 81.1 81.4 81.4 (55.788.3) (64.292.2) (66.494.2) (66.595.7) (66.896.0) (66.995.9) 18.7 8.2 4.0 1.8 0.7 0.5 (4.233.2) (0.025.5) (0.57.5) (0.43.2) (0.21.2) (0.20.8)

Sources: Astley and Garratt (1996, 1998).

movements in the DM/ exchange rate attributable to each shock correspond with the total DM/ movement.(1) In line with the whole sample period FEVD results, the historical decompositions suggest that the real appreciation of sterling against the Deutsche Mark in the late 1980s largely reflected positive real goods market (ie demand) shocks. This may have been related to the observed shift in relative domestic demand towards the United Kingdom at that time. Chart 6 Historical decomposition of DM/ real exchange rate movements
Log real DM/ exchange rate (news) 0.4

Demand shock component

0.3

0.2 Exchange rate (news) 0.1 Nominal shock component

+
0.0

Supply shock component 0.1

0.2 1986 87 88 89 90 91

Sources: Astley and Garratt (1996, 1998).

More generally, the impulse response results may suggest that the partial-equilibrium prediction that an exchange rate depreciation will generate a rise in relative consumer prices, through its effect on import prices, is not necessarily consistent with the result that would be obtained from a general-equilibrium analysis. This may reflect the fact that the shocks
(1) The total path is not the actual exchange rate movement. Rather, it is the actual movement minus the SVARs unconditional forecast, or base projection, formed on the basis of a few initial periods of shocks. It is this forecast error, or news, that is decomposed into the proportions attributable to each of the structural shocks after the initial periods.

Vector autoregression models

119

underlying sterling exchange rate movements affect UK (relative) consumer prices through a large number of channels, in addition to the import price channel. For example, structural shocks will also affect prices through their impact on the level of excess supply or demand in the economy; depending on the type of shock, such channels can either reinforce or offset the import price effects. (iv) Inflation and real disequilibria

Quah and Vahey (1995) used an SVAR to estimate a time series of core inflation in a system consisting of output and RPI inflation. The core inflation measure aimed to strip out erratic movements in inflation to obtain an estimate of underlying inflationary pressures. Quah and Vahey achieved this by modelling core inflation as the movements in RPI inflation associated with any shock, constrained to have no long-run effect on the level of output. This approach is consistent with a vertical long-run Phillips curve (see Chapter 3). The authors argue that their approach has a stronger economic content than alternative statistical methods of obtaining core inflation measures. Quah and Vahey found, among other things, that measured RPI inflation overstated core inflation around the time of the second oil-price shock and in the mid to late 1980s, but understated core inflation in the early 1980s. SVARs can also be used as an alternative way of modelling the real disequilibria mentioned in Chapters 2 and 3. Sterne and Bayoumi (1993, 1995) used a series of bivariate SVARs to estimate time series of output gaps for 21 OECD countries. They defined their output gap measure as the output movements associated with shocks constrained to have no long-run effects on output. Astley and Yates (1999) extended this approach by estimating, within a single SVAR representation of the UK economy, time series of the output gap and the unemployment gap.(1) Each gap was defined as the movements in output and unemployment associated with shocks that theory indicates should have no long-run effects on the variable. This approach aims to address several potential problems associated with alternative methods of producing such gap time series. In particular, the technique used by Astley and Yates is more closely based on economic theory than statistical approaches such as the Hodrick-Prescott filter, and takes account of the fact that the gaps are simultaneously and endogenously generated. Chart 7 presents point estimates of the resultant output and unemployment gaps.(2) As with the unemployment gaps derived using the Kalman-filter technique in Chapter 3, there is a considerable amount of uncertainty surrounding these estimates, and they are therefore best interpreted as indicating the broad pattern of movements. Comparing the unemployment gap generated by the SVAR with the estimates derived in Chapter 3 using the Kalman-filter technique, one can see that the profile of movements is quite similar, even if the levels differ. As noted in Chapter 1, an important reason for using a range of different estimation techniques is precisely to provide such tests of the robustness of the results derived from various models.
(1) (2) The Astley and Yates approach draws on the techniques used by Bean (1992) and Dolado and Jimeno (1995), which were used to examine the sources of EU unemployment movements. These estimates do not represent a Bank view as to the likely size of the output and unemployment gaps.

