1 Slides to Lecture 3 of Introductory Dynamic Macroeconomics. Linear Dynamic Models (Ch 2.5-2.8 of IDM) The error correction model (Ch 2.5 in IDM) In lecture 2 we • used the ADL model to make the concept of dynamic multiplier precise, • Saw also that the simple ADL model can be generalized, and • how macroeconomic hypotheses can be specified as ADLs. Ragnar Nymoen University of Oslo, Department of Economics In this lecture, start by recalling the typology: January 30, 2006 2 1 Table 2: A model typology. Type Equation Restrictions And concentrate on the ECM and the Homogenous ECM. These models are particularly useful in dynamic macroeconomics. ADL yt = β0 + β1xt + β2xt−1 + αyt−1 + εt. None Static yt = β0 + β1xt + εt. β2 = α = 0 Random walk yt = β0 + yt−1 + εt β1 = β2 = 0, α = 1. DL yt = β0 + β1xt + β2xt−1 + εt α=0 can be transformed to error correction form. Differenced data1 ∆yt = β0 + β1∆xt + εt β2 = −β1, α = 1 We then give an interpretation. ECM ∆yt = β0 + β1∆xt + (β1 + β2)xt−1 +(α − 1)yt−1 + εt None We start by showing how the ADL model Homogenous ECM ∆yt = β0 + β1∆xt +(α − 1)(yt−1 − xt−1) + εt 1 The difference operator ∆is defined as ∆zt ≡ zt − zt−1. 3 yt = β0 + β1xt + β2xt−1 + αyt−1 + εt. β1 + β2 = −(α − 1) 4 (1) The ECM Transformation of ADL to an error correction model (ECM) ∆yt = β0 + β1∆xt + (β1 + β2)xt−1 + (α − 1)yt−1 + εt brings to the open what is implicit in the ADL, namely that the growth of yt is explained by the growth of the explanatory variable and the past levels of xt and yt. Step 1: Subtract yt−1 on both sides of the ADL equation, which gives yt − yt−1 = β0 + β1xt + β2xt−1 + (α − 1)yt−1 + εt Step 2: Add and subtract β1xt−1 on the right hand side: yt − yt−1 = β0 + β1(xt − xt−1) + (β1 + β2)xt−1 + (α − 1)yt−1 + εt Step 3: Use the difference operator ∆, meaning for example ∆yt = yt − yt−1, and write the ECM as: ∆yt = β0 + β1∆xt + (β1 + β2)xt−1 + (α − 1)yt−1 + εt (3) The occurrence of both a variable’s growth and its level is a defining characteristic of genuinely dynamic models. In 1929 Frisch put this down as a definitional trait of dynamic models: (2) Step 1-3 is a re-parameterization: If the values y0, y1, y2, y3.... satisfy (1), they also satisfy (2). “A theoretical law which is such that it involves the notion of rate of change or the notion of speed of reaction (in terms of time) is a dynamic law. All other laws are static.” The transformation ADL to ECM shows that Frisch was right! 6 5 The name error correction, stems from the fact that If yt and xt are measured in logarithms. ∆yt and ∆xt are their respective growth rates (check the appendix to IDM if in doubt For example and the ECM version of the estimated ADL equation (1.18) in IDM is 7 embodies adjustments of yt with respect to deviations from the long-run equilibrium relationship between y and x. To make this interpretation clear, collect yt−1 and xt−1 inside a bracket: ½ Ct − Ct−1 Ct − Ct−1 ∆ ln Ct = ln(Ct/Ct−1) = ln(1 + )≈ . Ct−1 Ct−1 ∆ ln Ct = 0.04 + 0.13∆ ln(IN Ct) + 0.21 ln IN Ct−1 − 0.21 ln Ct−1 ∆yt = β0 + β1∆xt + (β1 + β2)xt−1 + (α − 1)yt−1 + εt (4) ¾ β + β2 ∆yt = β0 + β1∆xt − (1 − α) y − 1 + εt, (5) x 1−α t−1 and assume the following long-run relationship for a situation with constant growth rates (possibly zero) in y and x: y ∗ = k + γx, (6) where y ∗ denotes the steady-state equilibrium of yt . We want equation (5) to be consistent with a steady-state, and the slope coefficient γ must therefore be equal to the long-run multiplier of the ADL-ECM, that is: β + β2 γ = δlong−run ≡ 1 , −1<α<1 (7) 1−α 8 Since y ∗ = k + γx, the expression inside the brackets in (5) can be rewritten as β + β2 y− 1 (8) x = y − γx = y − y ∗ + k. 1−α Using (8) in (5) we obtain ∆yt = β0 − (1 − α)k + β1∆xt − (1 − α) {y − y ∗}t−1 + εt, As a point of detail, consider the steady-state situation with constant growth ∆xt = gx and ∆yt = gy , and zero disturbance: εt = 0 (the mean of the disturbance term). Imposing this in (9), and noting that {y − y ∗}t−1 by definition of the steady-state equilibrium, give − 1 < α < 1 (9) showing that ∆yt is explained by two factors: 1. the change in the explanatory variable, ∆xt, and 2. the correction of the last period’s disequilibrium y − y ∗. One can argue that it would be more logical to use the acronym ECM for equilibrium correction model, since the model implies that ∆yt is correcting in an equilibrating way. But the term error correction seems to have stuck. gy = β0 − (1 − α)k + β1gx, This means that k in the theoretical model y ∗ = k + γx depends on the two growth terms. −gy + β0 + β1gx k= , if − 1 < α < 1 (10) 1−α In the case of a stationary steady-state (no growth): with gx = gy = 0, (10) simplifies to k = β0/(1 − α). 10 9 The homogenous ECM The ECM shows that there are important points of correspondence between the dynamic ADL model and a static relationship for long-run: 1. A theoretical linear relationship, y ∗ = k + γx, represents the steady-state solution of the dynamic model (1). 2. The theoretical long-run slope coefficient γ is identical to the corresponding long-run multiplier 3. Conversely, if we are only interested in quantifying a long-run multiplier (rather than the whole sequence of dynamic multipliers), it can be found by using the identity in (7). As we have seen, the ECM is a “1-1” transformation of the ADL. It is a reparameterization, there are no parameter restrictions involved, and therefor no loss of information. If the long-run slope coefficient γ is restricted to 1, we get the Homogenous ECM, the last model in the Typology. γ = 1 ⇐⇒ β1 + β2 = −(α − 1) as in Typology The name refers to the property that a permanent unit increase in the exogenous variable leads to a unit long- run increase in y (as with homogeneity of degree one of equilibrium market prices in a general equilibrium model). γ = 1 may fail to match the properties of the data, but in our example of a Norwegian consumption function it fits quite well. From (4) we have ∆ ln Ct = 0.04 + 0.13∆ ln(IN Ct) − 0.21 {ln C − ln IN C}t−1 11 12 2 Static models reconsidered ( Ch 2.6 in IDM) As noted in the first lecture, a static relationship has two distinct interpretations in macroeconomics 1. As an (approximate) descriptions of dynamics This corresponds to simplifying the ADL by setting β2 = 0 and α = 0 in the Typology. In Frisch terminology, this is the case of “infinitely great speed of reaction”. 2. As a long-run relationship which applies to a steady-state situation. The validity of this interpretation only hinges on α being “less than one”. Frisch: “static laws basically express what would happen in the long-run if the static theory’s assumptions prevailed long enough for the phenomena to have time to have time to react in accordance with these assumptions.” The PPP hypothesis It is important to be aware of the two interpretations of static models, and to be able to distinguish between them. Consider for example the real-exchange rate, σ EP ∗ P where P ∗ is the foreign price level, E is the nominal exchange rate (Kroner/Euro) and P is the domestic price level. σ= The purchasing power parity hypothesis, PPP, says that the real exchange rate is constant. But what is the time perspective of the PPP hypothesis? 13 14 3 If PPP is taken as a short-run proposition, then σt = σt−1 = k, and, in the case of exogenous Et: ln Pt = − ln k + ln Et + ln Pt∗, a static price equation which says that the pass-through of a currency change on Pt is full and immediate If on the other hand, PPP is taken a hypothesis of the long-run behaviour, we have instead that σt = σ̄ in a steady-state situation, and Solution and simulation of dynamic models (Ch 2.7) The existence of a long-run multiplier, and thereby the validity of the correspondence between the ADL model and long-run relationships, depends on the autoregressive parameter α in (1) being different from unity. We next show that the parameter α is also crucial for the nature and type of solution of equation (1). ∆ ln Pt = gE + π ∗ where gE and π ∗ denote the long-run constant growth rates of the nominal exchange rates and of foreign prices. On this interpretation, PPP implies full long-run pass-through, but PPP has no implications about the short-run pass through of a change in Et on Pt. 15 3.1 Solution of ADL equations We loose little by considering the case of a deterministic autoregressive model, where we abstract from xt and xt−1, as well as from the disturbance term (εt). One way to achieve this simplification of (1) is to assume that both xt and εt are fixed at their respective constant means: 16 εt = 0 for t = 0, 1, ..., and The solution can be stable, unstable or explosive. xt = mx for t = 0, 1.... stable: often implicit in discussions about the effects of a changes in policy instruments, i.e., multiplier analysis. (example: interest rate effects on inflation) We write the simplified model as yt = β0 + Bmx + αyt−1, where B = β1 + β2. (11) In the following we proceed as if the coefficients β0, β1, β2 and α are known numbers (in practice they will be estimated). unstable: sometimes macroeconomists believe that the effects of a shock linger on after the shock itself has gone away, often referred to as hysteresis. (11) holds for t = 0, 1, 2, .... t = 0 is the initial period. When y0 is a fixed and known number