120

Economic models at the Bank of England

Chart 7 SVAR estimates of gaps in the United Kingdom

Output gap Percentage of GDP 1.0 0.5

+ _0.0
0.5 1.0 1.5 1975 Q3 79 Q3 83 Q3 87 Q3 91 Q3 95 Q3 Percentage points 4 2

Unemployment gap

+0 _
2 4 6 1975 Q3 79 Q3 83 Q3 87 Q3 91 Q3 95 Q3 Source: Astley and Yates (1999).

It is also apparent from Chart 7 that though the output and unemployment gaps derived from this method have tended to move in parallel with each other recently, there have been periods when their movements diverged. This may reflect measurement error in the data or economic considerations (eg labour hoarding), indicative of the restrictiveness of the conditions that are needed for the two gaps to contain the same information. Conclusions This chapter has outlined how the Bank uses VARs and SVARs in its economic analysis and forecasting to complement more traditional macroeconometric modelling methods. An important difference between SVARs and traditional macroeconometric models is that the latter focus directly on the relationships among economic variables, whereas the former try to uncover information about the shocks driving movements in the economic variables, and then make inferences about the co-movements between those variables. VARs and SVARs are particularly useful when the determinants of endogenous variable movements are unobservable. Equally, however, there are situations where more traditional approaches will have advantages over SVARsfor example, when the type of identifying restrictions used by SVARs appears inappropriate, or when insufficient realistic identifying restrictions are available. In addition, there are certain types of policy analyses (such as examining the effects of Taylor monetary policy rules) that are difficult to formulate in terms of a VAR/SVAR. The choice between using a VAR/SVAR or more structural econometric methods therefore has to be determined case by case.

Vector autoregression models

121

Annex Identification of VARs: the different schemes available This Annex details the alternative methods available for identifying VARs, so that the effects of exogenous shocks on the endogenous variables in the system can be examined. Section 1 showed that the basic VAR has the following representation: Zt = A1Zt-1 + A2Zt-2 + ...... + ApZt-p + t This can be rewritten in more compact form using the lag operator (L): A(L)Zt = t (A2) (A1)

The VAR approach views movements in the endogenous variables as fundamentally reflecting the effects of exogenous shocks hitting the economy. Unfortunately, the residuals, t, in equation (A2) are complex functions of the effects of these structural shocks, and therefore have no economic interpretation. The structural shocks can, however, be recovered via a transformation process that imposes a number of identifying restrictions on the VAR. This Annex explains the various methods available to estimate the exogenous shocks and their dynamic effects. The first step in all the methods is to invert the autoregressive representation of equation (A2) to obtain the following moving-average representation: Zt = B(L)t where B(L) = A(L)-1 This transformation does not introduce any economic structure. Hence equation (A3) is still a reduced-form representation. The aim is to obtain the structural moving-average representation: Zt = C(L)et which can be written in its full form as: Zt = C0et + C1et-1 + C2et-2 + C3et-3 + ........ + Cmet-m (A5) (A4) (A3)

Equation (A5) is obtained from equation (A3) by imposing two desirable properties.(1) First, the exogenous shocks, et, are transformed to be orthogonal to each other. This means that they can be interpreted as representing independent economic phenomena. Second, restrictions are imposed on the effects of the exogenous shocks on the endogenous variables, represented by the Ck vectors, so that they embody some form of economic structure.(2) For
(1) (2) See Hamilton (1994) for the details of this transformation. In equation (A5), the Ck represent vectors of parameters determining the current effects on the endogenous variables of the exogenous shocks (et) that occurred k periods ago, and m is the number of periods (potentially infinite) it takes for the effects of a shock on an endogenous variable to asymptote to its long-run effect.