there is a unique sequence of numbers y0, y1, y2, ... which is the solution of (11). explosive: “bubbles” in capital markets. The condition The solution is found by first solving (11) for y1, then for y2 and so on. This recursive method is the same as we used when we obtained the dynamic multipliers. Follow the steps on page 50 in IDM and derive the solution yt = (β0 + Bmx) t−1 X αs + αty0, t = 1, 2, ... −1 < α < 1 (13) is the necessary and sufficient condition for the existence of a (globally asymptotically) stable solution. (12) s=0 18 17 Stable solution −1 < α < 1. Since t−1 X 1 as t → ∞, we have: αs → 1−α 5.10 8 DFystable_ar.m8 DFystable_ar.8 s=0 (14) 1−α where y ∗ denotes the equilibrium of yt. The solution can also be expressed as (see page 52 of IDM). yt = y ∗ + αt(y0 − y ∗). 2 200 205 210 DFyunstable_201 200 205 210 DFyexplosive 15000 3.70 12500 (15) showing that the solution contains a trend, and that the initial condition exerts full influence on all solution periods. Marked contrast to stable case. Explosive solution When α is greater than unity in absolute value the solution is called explosive, for reasons that become obvious when you consult (12). 19 4 5.00 DFyunstable Unstable solution (hysteresis) When α = 1, we obtain from equation (12): yt = (β0 + β1mx)t + y0, t = 1, 2, ... 6 5.05 (β + β1mx) y∗ = 0 10000 3.65 7500 3.60 200 5000 205 200 210 220 230 Figure 1: Panel a) and b): Two stable solutions of (11), (corresponding to positive and negative values of α. Panel c): Two unstable solutions (corresponding to different initial conditions). Panel d). An explosive solution, See the text for details about each case. 20 3.2 ln(C), actual value 12.05 Simulation of dynamic models Solution of estimated ADL for Norwegian consumption 12.00 Solution based on: 11.95 a) observed values of l n(INC t ) and ln(INCt−1 ) for the period 1990(1)-2001(4). b) zero disturbance ε t for the period 1990(1)-2001(4). c) observed values of l n(Ct−1 ) and ln(INCt−1 ) in 1989(4) In applied macroeconomics it is not common to refer to the solution of a ADL model. The common term is instead simulation. 11.90 Specifically, economists use dynamic simulation to denote the case where the solution for period 1 is used to calculate the solution for period 2, and the solution for period 2 is in its turn used to find the solution for period 3, and so on. From what we have said above, in the case of a single ADL equation like 11.85 11.80 11.75 11.70 yt = β0 + β1xt + β2xt−1 + αyt−1 + εt, 11.65 dynamic simulation amounts to finding a solution of the model. 11.60 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 Figure 2: Solution of the ADL consumption function for the period 1990(1)2001(4). The actual values of log(Ct) are also shown, for comparison. Conveniently, the correspondence between solution and dynamic simulation also holds for complex equations (with several lags and/or several explanatory variables), as well as for the systems of dynamic equations which are used in practice. But use computer programs to find the solution. 21 4 22 Dynamic systems (IDM ch 2.8) It is often quite easy to bring a system of dynamic equations on a form with two reduced form ADL equations which are of the same form that we have considered above. In the Introduction of ICM, and in the seminar exercises we have already seen examples of dynamic systems: the cobweb model and the “habit model” of a perfectly competitive commodity market. The 2-equation dynamic system has two endogenous variables Ct and IN Ct, while Jt and εt are exogenous. The ADL in (16) has an endogenous variable on the right hand side. Want to find and ADL for Ct with only exogenous and predetermined variables on the right hand side, a so called final equation for Ct. In this example: simply substitute IN C from (17) to give: Ct = β̃0 + α̃Ct−1 + β̃2Jt + ε̃t We take the Keynesian model as our first macroeconomic example. Here, it is most convenient to use a linear specification of the consumption function. Ct = β0 + β1IN Ct + αCt−1 + εt (16) (18) β̃0 and α̃ are β0 and α divided by (1 − β1), and β̃2 = β1/(1 − β1). (18) is the final equation for Ct in this example. (17) If −1 < α̃ < 1 the solution is stable. In that case, there is also a stable solution for IN Ct, so we don’t have to derive a separate equation for IN Ct in order to check stability of income. where Jt denotes autonomous expenditure, and IN C is now interpreted as GDP. The impact multiplier of consumption with respect to autonomous expenditure is β̃2, and the long-run multiplier is β̃2/(1 − α̃). 23 24 together with the product market equilibrium condition IN Ct = Ct + Jt