122

Economic models at the Bank of England

a system of n variables, imposing these two desirable properties requires the imposition of n2 restrictions, in order to achieve exact identification and obtain the structural moving-average representation. Imposing the condition that the exogenous shocks should be orthogonal to each other provides n(n-1)/2 of these restrictions, leaving n(n+1)/2 restrictions to be found from elsewhere. Sims (1980), and many subsequent papers, generated the remaining restrictions in two steps. First, n restrictions were provided by imposing unit coefficients on the principal diagonal of the C0 matrixmeaning that the contemporaneous effect of the jth exogenous shock on the jth endogenous variable is the standard deviation of that shock. Second, the final n(n-1)/2 restrictions were provided by setting the elements above the principal diagonal of the C0 matrix to zero. This approach, which is known as the Choleski decomposition, implies that the contemporaneous interactions between the exogenous shocks and the endogenous variables are characterised by a Wold (1938) causal chain. However, Cooley and LeRoy (1985) pointed out several problems with Sims identification procedure. They argued that the particular economic structure implied by the restrictions was often not very plausible, and highlighted that the results obtained were often sensitive to the ordering of the variables in the system (ie the particular Wold causal chain selected). Unfortunately, economic theory generally provides little guidance as to the most appropriate ordering. These problems led to the development of a number of more sophisticated and flexible identification methods. These became known as structural VARs (SVARs), because their identifying restrictions could be traced to rigorous theoretical underpinnings more easily than the Sims approach. This explicit link to theory means that SVARs are less differentiated from traditional structural models than VARs. The distinction is that SVARs aim to utilise the subset of the theory-derived restrictions used by structural models that are least contentious (the ones that hold in a wide variety of models). The four main identification schemes are: (i) The Bernanke (1986) method. This approach follows that of Sims, by imposing n(n-1)/2 zero restrictions on the C0 matrix. Importantly, however, it is more general and flexible, since it does not impose a contemporaneous Wold causal chain.

(ii) The Blanchard and Quah (1989) method. This approach achieves identification by imposing zero restrictions on the long-run effect of structural shocks on the level of particular endogenous variables. This formulation requires that the levels of the endogenous variables that such restrictions are imposed on are non-stationary, but not cointegrated with any of the other non-stationary endogenous variables in the system, so that their first differences enter the reduced-form VAR. The long-run effect of the shocks on the level of the endogenous variable, denoted by C(1), then equals the sum of the (first-difference) Ck matrices in (A4): C(1) = Ck
k =1 m

(A6)

Vector autoregression models

123

The Blanchard and Quah approach achieves identification by imposing n(n-1)/2 theory-derived zero restrictions on the C(1) matrix.(1) The main advantage of this approach is that such restrictions are often less contentious than contemporaneous restrictions. (iii) The Gali (1992) method. This uses a combination of the Bernanke contemporaneous restrictions and the Blanchard and Quah long-run restrictions. (iv) The King, Plosser, Stock and Watson (1991) method. This takes account of any cointegrating relationships among the non-stationary variables included in the system. The presence of cointegrating relationships means that, rather than estimating a VAR such as (A1) in the first stage of the method, it is necessary to estimate a vector error-correction model (VECM) of the form: A(L)Yt = 'Yt-1 + t (A7)

where is the matrix of cointegrating vectors and is the loading matrix.(2) Cointegration signals that some of the non-stationary variables move together in the long run, because they are driven by a common set of permanent shocks. So if r cointegrating vectors are present in a system of n variables, then only n-r permanent shocks will be present. This means that the matrix of long-run effects of shocks, C(1) in equation (A6), will have dimensions of n x n-r, rather than n x n. The nature of the long-run relationships between the variables (the coefficients in the matrix) also imposes some structure on the C(1) matrix. King, Plosser, Stock and Watson show that the n-r permanent shocks can be identified by isolating the free parameters of the C(1) matrix into a smaller n-r x n-r matrix using the long-run and/or contemporaneous restrictions described above. The remaining transitory shocks driving the system can also be identified using contemporaneous restrictions, although this is not necessary if only the permanent shocks are of interest. Once identified, the structural moving-average representation can be used in several ways to examine the impact of the exogenous shocks on the endogenous variables. First, the time series of the exogenous shocks can easily be obtained. Second, impulse response functions can be used to trace out the dynamic effects of a particular exogenous shock on a particular endogenous variable over time. They are obtained by considering the same row of the Ck vectors in equation (A5) as k increases. Third, forecast-error variance decompositions can be used to determine the proportion of the forecast-error variance of each endogenous variable attributable to each shock at different forecast horizons. Hamilton (1994) shows that these are basically non-linear transformations of the impulse response functions. Finally, historical decompositions can be used to assess the relative importance of different exogenous shocks to sub-sample endogenous variable movements. As Burbidge and Harrison (1985) explain, historical decompositions combine the estimated structural-shock time series with the estimated impulse responses.
(1) (2) The remaining n restrictions are provided by normalising the variances of each of the orthogonal shocks to unity. The presence of such cointegrating relationships means that it is more complicated to invert a VECM than a VAR.